A binary adder

Suppose we wanted to build a device that could add two binary bits together. Such a device is known as a half-adder, and its gate circuit looks like this:

The Σ symbol represents the "sum" output of the half-adder, the sum's least significant bit (LSB). C_out represents the "carry" output of the half-adder, the sum's most significant bit (MSB).

If we were to implement this same function in ladder (relay) logic, it would look like this:

Either circuit is capable of adding two binary digits together. The mathematical "rules" of how to add bits together are intrinsic to the hard-wired logic of the circuits. If we wanted to perform a different arithmetic operation with binary bits, such as multiplication, we would have to construct another circuit. The above circuit designs will only perform one function: add two binary bits together. To make them do something else would take re-wiring, and perhaps different componentry.

In this sense, digital arithmetic circuits aren't much different from analog arithmetic (operational amplifier) circuits: they do exactly what they're wired to do, no more and no less. We are not, however, restricted to designing digital computer circuits in this manner. It is possible to embed the mathematical "rules" for any arithmetic operation in the form of digital data rather than in hard-wired connections between gates. The result is unparalleled flexibility in operation, giving rise to a whole new kind of digital device: the programmable computer.

While this chapter is by no means exhaustive, it provides what I believe is a unique and interesting look at the nature of programmable computer devices, starting with two devices often overlooked in introductory textbooks: look-up table memories and finite-state machines.

Look-up tables

Having learned about digital memory devices in the last chapter, we know that it is possible to store binary data within solid-state devices. Those storage "cells" within solid-state memory devices are easily addressed by driving the "address" lines of the device with the proper binary value(s). Suppose we had a ROM memory circuit written, or programmed, with certain data, such that the address lines of the ROM served as inputs and the data lines of the ROM served as outputs, generating the characteristic response of a particular logic function. Theoretically, we could program this ROM chip to emulate whatever logic function we wanted without having to alter any wire connections or gates.

Consider the following example of a 4 x 2 bit ROM memory (a very small memory!) programmed with the functionality of a half adder:

If this ROM has been written with the above data (representing a half-adder's truth table), driving the A and B address inputs will cause the respective memory cells in the ROM chip to be enabled, thus outputting the corresponding data as the Σ (Sum) and C_out bits. Unlike the half-adder circuit built of gates or relays, this device can be set up to perform any logic function at all with two inputs and two outputs, not just the half-adder function. To change the logic function, all we would need to do is write a different table of data to another ROM chip. We could even use an EPROM chip which could be re-written at will, giving the ultimate flexibility in function.

It is vitally important to recognize the significance of this principle as applied to digital circuitry. Whereas the half-adder built from gates or relays processes the input bits to arrive at a specific output, the ROM simply remembers what the outputs should be for any given combination of inputs. This is not much different from the "times tables" memorized in grade school: rather than having to calculate the product of 5 times 6 (5 + 5 + 5 + 5 + 5 + 5 = 30), school-children are taught to remember that 5 x 6 = 30, and then expected to recall this product from memory as needed. Likewise, rather than the logic function depending on the functional arrangement of hard-wired gates or relays (hardware), it depends solely on the data written into the memory (software).

Such a simple application, with definite outputs for every input, is called a look-up table, because the memory device simply "looks up" what the output(s) should to be for any given combination of inputs states.

This application of a memory device to perform logical functions is significant for several reasons:

• Software is much easier to change than hardware.
• Software can be archived on various kinds of memory media (disk, tape), thus providing an easy way to document and manipulate the function in a "virtual" form; hardware can only be "archived" abstractly in the form of some kind of graphical drawing.
• Software can be copied from one memory device (such as the EPROM chip) to another, allowing the ability for one device to "learn" its function from another device.
• Software such as the logic function example can be designed to perform functions that would be extremely difficult to emulate with discrete logic gates (or relays!).

The usefulness of a look-up table becomes more and more evident with increasing complexity of function. Suppose we wanted to build a 4-bit adder circuit using a ROM. We'd require a ROM with 8 address lines (two 4-bit numbers to be added together), plus 4 data lines (for the signed output):

With 256 addressable memory locations in this ROM chip, we would have a fair amount of programming to do, telling it what binary output to generate for each and every combination of binary inputs. We would also run the risk of making a mistake in our programming and have it output an incorrect sum, if we weren't careful. However, the flexibility of being able to configure this function (or any function) through software alone generally outweighs that costs.

Consider some of the advanced functions we could implement with the above "adder." We know that when we add two sets of numbers in 2's complement signed notation, we risk having the answer overflow. For instance, if we try to add 0111 (decimal 7) to 0110 (decimal 6) with only a 4-bit number field, the answer we'll get is 1001 (decimal -7) instead of the correct value, 13 (7 + 6), which cannot be expressed using 4 signed bits. If we wanted to, we could avoid the strange answers given in overflow conditions by programming this look-up table circuit to output something else in conditions where we know overflow will occur (that is, in any case where the real sum would exceed +7 or -8). One alternative might be to program the ROM to output the quantity 0111 (the maximum positive value that can be represented with 4 signed bits), or any other value that we determined to be more appropriate for the application than the typical overflowed "error" value that a regular adder circuit would output. It's all up to the programmer to decide what he or she wants this circuit to do, because we are no longer limited by the constraints of logic gate functions.

The possibilities don't stop at customized logic functions, either. By adding more address lines to the 256 x 4 ROM chip, we can expand the look-up table to include multiple functions:

With two more address lines, the ROM chip will have 4 times as many addresses as before (1024 instead of 256). This ROM could be programmed so that when A8 and A9 were both low, the output data represented the sum of the two 4-bit binary numbers input on address lines A0 through A7, just as we had with the previous 256 x 4 ROM circuit. For the addresses A8=1 and A9=0, it could be programmed to output the difference (subtraction) between the first 4-bit binary number (A0 through A3) and the second binary number (A4 through A7). For the addresses A8=0 and A9=1, we could program the ROM to output the difference (subtraction) of the two numbers in reverse order (second - first rather than first - second), and finally, for the addresses A8=1 and A9=1, the ROM could be programmed to compare the two inputs and output an indication of equality or inequality. What we will have then is a device that can perform four different arithmetical operations on 4-bit binary numbers, all by "looking up" the answers programmed into it.

If we had used a ROM chip with more than two additional address lines, we could program it with a wider variety of functions to perform on the two 4-bit inputs. There are a number of operations peculiar to binary data (such as parity check or Exclusive-ORing of bits) that we might find useful to have programmed in such a look-up table.

Devices such as this, which can perform a variety of arithmetical tasks as dictated by a binary input code, are known as Arithmetic Logic Units (ALUs), and they comprise one of the essential components of computer technology. Although modern ALUs are more often constructed from very complex combinational logic (gate) circuits for reasons of speed, it should be comforting to know that the exact same functionality may be duplicated with a "dumb" ROM chip programmed with the appropriate look-up table(s). In fact, this exact approach was used by IBM engineers in 1959 with the development of the IBM 1401 and 1620 computers, which used look-up tables to perform addition, rather than binary adder circuitry. The machine was fondly known as the "CADET," which stood for "Can't Add, Doesn't Even Try."

A very common application for look-up table ROMs is in control systems where a custom mathematical function needs to be represented. Such an application is found in computer-controlled fuel injection systems for automobile engines, where the proper air/fuel mixture ratio for efficient and clean operation changes with several environmental and operational variables. Tests performed on engines in research laboratories determine what these ideal ratios are for varying conditions of engine load, ambient air temperature, and barometric air pressure. The variables are measured with sensor transducers, their analog outputs converted to digital signals with A/D circuitry, and those parallel digital signals used as address inputs to a high-capacity ROM chip programmed to output the optimum digital value for air/fuel ratio for any of these given conditions.

Sometimes, ROMs are used to provide one-dimensional look-up table functions, for "correcting" digitized signal values so that they more accurately represent their real-world significance. An example of such a device is a thermocouple transmitter, which measures the millivoltage signal generated by a junction of dissimilar metals and outputs a signal which is supposed to directly correspond to that junction temperature. Unfortunately, thermocouple junctions do not have perfectly linear temperature/voltage responses, and so the raw voltage signal is not perfectly proportional to temperature. By digitizing the voltage signal (A/D conversion) and sending that digital value to the address of a ROM programmed with the necessary correction values, the ROM's programming could eliminate some of the nonlinearity of the thermocouple's temperature-to-millivoltage relationship, so that the final output of the device would be more accurate. The popular instrumentation term for such a look-up table is a digital characterizer.

Another application for look-up tables is in special code translation. A 128 x 8 ROM, for instance, could be used to translate 7-bit ASCII code to 8-bit EBCDIC code:

Again, all that is required is for the ROM chip to be properly programmed with the necessary data so that each valid ASCII input will produce a corresponding EBCDIC output code.

Finite-state machines

Feedback is a fascinating engineering principle. It can turn a rather simple device or process into something substantially more complex. We've seen the effects of feedback intentionally integrated into circuit designs with some rather astounding effects:

• Comparator + negative feedback -----------> controllable-gain amplifier
• Comparator + positive feedback -----------> comparator with hysteresis
• Combinational logic + positive feedback --> multivibrator

In the field of process instrumentation, feedback is used to transform a simple measurement system into something capable of control:

• Measurement system + negative feedback ---> closed-loop control system

Feedback, both positive and negative, has the tendency to add whole new dynamics to the operation of a device or system. Sometimes, these new dynamics find useful application, while other times they are merely interesting. With look-up tables programmed into memory devices, feedback from the data outputs back to the address inputs creates a whole new type of device: the Finite State Machine, or FSM:

The above circuit illustrates the basic idea: the data stored at each address becomes the next storage location that the ROM gets addressed to. The result is a specific sequence of binary numbers (following the sequence programmed into the ROM) at the output, over time. To avoid signal timing problems, though, we need to connect the data outputs back to the address inputs through a 4-bit D-type flip-flop, so that the sequence takes place step by step to the beat of a controlled clock pulse:

An analogy for the workings of such a device might be an array of post-office boxes, each one with an identifying number on the door (the address), and each one containing a piece of paper with the address of another P.O. box written on it (the data). A person, opening the first P.O. box, would find in it the address of the next P.O. box to open. By storing a particular pattern of addresses in the P.O. boxes, we can dictate the sequence in which each box gets opened, and therefore the sequence of which paper gets read.

Having 16 addressable memory locations in the ROM, this Finite State Machine would have 16 different stable "states" in which it could latch. In each of those states, the identity of the next state would be programmed in to the ROM, awaiting the signal of the next clock pulse to be fed back to the ROM as an address. One useful application of such an FSM would be to generate an arbitrary count sequence, such as Gray Code:

Address  -----> Data          Gray Code count sequence:
0000   -------> 0001                 0   0000
0001   -------> 0011                 1   0001
0010   -------> 0110                 2   0011
0011   -------> 0010                 3   0010
0100   -------> 1100                 4   0110
0101   -------> 0100                 5   0111
0110   -------> 0111                 6   0101
0111   -------> 0101                 7   0100
1000   -------> 0000                 8   1100
1001   -------> 1000                 9   1101
1010   -------> 1011                10   1111
1011   -------> 1001                11   1110
1100   -------> 1101                12   1010
1101   -------> 1111                13   1011
1110   -------> 1010                14   1001
1111   -------> 1110                15   1000

Try to follow the Gray Code count sequence as the FSM would do it: starting at 0000, follow the data stored at that address (0001) to the next address, and so on (0011), and so on (0010), and so on (0110), etc. The result, for the program table shown, is that the sequence of addressing jumps around from address to address in what looks like a haphazard fashion, but when you check each address that is accessed, you will find that it follows the correct order for 4-bit Gray code. When the FSM arrives at its last programmed state (address 1000), the data stored there is 0000, which starts the whole sequence over again at address 0000 in step with the next clock pulse.

We could expand on the capabilities of the above circuit by using a ROM with more address lines, and adding more programming data:

Now, just like the look-up table adder circuit that we turned into an Arithmetic Logic Unit (+, -, x, / functions) by utilizing more address lines as "function control" inputs, this FSM counter can be used to generate more than one count sequence, a different sequence programmed for the four feedback bits (A0 through A3) for each of the two function control line input combinations (A4 = 0 or 1).

Address  -----> Data            Address  -----> Data
00000  -------> 0001            10000  -------> 0001
00001  -------> 0010            10001  -------> 0011
00010  -------> 0011            10010  -------> 0110
00011  -------> 0100            10011  -------> 0010
00100  -------> 0101            10100  -------> 1100
00101  -------> 0110            10101  -------> 0100
00110  -------> 0111            10110  -------> 0111
00111  -------> 1000            10111  -------> 0101
01000  -------> 1001            11000  -------> 0000
01001  -------> 1010            11001  -------> 1000
01010  -------> 1011            11010  -------> 1011
01011  -------> 1100            11011  -------> 1001
01100  -------> 1101            11100  -------> 1101
01101  -------> 1110            11101  -------> 1111
01110  -------> 1111            11110  -------> 1010
01111  -------> 0000            11111  -------> 1110

If A4 is 0, the FSM counts in binary; if A4 is 1, the FSM counts in Gray Code. In either case, the counting sequence is arbitrary: determined by the whim of the programmer. For that matter, the counting sequence doesn't even have to have 16 steps, as the programmer may decide to have the sequence recycle to 0000 at any one of the steps at all. It is a completely flexible counting device, the behavior strictly determined by the software (programming) in the ROM.

We can expand on the capabilities of the FSM even more by utilizing a ROM chip with additional address input and data output lines. Take the following circuit, for example:

Here, the D0 through D3 data outputs are used exclusively for feedback to the A0 through A3 address lines. Date output lines D4 through D7 can be programmed to output something other than the FSM's "state" value. Being that four data output bits are being fed back to four address bits, this is still a 16-state device. However, having the output data come from other data output lines gives the programmer more freedom to configure functions than before. In other words, this device can do far more than just count! The programmed output of this FSM is dependent not only upon the state of the feedback address lines (A0 through A3), but also the states of the input lines (A4 through A7). The D-type flip/flop's clock signal input does not have to come from a pulse generator, either. To make things more interesting, the flip/flop could be wired up to clock on some external event, so that the FSM goes to the next state only when an input signal tells it to.

Now we have a device that better fulfills the meaning of the word "programmable." The data written to the ROM is a program in the truest sense: the outputs follow a pre-established order based on the inputs to the device and which "step" the device is on in its sequence. This is very close to the operating design of the Turing Machine, a theoretical computing device invented by Alan Turing, mathematically proven to be able to solve any known arithmetic problem, given enough memory capacity.

Microprocessors

Early computer science pioneers such as Alan Turing and John Von Neumann postulated that for a computing device to be really useful, it not only had to be able to generate specific outputs as dictated by programmed instructions, but it also had to be able to write data to memory, and be able to act on that data later. Both the program steps and the processed data were to reside in a common memory "pool," thus giving way to the label of the stored-program computer. Turing's theoretical machine utilized a sequential-access tape, which would store data for a control circuit to read, the control circuit re-writing data to the tape and/or moving the tape to a new position to read more data. Modern computers use random-access memory devices instead of sequential-access tapes to accomplish essentially the same thing, except with greater capability.

A helpful illustration is that of early automatic machine tool control technology. Called open-loop, or sometimes just NC (numerical control), these control systems would direct the motion of a machine tool such as a lathe or a mill by following instructions programmed as holes in paper tape. The tape would be run one direction through a "read" mechanism, and the machine would blindly follow the instructions on the tape without regard to any other conditions. While these devices eliminated the burden of having to have a human machinist direct every motion of the machine tool, it was limited in usefulness. Because the machine was blind to the real world, only following the instructions written on the tape, it could not compensate for changing conditions such as expansion of the metal or wear of the mechanisms. Also, the tape programmer had to be acutely aware of the sequence of previous instructions in the machine's program to avoid troublesome circumstances (such as telling the machine tool to move the drill bit laterally while it is still inserted into a hole in the work), since the device had no memory other than the tape itself, which was read-only. Upgrading from a simple tape reader to a Finite State control design gave the device a sort of memory that could be used to keep track of what it had already done (through feedback of some of the data bits to the address bits), so at least the programmer could decide to have the circuit remember "states" that the machine tool could be in (such as "coolant on," or tool position). However, there was still room for improvement.

The ultimate approach is to have the program give instructions which would include the writing of new data to a read/write (RAM) memory, which the program could easily recall and process. This way, the control system could record what it had done, and any sensor-detectable process changes, much in the same way that a human machinist might jot down notes or measurements on a scratch-pad for future reference in his or her work. This is what is referred to as CNC, or Closed-loop Numerical Control.

Engineers and computer scientists looked forward to the possibility of building digital devices that could modify their own programming, much the same as the human brain adapts the strength of inter-neural connections depending on environmental experiences (that is why memory retention improves with repeated study, and behavior is modified through consequential feedback). Only if the computer's program were stored in the same writable memory "pool" as the data would this be practical. It is interesting to note that the notion of a self-modifying program is still considered to be on the cutting edge of computer science. Most computer programming relies on rather fixed sequences of instructions, with a separate field of data being the only information that gets altered.

To facilitate the stored-program approach, we require a device that is much more complex than the simple FSM, although many of the same principles apply. First, we need read/write memory that can be easily accessed: this is easy enough to do. Static or dynamic RAM chips do the job well, and are inexpensive. Secondly, we need some form of logic to process the data stored in memory. Because standard and Boolean arithmetic functions are so useful, we can use an Arithmetic Logic Unit (ALU) such as the look-up table ROM example explored earlier. Finally, we need a device that controls how and where data flows between the memory, the ALU, and the outside world. This so-called Control Unit is the most mysterious piece of the puzzle yet, being comprised of tri-state buffers (to direct data to and from buses) and decoding logic which interprets certain binary codes as instructions to carry out. Sample instructions might be something like: "add the number stored at memory address 0010 with the number stored at memory address 1101," or, "determine the parity of the data in memory address 0111." The choice of which binary codes represent which instructions for the Control Unit to decode is largely arbitrary, just as the choice of which binary codes to use in representing the letters of the alphabet in the ASCII standard was largely arbitrary. ASCII, however, is now an internationally recognized standard, whereas control unit instruction codes are almost always manufacturer-specific.

Putting these components together (read/write memory, ALU, and control unit) results in a digital device that is typically called a processor. If minimal memory is used, and all the necessary components are contained on a single integrated circuit, it is called a microprocessor. When combined with the necessary bus-control support circuitry, it is known as a Central Processing Unit, or CPU.

CPU operation is summed up in the so-called fetch/execute cycle. Fetch means to read an instruction from memory for the Control Unit to decode. A small binary counter in the CPU (known as the program counter or instruction pointer) holds the address value where the next instruction is stored in main memory. The Control Unit sends this binary address value to the main memory's address lines, and the memory's data output is read by the Control Unit to send to another holding register. If the fetched instruction requires reading more data from memory (for example, in adding two numbers together, we have to read both the numbers that are to be added from main memory or from some other source), the Control Unit appropriately addresses the location of the requested data and directs the data output to ALU registers. Next, the Control Unit would execute the instruction by signaling the ALU to do whatever was requested with the two numbers, and direct the result to another register called the accumulator. The instruction has now been "fetched" and "executed," so the Control Unit now increments the program counter to step the next instruction, and the cycle repeats itself.

        Microprocessor (CPU)

--------------------------------------
|       ** Program counter **        |
|  (increments address value sent to |
|  external memory chip(s) to fetch  |==========> Address bus
|  the next instruction)             |          (to RAM memory)
--------------------------------------
|        ** Control Unit **          |<=========> Control Bus
|  (decodes instructions read from   | (to all devices sharing
|  program in memory, enables flow   | address and/or data busses;
|  of data to and from ALU, internal | arbitrates all bus communi-
|  registers, and external devices)  | cations)
--------------------------------------
| ** Arithmetic Logic Unit (ALU) **  |
|    (performs all mathematical      |
|     calculations and Boolean       |
|     functions)                     |
--------------------------------------
|          ** Registers **           |
|    (small read/write memories for  |<=========> Data Bus
|     holding instruction codes,     | (from RAM memory and other
|     error codes, ALU data, etc;    |  external devices)
|     includes the "accumulator")    |
--------------------------------------

As one might guess, carrying out even simple instructions is a tedious process. Several steps are necessary for the Control Unit to complete the simplest of mathematical procedures. This is especially true for arithmetic procedures such as exponents, which involve repeated executions ("iterations") of simpler functions. Just imagine the sheer quantity of steps necessary within the CPU to update the bits of information for the graphic display on a flight simulator game! The only thing which makes such a tedious process practical is the fact that microprocessor circuits are able to repeat the fetch/execute cycle with great speed.

In some microprocessor designs, there are minimal programs stored within a special ROM memory internal to the device (called microcode) which handle all the sub-steps necessary to carry out more complex math operations. This way, only a single instruction has to be read from the program RAM to do the task, and the programmer doesn't have to deal with trying to tell the microprocessor how to do every minute step. In essence, its a processor inside of a processor; a program running inside of a program.

Microprocessor programming

The "vocabulary" of instructions which any particular microprocessor chip possesses is specific to that model of chip. An Intel 80386, for example, uses a completely different set of binary codes than a Motorola 68020, for designating equivalent functions. Unfortunately, there are no standards in place for microprocessor instructions. This makes programming at the very lowest level very confusing and specialized.

When a human programmer develops a set of instructions to directly tell a microprocessor how to do something (like automatically control the fuel injection rate to an engine), they're programming in the CPU's own "language." This language, which consists of the very same binary codes which the Control Unit inside the CPU chip decodes to perform tasks, is often referred to as machine language. While machine language software can be "worded" in binary notation, it is often written in hexadecimal form, because it is easier for human beings to work with. For example, I'll present just a few of the common instruction codes for the Intel 8080 micro-processor chip:

Hexadecimal    Binary              Instruction description
-----------   --------   -----------------------------------------
|   7B        01111011   Move contents of register A to register E
|
|   87        10000111   Add contents of register A to register D
|
|   1C        00011100   Increment the contents of register E by 1
|
|   D3        11010011   Output byte of data to data bus

Even with hexadecimal notation, these instructions can be easily confused and forgotten. For this purpose, another aid for programmers exists called assembly language. With assembly language, two to four letter mnemonic words are used in place of the actual hex or binary code for describing program steps. For example, the instruction 7B for the Intel 8080 would be "MOV A,E" in assembly language. The mnemonics, of course, are useless to the microprocessor, which can only understand binary codes, but it is an expedient way for programmers to manage the writing of their programs on paper or text editor (word processor). There are even programs written for computers called assemblers which understand these mnemonics, translating them to the appropriate binary codes for a specified target microprocessor, so that the programmer can write a program in the computer's native language without ever having to deal with strange hex or tedious binary code notation.

Once a program is developed by a person, it must be written into memory before a microprocessor can execute it. If the program is to be stored in ROM (which some are), this can be done with a special machine called a ROM programmer, or (if you're masochistic), by plugging the ROM chip into a breadboard, powering it up with the appropriate voltages, and writing data by making the right wire connections to the address and data lines, one at a time, for each instruction. If the program is to be stored in volatile memory, such as the operating computer's RAM memory, there may be a way to type it in by hand through that computer's keyboard (some computers have a mini-program stored in ROM which tells the microprocessor how to accept keystrokes from a keyboard and store them as commands in RAM), even if it is too dumb to do anything else. Many "hobby" computer kits work like this. If the computer to be programmed is a fully-functional personal computer with an operating system, disk drives, and the whole works, you can simply command the assembler to store your finished program onto a disk for later retrieval. To "run" your program, you would simply type your program's filename at the prompt, press the Enter key, and the microprocessor's Program Counter register would be set to point to the location ("address") on the disk where the first instruction is stored, and your program would run from there.

Although programming in machine language or assembly language makes for fast and highly efficient programs, it takes a lot of time and skill to do so for anything but the simplest tasks, because each machine language instruction is so crude. The answer to this is to develop ways for programmers to write in "high level" languages, which can more efficiently express human thought. Instead of typing in dozens of cryptic assembly language codes, a programmer writing in a high-level language would be able to write something like this . . .

Print "Hello, world!"

. . . and expect the computer to print "Hello, world!" with no further instruction on how to do so. This is a great idea, but how does a microprocessor understand such "human" thinking when its vocabulary is so limited?

The answer comes in two different forms: interpretation, or compilation. Just like two people speaking different languages, there has to be some way to transcend the language barrier in order for them to converse. A translator is needed to translate each person's words to the other person's language, one way at a time. For the microprocessor, this means another program, written by another programmer in machine language, which recognizes the ASCII character patterns of high-level commands such as Print (P-r-i-n-t) and can translate them into the necessary bite-size steps that the microprocessor can directly understand. If this translation is done during program execution, just like a translator intervening between two people in a live conversation, it is called "interpretation." On the other hand, if the entire program is translated to machine language in one fell swoop, like a translator recording a monologue on paper and then translating all the words at one sitting into a written document in the other language, the process is called "compilation."

Interpretation is simple, but makes for a slow-running program because the microprocessor has to continually translate the program between steps, and that takes time. Compilation takes time initially to translate the whole program into machine code, but the resulting machine code needs no translation after that and runs faster as a consequence. Programming languages such as BASIC and FORTH are interpreted. Languages such as C, C++, FORTRAN, and PASCAL are compiled. Compiled languages are generally considered to be the languages of choice for professional programmers, because of the efficiency of the final product.

Naturally, because machine language vocabularies vary widely from microprocessor to microprocessor, and since high-level languages are designed to be as universal as possible, the interpreting and compiling programs necessary for language translation must be microprocessor-specific. Development of these interpreters and compilers is a most impressive feat: the people who make these programs most definitely earn their keep, especially when you consider the work they must do to keep their software product current with the rapidly-changing microprocessor models appearing on the market!

To mitigate this difficulty, the trend-setting manufacturers of microprocessor chips (most notably, Intel and Motorola) try to design their new products to be backwardly compatible with their older products. For example, the entire instruction set for the Intel 80386 chip is contained within the latest Pentium IV chips, although the Pentium chips have additional instructions that the 80386 chips lack. What this means is that machine-language programs (compilers, too) written for 80386 computers will run on the latest and greatest Intel Pentium IV CPU, but machine-language programs written specifically to take advantage of the Pentium's larger instruction set will not run on an 80386, because the older CPU simply doesn't have some of those instructions in its vocabulary: the Control Unit inside the 80386 cannot decode them.

Building on this theme, most compilers have settings that allow the programmer to select which CPU type he or she wants to compile machine-language code for. If they select the 80386 setting, the compiler will perform the translation using only instructions known to the 80386 chip; if they select the Pentium setting, the compiler is free to make use of all instructions known to Pentiums. This is analogous to telling a translator what minimum reading level their audience will be: a document translated for a child will be understandable to an adult, but a document translated for an adult may very well be gibberish to a child.

Read More......

Why digital?

Although many textbooks provide good introductions to digital memory technology, I intend to make this chapter unique in presenting both past and present technologies to some degree of detail. While many of these memory designs are obsolete, their foundational principles are still quite interesting and educational, and may even find re-application in the memory technologies of the future.

The basic goal of digital memory is to provide a means to store and access binary data: sequences of 1's and 0's. The digital storage of information holds advantages over analog techniques much the same as digital communication of information holds advantages over analog communication. This is not to say that digital data storage is unequivocally superior to analog, but it does address some of the more common problems associated with analog techniques and thus finds immense popularity in both consumer and industrial applications. Digital data storage also complements digital computation technology well, and thus finds natural application in the world of computers.

The most evident advantage of digital data storage is the resistance to corruption. Suppose that we were going to store a piece of data regarding the magnitude of a voltage signal by means of magnetizing a small chunk of magnetic material. Since many magnetic materials retain their strength of magnetization very well over time, this would be a logical media candidate for long-term storage of this particular data (in fact, this is precisely how audio and video tape technology works: thin plastic tape is impregnated with particles of iron-oxide material, which can be magnetized or demagnetized via the application of a magnetic field from an electromagnet coil. The data is then retrieved from the tape by moving the magnetized tape past another coil of wire, the magnetized spots on the tape inducing voltage in that coil, reproducing the voltage waveform initially used to magnetize the tape).

If we represent an analog signal by the strength of magnetization on spots of the tape, the storage of data on the tape will be susceptible to the smallest degree of degradation of that magnetization. As the tape ages and the magnetization fades, the analog signal magnitude represented on the tape will appear to be less than what it was when we first recorded the data. Also, if any spurious magnetic fields happen to alter the magnetization on the tape, even if its only by a small amount, that altering of field strength will be interpreted upon re-play as an altering (or corruption) of the signal that was recorded. Since analog signals have infinite resolution, the smallest degree of change will have an impact on the integrity of the data storage.

If we were to use that same tape and store the data in binary digital form, however, the strength of magnetization on the tape would fall into two discrete levels: "high" and "low," with no valid in-between states. As the tape aged or was exposed to spurious magnetic fields, those same locations on the tape would experience slight alteration of magnetic field strength, but unless the alterations were extreme, no data corruption would occur upon re-play of the tape. By reducing the resolution of the signal impressed upon the magnetic tape, we've gained significant immunity to the kind of degradation and "noise" typically plaguing stored analog data. On the other hand, our data resolution would be limited to the scanning rate and the number of bits output by the A/D converter which interpreted the original analog signal, so the reproduction wouldn't necessarily be "better" than with analog, merely more rugged. With the advanced technology of modern A/D's, though, the tradeoff is acceptable for most applications.

Also, by encoding different types of data into specific binary number schemes, digital storage allows us to archive a wide variety of information that is often difficult to encode in analog form. Text, for example, is represented quite easily with the binary ASCII code, seven bits for each character, including punctuation marks, spaces, and carriage returns. A wider range of text is encoded using the Unicode standard, in like manner. Any kind of numerical data can be represented using binary notation on digital media, and any kind of information that can be encoded in numerical form (which almost any kind can!) is storable, too. Techniques such as parity and checksum error detection can be employed to further guard against data corruption, in ways that analog does not lend itself to.

Digital memory terms and concepts

When we store information in some kind of circuit or device, we not only need some way to store and retrieve it, but also to locate precisely where in the device that it is. Most, if not all, memory devices can be thought of as a series of mail boxes, folders in a file cabinet, or some other metaphor where information can be located in a variety of places. When we refer to the actual information being stored in the memory device, we usually refer to it as the data. The location of this data within the storage device is typically called the address, in a manner reminiscent of the postal service.

With some types of memory devices, the address in which certain data is stored can be called up by means of parallel data lines in a digital circuit (we'll discuss this in more detail later in this lesson). With other types of devices, data is addressed in terms of an actual physical location on the surface of some type of media (the tracks and sectors of circular computer disks, for instance). However, some memory devices such as magnetic tapes have a one-dimensional type of data addressing: if you want to play your favorite song in the middle of a cassette tape album, you have to fast-forward to that spot in the tape, arriving at the proper spot by means of trial-and-error, judging the approximate area by means of a counter that keeps track of tape position, and/or by the amount of time it takes to get there from the beginning of the tape. The access of data from a storage device falls roughly into two categories: random access and sequential access. Random access means that you can quickly and precisely address a specific data location within the device, and non-random simply means that you cannot. A vinyl record platter is an example of a random-access device: to skip to any song, you just position the stylus arm at whatever location on the record that you want (compact audio disks so the same thing, only they do it automatically for you). Cassette tape, on the other hand, is sequential. You have to wait to go past the other songs in sequence before you can access or address the song that you want to skip to.

The process of storing a piece of data to a memory device is called writing, and the process of retrieving data is called reading. Memory devices allowing both reading and writing are equipped with a way to distinguish between the two tasks, so that no mistake is made by the user (writing new information to a device when all you wanted to do is see what was stored there). Some devices do not allow for the writing of new data, and are purchased "pre-written" from the manufacturer. Such is the case for vinyl records and compact audio disks, and this is typically referred to in the digital world as read-only memory, or ROM. Cassette audio and video tape, on the other hand, can be re-recorded (re-written) or purchased blank and recorded fresh by the user. This is often called read-write memory.

Another distinction to be made for any particular memory technology is its volatility, or data storage permanence without power. Many electronic memory devices store binary data by means of circuits that are either latched in a "high" or "low" state, and this latching effect holds only as long as electric power is maintained to those circuits. Such memory would be properly referred to as volatile. Storage media such as magnetized disk or tape is nonvolatile, because no source of power is needed to maintain data storage. This is often confusing for new students of computer technology, because the volatile electronic memory typically used for the construction of computer devices is commonly and distinctly referred to as RAM (Random Access Memory). While RAM memory is typically randomly-accessed, so is virtually every other kind of memory device in the computer! What "RAM" really refers to is the volatility of the memory, and not its mode of access. Nonvolatile memory integrated circuits in personal computers are commonly (and properly) referred to as ROM (Read-Only Memory), but their data contents are accessed randomly, just like the volatile memory circuits!

Finally, there needs to be a way to denote how much data can be stored by any particular memory device. This, fortunately for us, is very simple and straightforward: just count up the number of bits (or bytes, 1 byte = 8 bits) of total data storage space. Due to the high capacity of modern data storage devices, metric prefixes are generally affixed to the unit of bytes in order to represent storage space: 1.6 Gigabytes is equal to 1.6 billion bytes, or 12.8 billion bits, of data storage capacity. The only caveat here is to be aware of rounded numbers. Because the storage mechanisms of many random-access memory devices are typically arranged so that the number of "cells" in which bits of data can be stored appears in binary progression (powers of 2), a "one kilobyte" memory device most likely contains 1024 (2 to the power of 10) locations for data bytes rather than exactly 1000. A "64 kbyte" memory device actually holds 65,536 bytes of data (2 to the 16th power), and should probably be called a "66 Kbyte" device to be more precise. When we round numbers in our base-10 system, we fall out of step with the round equivalents in the base-2 system.

Modern nonmechanical memory

Now we can proceed to studying specific types of digital storage devices. To start, I want to explore some of the technologies which do not require any moving parts. These are not necessarily the newest technologies, as one might suspect, although they will most likely replace moving-part technologies in the future.

A very simple type of electronic memory is the bistable multivibrator. Capable of storing a single bit of data, it is volatile (requiring power to maintain its memory) and very fast. The D-latch is probably the simplest implementation of a bistable multivibrator for memory usage, the D input serving as the data "write" input, the Q output serving as the "read" output, and the enable input serving as the read/write control line:

If we desire more than one bit's worth of storage (and we probably do), we'll have to have many latches arranged in some kind of an array where we can selectively address which one (or which set) we're reading from or writing to. Using a pair of tristate buffers, we can connect both the data write input and the data read output to a common data bus line, and enable those buffers to either connect the Q output to the data line (READ), connect the D input to the data line (WRITE), or keep both buffers in the High-Z state to disconnect D and Q from the data line (unaddressed mode). One memory "cell" would look like this, internally:

When the address enable input is 0, both tristate buffers will be placed in high-Z mode, and the latch will be disconnected from the data input/output (bus) line. Only when the address enable input is active (1) will the latch be connected to the data bus. Every latch circuit, of course, will be enabled with a different "address enable" (AE) input line, which will come from a 1-of-n output decoder:

In the above circuit, 16 memory cells are individually addressed with a 4-bit binary code input into the decoder. If a cell is not addressed, it will be disconnected from the 1-bit data bus by its internal tristate buffers: consequently, data cannot be either written or read through the bus to or from that cell. Only the cell circuit that is addressed by the 4-bit decoder input will be accessible through the data bus.

This simple memory circuit is random-access and volatile. Technically, it is known as a static RAM. Its total memory capacity is 16 bits. Since it contains 16 addresses and has a data bus that is 1 bit wide, it would be designated as a 16 x 1 bit static RAM circuit. As you can see, it takes an incredible number of gates (and multiple transistors per gate!) to construct a practical static RAM circuit. This makes the static RAM a relatively low-density device, with less capacity than most other types of RAM technology per unit IC chip space. Because each cell circuit consumes a certain amount of power, the overall power consumption for a large array of cells can be quite high. Early static RAM banks in personal computers consumed a fair amount of power and generated a lot of heat, too. CMOS IC technology has made it possible to lower the specific power consumption of static RAM circuits, but low storage density is still an issue.

To address this, engineers turned to the capacitor instead of the bistable multivibrator as a means of storing binary data. A tiny capacitor could serve as a memory cell, complete with a single MOSFET transistor for connecting it to the data bus for charging (writing a 1), discharging (writing a 0), or reading. Unfortunately, such tiny capacitors have very small capacitances, and their charge tends to "leak" away through any circuit impedances quite rapidly. To combat this tendency, engineers designed circuits internal to the RAM memory chip which would periodically read all cells and recharge (or "refresh") the capacitors as needed. Although this added to the complexity of the circuit, it still required far less componentry than a RAM built of multivibrators. They called this type of memory circuit a dynamic RAM, because of its need of periodic refreshing.

Recent advances in IC chip manufacturing has led to the introduction of flash memory, which works on a capacitive storage principle like the dynamic RAM, but uses the insulated gate of a MOSFET as the capacitor itself.

Before the advent of transistors (especially the MOSFET), engineers had to implement digital circuitry with gates constructed from vacuum tubes. As you can imagine, the enormous comparative size and power consumption of a vacuum tube as compared to a transistor made memory circuits like static and dynamic RAM a practical impossibility. Other, rather ingenious, techniques to store digital data without the use of moving parts were developed.

Historical, nonmechanical memory technologies

Perhaps the most ingenious technique was that of the delay line. A delay line is any kind of device which delays the propagation of a pulse or wave signal. If you've ever heard a sound echo back and forth through a canyon or cave, you've experienced an audio delay line: the noise wave travels at the speed of sound, bouncing off of walls and reversing direction of travel. The delay line "stores" data on a very temporary basis if the signal is not strengthened periodically, but the very fact that it stores data at all is a phenomenon exploitable for memory technology.

Early computer delay lines used long tubes filled with liquid mercury, which was used as the physical medium through which sound waves traveled along the length of the tube. An electrical/sound transducer was mounted at each end, one to create sound waves from electrical impulses, and the other to generate electrical impulses from sound waves. A stream of serial binary data was sent to the transmitting transducer as a voltage signal. The sequence of sound waves would travel from left to right through the mercury in the tube and be received by the transducer at the other end. The receiving transducer would receive the pulses in the same order as they were transmitted:

A feedback circuit connected to the receiving transducer would drive the transmitting transducer again, sending the same sequence of pulses through the tube as sound waves, storing the data as long as the feedback circuit continued to function. The delay line functioned like a first-in-first-out (FIFO) shift register, and external feedback turned that shift register behavior into a ring counter, cycling the bits around indefinitely.

The delay line concept suffered numerous limitations from the materials and technology that were then available. The EDVAC computer of the early 1950's used 128 mercury-filled tubes, each one about 5 feet long and storing a maximum of 384 bits. Temperature changes would affect the speed of sound in the mercury, thus skewing the time delay in each tube and causing timing problems. Later designs replaced the liquid mercury medium with solid rods of glass, quartz, or special metal that delayed torsional (twisting) waves rather than longitudinal (lengthwise) waves, and operated at much higher frequencies.

One such delay line used a special nickel-iron-titanium wire (chosen for its good temperature stability) about 95 feet in length, coiled to reduce the overall package size. The total delay time from one end of the wire to the other was about 9.8 milliseconds, and the highest practical clock frequency was 1 MHz. This meant that approximately 9800 bits of data could be stored in the delay line wire at any given time. Given different means of delaying signals which wouldn't be so susceptible to environmental variables (such as serial pulses of light within a long optical fiber), this approach might someday find re-application.

Another approach experimented with by early computer engineers was the use of a cathode ray tube (CRT), the type commonly used for oscilloscope, radar, and television viewscreens, to store binary data. Normally, the focused and directed electron beam in a CRT would be used to make bits of phosphor chemical on the inside of the tube glow, thus producing a viewable image on the screen. In this application, however, the desired result was the creation of an electric charge on the glass of the screen by the impact of the electron beam, which would then be detected by a metal grid placed directly in front of the CRT. Like the delay line, the so-called Williams Tube memory needed to be periodically refreshed with external circuitry to retain its data. Unlike the delay line mechanisms, it was virtually immune to the environmental factors of temperature and vibration. The IBM model 701 computer sported a Williams Tube memory with 4 Kilobyte capacity and a bad habit of "overcharging" bits on the tube screen with successive re-writes so that false "1" states might overflow to adjacent spots on the screen.

The next major advance in computer memory came when engineers turned to magnetic materials as a means of storing binary data. It was discovered that certain compounds of iron, namely "ferrite," possessed hysteresis curves that were almost square:

Shown on a graph with the strength of the applied magnetic field on the horizontal axis (field intensity), and the actual magnetization (orientation of electron spins in the ferrite material) on the vertical axis (flux density), ferrite won't become magnetized one direction until the applied field exceeds a critical threshold value. Once that critical value is exceeded, the electrons in the ferrite "snap" into magnetic alignment and the ferrite becomes magnetized. If the applied field is then turned off, the ferrite maintains full magnetism. To magnetize the ferrite in the other direction (polarity), the applied magnetic field must exceed the critical value in the opposite direction. Once that critical value is exceeded, the electrons in the ferrite "snap" into magnetic alignment in the opposite direction. Once again, if the applied field is then turned off, the ferrite maintains full magnetism. To put it simply, the magnetization of a piece of ferrite is "bistable."

Exploiting this strange property of ferrite, we can use this natural magnetic "latch" to store a binary bit of data. To set or reset this "latch," we can use electric current through a wire or coil to generate the necessary magnetic field, which will then be applied to the ferrite. Jay Forrester of MIT applied this principle in inventing the magnetic "core" memory, which became the dominant computer memory technology during the 1970's.

A grid of wires, electrically insulated from one another, crossed through the center of many ferrite rings, each of which being called a "core." As DC current moved through any wire from the power supply to ground, a circular magnetic field was generated around that energized wire. The resistor values were set so that the amount of current at the regulated power supply voltage would produce slightly more than 1/2 the critical magnetic field strength needed to magnetize any one of the ferrite rings. Therefore, if column #4 wire was energized, all the cores on that column would be subjected to the magnetic field from that one wire, but it would not be strong enough to change the magnetization of any of those cores. However, if column #4 wire and row #5 wire were both energized, the core at that intersection of column #4 and row #5 would be subjected to a sum of those two magnetic fields: a magnitude strong enough to "set" or "reset" the magnetization of that core. In other words, each core was addressed by the intersection of row and column. The distinction between "set" and "reset" was the direction of the core's magnetic polarity, and that bit value of data would be determined by the polarity of the voltages (with respect to ground) that the row and column wires would be energized with.

The following photograph shows a core memory board from a Data General brand, "Nova" model computer, circa late 1960's or early 1970's. It had a total storage capacity of 4 kbytes (that's kilobytes, not megabytes!). A ball-point pen is shown for size comparison:

The electronic components seen around the periphery of this board are used for "driving" the column and row wires with current, and also to read the status of a core. A close-up photograph reveals the ring-shaped cores, through which the matrix wires thread. Again, a ball-point pen is shown for size comparison:

A core memory board of later design (circa 1971) is shown in the next photograph. Its cores are much smaller and more densely packed, giving more memory storage capacity than the former board (8 kbytes instead of 4 kbytes):

And, another close-up of the cores:

Writing data to core memory was easy enough, but reading that data was a bit of a trick. To facilitate this essential function, a "read" wire was threaded through all the cores in a memory matrix, one end of it being grounded and the other end connected to an amplifier circuit. A pulse of voltage would be generated on this "read" wire if the addressed core changed states (from 0 to 1, or 1 to 0). In other words, to read a core's value, you had to write either a 1 or a 0 to that core and monitor the voltage induced on the read wire to see if the core changed. Obviously, if the core's state was changed, you would have to re-set it back to its original state, or else the data would have been lost. This process is known as a destructive read, because data may be changed (destroyed) as it is read. Thus, refreshing is necessary with core memory, although not in every case (that is, in the case of the core's state not changing when either a 1 or a 0 was written to it).

One major advantage of core memory over delay lines and Williams Tubes was nonvolatility. The ferrite cores maintained their magnetization indefinitely, with no power or refreshing required. It was also relatively easy to build, denser, and physically more rugged than any of its predecessors. Core memory was used from the 1960's until the late 1970's in many computer systems, including the computers used for the Apollo space program, CNC machine tool control computers, business ("mainframe") computers, and industrial control systems. Despite the fact that core memory is long obsolete, the term "core" is still used sometimes with reference to a computer's RAM memory.

All the while that delay lines, Williams Tube, and core memory technologies were being invented, the simple static RAM was being improved with smaller active component (vacuum tube or transistor) technology. Static RAM was never totally eclipsed by its competitors: even the old ENIAC computer of the 1950's used vacuum tube ring-counter circuitry for data registers and computation. Eventually though, smaller and smaller scale IC chip manufacturing technology gave transistors the practical edge over other technologies, and core memory became a museum piece in the 1980's.

One last attempt at a magnetic memory better than core was the bubble memory. Bubble memory took advantage of a peculiar phenomenon in a mineral called garnet, which, when arranged in a thin film and exposed to a constant magnetic field perpendicular to the film, supported tiny regions of oppositely-magnetized "bubbles" that could be nudged along the film by prodding with other external magnetic fields. "Tracks" could be laid on the garnet to focus the movement of the bubbles by depositing magnetic material on the surface of the film. A continuous track was formed on the garnet which gave the bubbles a long loop in which to travel, and motive force was applied to the bubbles with a pair of wire coils wrapped around the garnet and energized with a 2-phase voltage. Bubbles could be created or destroyed with a tiny coil of wire strategically placed in the bubbles' path.

The presence of a bubble represented a binary "1" and the absence of a bubble represented a binary "0." Data could be read and written in this chain of moving magnetic bubbles as they passed by the tiny coil of wire, much the same as the read/write "head" in a cassette tape player, reading the magnetization of the tape as it moves. Like core memory, bubble memory was nonvolatile: a permanent magnet supplied the necessary background field needed to support the bubbles when the power was turned off. Unlike core memory, however, bubble memory had phenomenal storage density: millions of bits could be stored on a chip of garnet only a couple of square inches in size. What killed bubble memory as a viable alternative to static and dynamic RAM was its slow, sequential data access. Being nothing more than an incredibly long serial shift register (ring counter), access to any particular portion of data in the serial string could be quite slow compared to other memory technologies.

An electrostatic equivalent of the bubble memory is the Charge-Coupled Device (CCD) memory, an adaptation of the CCD devices used in digital photography. Like bubble memory, the bits are serially shifted along channels on the substrate material by clock pulses. Unlike bubble memory, the electrostatic charges decay and must be refreshed. CCD memory is therefore volatile, with high storage density and sequential access. Interesting, isn't it? The old Williams Tube memory was adapted from CRT viewing technology, and CCD memory from video recording technology.

Read-only memory

Read-only memory (ROM) is similar in design to static or dynamic RAM circuits, except that the "latching" mechanism is made for one-time (or limited) operation. The simplest type of ROM is that which uses tiny "fuses" which can be selectively blown or left alone to represent the two binary states. Obviously, once one of the little fuses is blown, it cannot be made whole again, so the writing of such ROM circuits is one-time only. Because it can be written (programmed) once, these circuits are sometimes referred to as PROMs (Programmable Read-Only Memory).

However, not all writing methods are as permanent as blown fuses. If a transistor latch can be made which is resettable only with significant effort, a memory device that's something of a cross between a RAM and a ROM can be built. Such a device is given a rather oxymoronic name: the EPROM (Erasable Programmable Read-Only Memory). EPROMs come in two basic varieties: Electrically-erasable (EEPROM) and Ultraviolet-erasable (UV/EPROM). Both types of EPROMs use capacitive charge MOSFET devices to latch on or off. UV/EPROMs are "cleared" by long-term exposure to ultraviolet light. They are easy to identify: they have a transparent glass window which exposes the silicon chip material to light. Once programmed, you must cover that glass window with tape to prevent ambient light from degrading the data over time. EPROMs are often programmed using higher signal voltages than what is used during "read-only" mode.

Memory with moving parts: "Drives"

The earliest forms of digital data storage involving moving parts was that of the punched paper card. Joseph Marie Jacquard invented a weaving loom in 1780 which automatically followed weaving instructions set by carefully placed holes in paper cards. This same technology was adapted to electronic computers in the 1950's, with the cards being read mechanically (metal-to-metal contact through the holes), pneumatically (air blown through the holes, the presence of a hole sensed by air nozzle backpressure), or optically (light shining through the holes).

An improvement over paper cards is the paper tape, still used in some industrial environments (notably the CNC machine tool industry), where data storage and speed demands are low and ruggedness is highly valued. Instead of wood-fiber paper, mylar material is often used, with optical reading of the tape being the most popular method.

Magnetic tape (very similar to audio or video cassette tape) was the next logical improvement in storage media. It is still widely used today, as a means to store "backup" data for archiving and emergency restoration for other, faster methods of data storage. Like paper tape, magnetic tape is sequential access, rather than random access. In early home computer systems, regular audio cassette tape was used to store data in modulated form, the binary 1's and 0's represented by different frequencies (similar to FSK data communication). Access speed was terribly slow (if you were reading ASCII text from the tape, you could almost keep up with the pace of the letters appearing on the computer's screen!), but it was cheap and fairly reliable.

Tape suffered the disadvantage of being sequential access. To address this weak point, magnetic storage "drives" with disk- or drum-shaped media were built. An electric motor provided constant-speed motion. A movable read/write coil (also known as a "head") was provided which could be positioned via servo-motors to various locations on the height of the drum or the radius of the disk, giving access that is almost random (you might still have to wait for the drum or disk to rotate to the proper position once the read/write coil has reached the right location).

The disk shape lent itself best to portable media, and thus the floppy disk was born. Floppy disks (so-called because the magnetic media is thin and flexible) were originally made in 8-inch diameter formats. Later, the 5-1/4 inch variety was introduced, which was made practical by advances in media particle density. All things being equal, a larger disk has more space upon which to write data. However, storage density can be improved by making the little grains of iron-oxide material on the disk substrate smaller. Today, the 3-1/2 inch floppy disk is the preeminent format, with a capacity of 1.44 Mbytes (2.88 Mbytes on SCSI drives). Other portable drive formats are becoming popular, with IoMega's 100 Mbyte "ZIP" and 1 Gbyte "JAZ" disks appearing as original equipment on some personal computers.

Still, floppy drives have the disadvantage of being exposed to harsh environments, being constantly removed from the drive mechanism which reads, writes, and spins the media. The first disks were enclosed units, sealed from all dust and other particulate matter, and were definitely not portable. Keeping the media in an enclosed environment allowed engineers to avoid dust altogether, as well as spurious magnetic fields. This, in turn, allowed for much closer spacing between the head and the magnetic material, resulting in a much tighter-focused magnetic field to write data to the magnetic material.

The following photograph shows a hard disk drive "platter" of approximately 30 Mbytes storage capacity. A ball-point pen has been set near the bottom of the platter for size reference:

Modern disk drives use multiple platters made of hard material (hence the name, "hard drive") with multiple read/write heads for every platter. The gap between head and platter is much smaller than the diameter of a human hair. If the hermetically-sealed environment inside a hard disk drive is contaminated with outside air, the hard drive will be rendered useless. Dust will lodge between the heads and the platters, causing damage to the surface of the media.

Here is a hard drive with four platters, although the angle of the shot only allows viewing of the top platter. This unit is complete with drive motor, read/write heads, and associated electronics. It has a storage capacity of 340 Mbytes, and is about the same length as the ball-point pen shown in the previous photograph:

While it is inevitable that non-moving-part technology will replace mechanical drives in the future, current state-of-the-art electromechanical drives continue to rival "solid-state" nonvolatile memory devices in storage density, and at a lower cost. In 1998, a 250 Mbyte hard drive was announced that was approximately the size of a quarter (smaller than the metal platter hub in the center of the last hard disk photograph)! In any case, storage density and reliability will undoubtedly continue to improve.

An incentive for digital data storage technology advancement was the advent of digitally encoded music. A joint venture between Sony and Phillips resulted in the release of the "compact audio disc" (CD) to the public in the late 1980's. This technology is a read-only type, the media being a transparent plastic disc backed by a thin film of aluminum. Binary bits are encoded as pits in the plastic which vary the path length of a low-power laser beam. Data is read by the low-power laser (the beam of which can be focused more precisely than normal light) reflecting off the aluminum to a photocell receiver.

The advantages of CDs over magnetic tape are legion. Being digital, the information is highly resistant to corruption. Being non-contact in operation, there is no wear incurred through playing. Being optical, they are immune to magnetic fields (which can easily corrupt data on magnetic tape or disks). It is possible to purchase CD "burner" drives which contain the high-power laser necessary to write to a blank disc.

Following on the heels of the music industry, the video entertainment industry has leveraged the technology of optical storage with the introduction of the Digital Video Disc, or DVD. Using a similar-sized plastic disc as the music CD, a DVD employs closer spacing of pits to achieve much greater storage density. This increased density allows feature-length movies to be encoded on DVD media, complete with trivia information about the movie, director's notes, and so on.

Much effort is being directed toward the development of a practical read/write optical disc (CD-W). Success has been found in using chemical substances whose color may be changed through exposure to bright laser light, then "read" by lower-intensity light. These optical discs are immediately identified by their characteristically colored surfaces, as opposed to the silver-colored underside of a standard CD.

Read More......

Introduction

In the design of large and complex digital systems, it is often necessary to have one device communicate digital information to and from other devices. One advantage of digital information is that it tends to be far more resistant to transmitted and interpreted errors than information symbolized in an analog medium. This accounts for the clarity of digitally-encoded telephone connections, compact audio disks, and for much of the enthusiasm in the engineering community for digital communications technology. However, digital communication has its own unique pitfalls, and there are multitudes of different and incompatible ways in which it can be sent. Hopefully, this chapter will enlighten you as to the basics of digital communication, its advantages, disadvantages, and practical considerations.

Suppose we are given the task of remotely monitoring the level of a water storage tank. Our job is to design a system to measure the level of water in the tank and send this information to a distant location so that other people may monitor it. Measuring the tank's level is quite easy, and can be accomplished with a number of different types of instruments, such as float switches, pressure transmitters, ultrasonic level detectors, capacitance probes, strain gauges, or radar level detectors.

For the sake of this illustration, we will use an analog level-measuring device with an output signal of 4-20 mA. 4 mA represents a tank level of 0%, 20 mA represents a tank level of 100%, and anything in between 4 and 20 mA represents a tank level proportionately between 0% and 100%. If we wanted to, we could simply send this 4-20 milliamp analog current signal to the remote monitoring location by means of a pair of copper wires, where it would drive a panel meter of some sort, the scale of which was calibrated to reflect the depth of water in the tank, in whatever units of measurement preferred.

This analog communication system would be simple and robust. For many applications, it would suffice for our needs perfectly. But, it is not the only way to get the job done. For the purposes of exploring digital techniques, we'll explore other methods of monitoring this hypothetical tank, even though the analog method just described might be the most practical.

The analog system, as simple as it may be, does have its limitations. One of them is the problem of analog signal interference. Since the tank's water level is symbolized by the magnitude of DC current in the circuit, any "noise" in this signal will be interpreted as a change in the water level. With no noise, a plot of the current signal over time for a steady tank level of 50% would look like this:

If the wires of this circuit are arranged too close to wires carrying 60 Hz AC power, for example, inductive and capacitive coupling may create a false "noise" signal to be introduced into this otherwise DC circuit. Although the low impedance of a 4-20 mA loop (250 Ω, typically) means that small noise voltages are significantly loaded (and thereby attenuated by the inefficiency of the capacitive/inductive coupling formed by the power wires), such noise can be significant enough to cause measurement problems:

The above example is a bit exaggerated, but the concept should be clear: any electrical noise introduced into an analog measurement system will be interpreted as changes in the measured quantity. One way to combat this problem is to symbolize the tank's water level by means of a digital signal instead of an analog signal. We can do this really crudely by replacing the analog transmitter device with a set of water level switches mounted at different heights on the tank:

Each of these switches is wired to close a circuit, sending current to individual lamps mounted on a panel at the monitoring location. As each switch closed, its respective lamp would light, and whoever looked at the panel would see a 5-lamp representation of the tank's level.

Being that each lamp circuit is digital in nature -- either 100% on or 100% off -- electrical interference from other wires along the run have much less effect on the accuracy of measurement at the monitoring end than in the case of the analog signal. A huge amount of interference would be required to cause an "off" signal to be interpreted as an "on" signal, or vice versa. Relative resistance to electrical interference is an advantage enjoyed by all forms of digital communication over analog.

Now that we know digital signals are far more resistant to error induced by "noise," let's improve on this tank level measurement system. For instance, we could increase the resolution of this tank gauging system by adding more switches, for more precise determination of water level. Suppose we install 16 switches along the tank's height instead of five. This would significantly improve our measurement resolution, but at the expense of greatly increasing the quantity of wires needing to be strung between the tank and the monitoring location. One way to reduce this wiring expense would be to use a priority encoder to take the 16 switches and generate a binary number which represented the same information:

Now, only 4 wires (plus any ground and power wires necessary) are needed to communicate the information, as opposed to 16 wires (plus any ground and power wires). At the monitoring location, we would need some kind of display device that could accept the 4-bit binary data and generate an easy-to-read display for a person to view. A decoder, wired to accept the 4-bit data as its input and light 1-of-16 output lamps, could be used for this task, or we could use a 4-bit decoder/driver circuit to drive some kind of numerical digit display.

Still, a resolution of 1/16 tank height may not be good enough for our application. To better resolve the water level, we need more bits in our binary output. We could add still more switches, but this gets impractical rather quickly. A better option would be to re-attach our original analog transmitter to the tank and electronically convert its 4-20 milliamp analog output into a binary number with far more bits than would be practical using a set of discrete level switches. Since the electrical noise we're trying to avoid is encountered along the long run of wire from the tank to the monitoring location, this A/D conversion can take place at the tank (where we have a "clean" 4-20 mA signal). There are a variety of methods to convert an analog signal to digital, but we'll skip an in-depth discussion of those techniques and concentrate on the digital signal communication itself.

The type of digital information being sent from our tank instrumentation to the monitoring instrumentation is referred to as parallel digital data. That is, each binary bit is being sent along its own dedicated wire, so that all bits arrive at their destination simultaneously. This obviously necessitates the use of at least one wire per bit to communicate with the monitoring location. We could further reduce our wiring needs by sending the binary data along a single channel (one wire + ground), so that each bit is communicated one at a time. This type of information is referred to as serial digital data.

We could use a multiplexer or a shift register to take the parallel data from the A/D converter (at the tank transmitter), and convert it to serial data. At the receiving end (the monitoring location) we could use a demultiplexer or another shift register to convert the serial data to parallel again for use in the display circuitry. The exact details of how the mux/demux or shift register pairs are maintained in synchronization is, like A/D conversion, a topic for another lesson. Fortunately, there are digital IC chips called UARTs (Universal Asynchronous Receiver-Transmitters) that handle all these details on their own and make the designer's life much simpler. For now, we must continue to focus our attention on the matter at hand: how to communicate the digital information from the tank to the monitoring location.

Networks and busses

This collection of wires that I keep referring to between the tank and the monitoring location can be called a bus or a network. The distinction between these two terms is more semantic than technical, and the two may be used interchangeably for all practical purposes. In my experience, the term "bus" is usually used in reference to a set of wires connecting digital components within the enclosure of a computer device, and "network" is for something that is physically more widespread. In recent years, however, the word "bus" has gained popularity in describing networks that specialize in interconnecting discrete instrumentation sensors over long distances ("Fieldbus" and "Profibus" are two examples). In either case, we are making reference to the means by which two or more digital devices are connected together so that data can be communicated between them.

Names like "Fieldbus" or "Profibus" encompass not only the physical wiring of the bus or network, but also the specified voltage levels for communication, their timing sequences (especially for serial data transmission), connector pinout specifications, and all other distinguishing technical features of the network. In other words, when we speak of a certain type of bus or network by name, we're actually speaking of a communications standard, roughly analogous to the rules and vocabulary of a written language. For example, before two or more people can become pen-pals, they must be able to write to one another in a common format. To merely have a mail system that is able to deliver their letters to each other is not enough. If they agree to write to each other in French, they agree to hold to the conventions of character set, vocabulary, spelling, and grammar that is specified by the standard of the French language. Likewise, if we connect two Profibus devices together, they will be able to communicate with each other only because the Profibus standard has specified such important details as voltage levels, timing sequences, etc. Simply having a set of wires strung between multiple devices is not enough to construct a working system (especially if the devices were built by different manufacturers!).

To illustrate in detail, let's design our own bus standard. Taking the crude water tank measurement system with five switches to detect varying levels of water, and using (at least) five wires to conduct the signals to their destination, we can lay the foundation for the mighty BogusBus:

The physical wiring for the BogusBus consists of seven wires between the transmitter device (switches) and the receiver device (lamps). The transmitter consists of all components and wiring connections to the left of the leftmost connectors (the "-->>--" symbols). Each connector symbol represents a complementary male and female element. The bus wiring consists of the seven wires between the connector pairs. Finally, the receiver and all of its constituent wiring lies to the right of the rightmost connectors. Five of the network wires (labeled 1 through 5) carry the data while two of those wires (labeled +V and -V) provide connections for DC power supplies. There is a standard for the 7-pin connector plugs, as well. The pin layout is asymmetrical to prevent "backward" connection.

In order for manufacturers to receive the awe-inspiring "BogusBus-compliant" certification on their products, they would have to comply with the specifications set by the designers of BogusBus (most likely another company, which designed the bus for a specific task and ended up marketing it for a wide variety of purposes). For instance, all devices must be able to use the 24 Volt DC supply power of BogusBus: the switch contacts in the transmitter must be rated for switching that DC voltage, the lamps must definitely be rated for being powered by that voltage, and the connectors must be able to handle it all. Wiring, of course, must be in compliance with that same standard: lamps 1 through 5, for example, must be wired to the appropriate pins so that when LS4 of Manufacturer XYZ's transmitter closes, lamp 4 of Manufacturer ABC's receiver lights up, and so on. Since both transmitter and receiver contain DC power supplies rated at an output of 24 Volts, all transmitter/receiver combinations (from all certified manufacturers) must have power supplies that can be safely wired in parallel. Consider what could happen if Manufacturer XYZ made a transmitter with the negative (-) side of their 24VDC power supply attached to earth ground and Manufacturer ABC made a receiver with the positive (+) side of their 24VDC power supply attached to earth ground. If both earth grounds are relatively "solid" (that is, a low resistance between them, such as might be the case if the two grounds were made on the metal structure of an industrial building), the two power supplies would short-circuit each other!

BogusBus, of course, is a completely hypothetical and very impractical example of a digital network. It has incredibly poor data resolution, requires substantial wiring to connect devices, and communicates in only a single direction (from transmitter to receiver). It does, however, suffice as a tutorial example of what a network is and some of the considerations associated with network selection and operation.

There are many types of buses and networks that you might come across in your profession. Each one has its own applications, advantages, and disadvantages. It is worthwhile to associate yourself with some of the "alphabet soup" that is used to label the various designs:

Short-distance busses

PC/AT Bus used in early IBM-compatible computers to connect peripheral devices such as disk drive and sound cards to the motherboard of the computer.

PCI Another bus used in personal computers, but not limited to IBM-compatibles. Much faster than PC/AT. Typical data transfer rate of 100 Mbytes/second (32 bit) and 200 Mbytes/second (64 bit).

PCMCIA A bus designed to connect peripherals to laptop and notebook sized personal computers. Has a very small physical "footprint," but is considerably slower than other popular PC buses.

VME A high-performance bus (co-designed by Motorola, and based on Motorola's earlier Versa-Bus standard) for constructing versatile industrial and military computers, where multiple memory, peripheral, and even microprocessor cards could be plugged in to a passive "rack" or "card cage" to facilitate custom system designs. Typical data transfer rate of 50 Mbytes/second (64 bits wide).

VXI Actually an expansion of the VME bus, VXI (VME eXtension for Instrumentation) includes the standard VME bus along with connectors for analog signals between cards in the rack.

S-100 Sometimes called the Altair bus, this bus standard was the product of a conference in 1976, intended to serve as an interface to the Intel 8080 microprocessor chip. Similar in philosophy to the VME, where multiple function cards could be plugged in to a passive "rack," facilitating the construction of custom systems.

MC6800 The Motorola equivalent of the Intel-centric S-100 bus, designed to interface peripheral devices to the popular Motorola 6800 microprocessor chip.

STD Stands for Simple-To-Design, and is yet another passive "rack" similar to the PC/AT bus, and lends itself well toward designs based on IBM-compatible hardware. Designed by Pro-Log, it is 8 bits wide (parallel), accommodating relatively small (4.5 inch by 6.5 inch) circuit cards.

Multibus I and II Another bus intended for the flexible design of custom computer systems, designed by Intel. 16 bits wide (parallel).

CompactPCI An industrial adaptation of the personal computer PCI standard, designed as a higher-performance alternative to the older VME bus. At a bus clock speed of 66 MHz, data transfer rates are 200 Mbytes/ second (32 bit) or 400 Mbytes/sec (64 bit).

Microchannel Yet another bus, this one designed by IBM for their ill-fated PS/2 series of computers, intended for the interfacing of PC motherboards to peripheral devices.

IDE A bus used primarily for connecting personal computer hard disk drives with the appropriate peripheral cards. Widely used in today's personal computers for hard drive and CD-ROM drive interfacing.

SCSI An alternative (technically superior to IDE) bus used for personal computer disk drives. SCSI stands for Small Computer System Interface. Used in some IBM-compatible PC's, as well as Macintosh (Apple), and many mini and mainframe business computers. Used to interface hard drives, CD-ROM drives, floppy disk drives, printers, scanners, modems, and a host of other peripheral devices. Speeds up to 1.5 Mbytes per second for the original standard.

GPIB (IEEE 488) General Purpose Interface Bus, also known as HPIB or IEEE 488, which was intended for the interfacing of electronic test equipment such as oscilloscopes and multimeters to personal computers. 8 bit wide address/data "path" with 8 additional lines for communications control.

Centronics parallel Widely used on personal computers for interfacing printer and plotter devices. Sometimes used to interface with other peripheral devices, such as external ZIP (100 Mbyte floppy) disk drives and tape drives.

USB Universal Serial Bus, which is intended to interconnect many external peripheral devices (such as keyboards, modems, mice, etc.) to personal computers. Long used on Macintosh PC's, it is now being installed as new equipment on IBM-compatible machines.

FireWire (IEEE 1394) A high-speed serial network capable of operating at 100, 200, or 400 Mbps with versatile features such as "hot swapping" (adding or removing devices with the power on) and flexible topology. Designed for high-performance personal computer interfacing.

Bluetooth A radio-based communications network designed for office linking of computer devices. Provisions for data security designed into this network standard.

Extended-distance networks

20 mA current loop Not to be confused with the common instrumentation 4-20 mA analog standard, this is a digital communications network based on interrupting a 20 mA (or sometimes 60 mA) current loop to represent binary data. Although the low impedance gives good noise immunity, it is susceptible to wiring faults (such as breaks) which would fail the entire network.

RS-232C The most common serial network used in computer systems, often used to link peripheral devices such as printers and mice to a personal computer. Limited in speed and distance (typically 45 feet and 20 kbps, although higher speeds can be run with shorter distances). I've been able to run RS-232 reliably at speeds in excess of 100 kbps, but this was using a cable only 6 feet long! RS-232C is often referred to simply as RS-232 (no "C").

RS-422A/RS-485 Two serial networks designed to overcome some of the distance and versatility limitations of RS-232C. Used widely in industry to link serial devices together in electrically "noisy" plant environments. Much greater distance and speed limitations than RS-232C, typically over half a mile and at speeds approaching 10 Mbps.

Ethernet (IEEE 802.3) A high-speed network which links computers and some types of peripheral devices together. "Normal" Ethernet runs at a speed of 10 million bits/second, and "Fast" Ethernet runs at 100 million bits/second. The slower (10 Mbps) Ethernet has been implemented in a variety of means on copper wire (thick coax = "10BASE5", thin coax = "10BASE2", twisted-pair = "10BASE-T"), radio, and on optical fiber ("10BASE-F"). The Fast Ethernet has also been implemented on a few different means (twisted-pair, 2 pair = 100BASE-TX; twisted-pair, 4 pair = 100BASE-T4; optical fiber = 100BASE-FX).

Token ring Another high-speed network linking computer devices together, using a philosophy of communication that is much different from Ethernet, allowing for more precise response times from individual network devices, and greater immunity to network wiring damage.

FDDI A very high-speed network exclusively implemented on fiber-optic cabling.

Modbus/Modbus Plus Originally implemented by the Modicon corporation, a large maker of Programmable Logic Controllers (PLCs) for linking remote I/O (Input/Output) racks with a PLC processor. Still quite popular.

Profibus Originally implemented by the Siemens corporation, another large maker of PLC equipment.

Foundation Fieldbus A high-performance bus expressly designed to allow multiple process instruments (transmitters, controllers, valve positioners) to communicate with host computers and with each other. May ultimately displace the 4-20 mA analog signal as the standard means of interconnecting process control instrumentation in the future.

Data flow

Buses and networks are designed to allow communication to occur between individual devices that are interconnected. The flow of information, or data, between nodes can take a variety of forms:

With simplex communication, all data flow is unidirectional: from the designated transmitter to the designated receiver. BogusBus is an example of simplex communication, where the transmitter sent information to the remote monitoring location, but no information is ever sent back to the water tank. If all we want to do is send information one-way, then simplex is just fine. Most applications, however, demand more:

With duplex communication, the flow of information is bidirectional for each device. Duplex can be further divided into two sub-categories:

Half-duplex communication may be likened to two tin cans on the ends of a single taut string: Either can may be used to transmit or receive, but not at the same time. Full-duplex communication is more like a true telephone, where two people can talk at the same time and hear one another simultaneously, the mouthpiece of one phone transmitting the the earpiece of the other, and vice versa. Full-duplex is often facilitated through the use of two separate channels or networks, with an individual set of wires for each direction of communication. It is sometimes accomplished by means of multiple-frequency carrier waves, especially in radio links, where one frequency is reserved for each direction of communication.

Electrical signal types

With BogusBus, our signals were very simple and straightforward: each signal wire (1 through 5) carried a single bit of digital data, 0 Volts representing "off" and 24 Volts DC representing "on." Because all the bits arrived at their destination simultaneously, we would call BogusBus a parallel network technology. If we were to improve the performance of BogusBus by adding binary encoding (to the transmitter end) and decoding (to the receiver end), so that more steps of resolution were available with fewer wires, it would still be a parallel network. If, however, we were to add a parallel-to-serial converter at the transmitter end and a serial-to-parallel converter at the receiver end, we would have something quite different.

It is primarily with the use of serial technology that we are forced to invent clever ways to transmit data bits. Because serial data requires us to send all data bits through the same wiring channel from transmitter to receiver, it necessitates a potentially high frequency signal on the network wiring. Consider the following illustration: a modified BogusBus system is communicating digital data in parallel, binary-encoded form. Instead of 5 discrete bits like the original BogusBus, we're sending 8 bits from transmitter to receiver. The A/D converter on the transmitter side generates a new output every second. That makes for 8 bits per second of data being sent to the receiver. For the sake of illustration, let's say that the transmitter is bouncing between an output of 10101010 and 10101011 every update (once per second):

Since only the least significant bit (Bit 1) is changing, the frequency on that wire (to ground) is only 1/2 Hertz. In fact, no matter what numbers are being generated by the A/D converter between updates, the frequency on any wire in this modified BogusBus network cannot exceed 1/2 Hertz, because that's how fast the A/D updates its digital output. 1/2 Hertz is pretty slow, and should present no problems for our network wiring.

On the other hand, if we used an 8-bit serial network, all data bits must appear on the single channel in sequence. And these bits must be output by the transmitter within the 1-second window of time between A/D converter updates. Therefore, the alternating digital output of 10101010 and 10101011 (once per second) would look something like this:

The frequency of our BogusBus signal is now approximately 4 Hertz instead of 1/2 Hertz, an eightfold increase! While 4 Hertz is still fairly slow, and does not constitute an engineering problem, you should be able to appreciate what might happen if we were transmitting 32 or 64 bits of data per update, along with the other bits necessary for parity checking and signal synchronization, at an update rate of thousands of times per second! Serial data network frequencies start to enter the radio range, and simple wires begin to act as antennas, pairs of wires as transmission lines, with all their associated quirks due to inductive and capacitive reactances.

What is worse, the signals that we're trying to communicate along a serial network are of a square-wave shape, being binary bits of information. Square waves are peculiar things, being mathematically equivalent to an infinite series of sine waves of diminishing amplitude and increasing frequency. A simple square wave at 10 kHz is actually "seen" by the capacitance and inductance of the network as a series of multiple sine-wave frequencies which extend into the hundreds of kHz at significant amplitudes. What we receive at the other end of a long 2-conductor network won't look like a clean square wave anymore, even under the best of conditions!

When engineers speak of network bandwidth, they're referring to the practical frequency limit of a network medium. In serial communication, bandwidth is a product of data volume (binary bits per transmitted "word") and data speed ("words" per second). The standard measure of network bandwidth is bits per second, or bps. An obsolete unit of bandwidth known as the baud is sometimes falsely equated with bits per second, but is actually the measure of signal level changes per second. Many serial network standards use multiple voltage or current level changes to represent a single bit, and so for these applications bps and baud are not equivalent.

The general BogusBus design, where all bits are voltages referenced to a common "ground" connection, is the worst-case situation for high-frequency square wave data communication. Everything will work well for short distances, where inductive and capacitive effects can be held to a minimum, but for long distances this method will surely be problematic:

A robust alternative to the common ground signal method is the differential voltage method, where each bit is represented by the difference of voltage between a ground-isolated pair of wires, instead of a voltage between one wire and a common ground. This tends to limit the capacitive and inductive effects imposed upon each signal and the tendency for the signals to be corrupted due to outside electrical interference, thereby significantly improving the practical distance of a serial network:

The triangular amplifier symbols represent differential amplifiers, which output a voltage signal between two wires, neither one electrically common with ground. Having eliminated any relation between the voltage signal and ground, the only significant capacitance imposed on the signal voltage is that existing between the two signal wires. Capacitance between a signal wire and a grounded conductor is of much less effect, because the capacitive path between the two signal wires via a ground connection is two capacitances in series (from signal wire #1 to ground, then from ground to signal wire #2), and series capacitance values are always less than any of the individual capacitances. Furthermore, any "noise" voltage induced between the signal wires and earth ground by an external source will be ignored, because that noise voltage will likely be induced on both signal wires in equal measure, and the receiving amplifier only responds to the differential voltage between the two signal wires, rather than the voltage between any one of them and earth ground.

RS-232C is a prime example of a ground-referenced serial network, while RS-422A is a prime example of a differential voltage serial network. RS-232C finds popular application in office environments where there is little electrical interference and wiring distances are short. RS-422A is more widely used in industrial applications where longer wiring distances and greater potential for electrical interference from AC power wiring exists.

However, a large part of the problem with digital network signals is the square-wave nature of such voltages, as was previously mentioned. If only we could avoid square waves all together, we could avoid many of their inherent difficulties in long, high-frequency networks. One way of doing this is to modulate a sine wave voltage signal with our digital data. "Modulation" means that magnitude of one signal has control over some aspect of another signal. Radio technology has incorporated modulation for decades now, in allowing an audio-frequency voltage signal to control either the amplitude (AM) or frequency (FM) of a much higher frequency "carrier" voltage, which is then send to the antenna for transmission. The frequency-modulation (FM) technique has found more use in digital networks than amplitude-modulation (AM), except that its referred to as Frequency Shift Keying (FSK). With simple FSK, sine waves of two distinct frequencies are used to represent the two binary states, 1 and 0:

Due to the practical problems of getting the low/high frequency sine waves to begin and end at the zero crossover points for any given combination of 0's and 1's, a variation of FSK called phase-continuous FSK is sometimes used, where the consecutive combination of a low/high frequency represents one binary state and the combination of a high/low frequency represents the other. This also makes for a situation where each bit, whether it be 0 or 1, takes exactly the same amount of time to transmit along the network:

With sine wave signal voltages, many of the problems encountered with square wave digital signals are minimized, although the circuitry required to modulate (and demodulate) the network signals is more complex and expensive.

Optical data communication

A modern alternative to sending (binary) digital information via electric voltage signals is to use optical (light) signals. Electrical signals from digital circuits (high/low voltages) may be converted into discrete optical signals (light or no light) with LEDs or solid-state lasers. Likewise, light signals can be translated back into electrical form through the use of photodiodes or phototransistors for introduction into the inputs of gate circuits.

Transmitting digital information in optical form may be done in open air, simply by aiming a laser at a photodetector at a remote distance, but interference with the beam in the form of temperature inversion layers, dust, rain, fog, and other obstructions can present significant engineering problems:

One way to avoid the problems of open-air optical data transmission is to send the light pulses down an ultra-pure glass fiber. Glass fibers will "conduct" a beam of light much as a copper wire will conduct electrons, with the advantage of completely avoiding all the associated problems of inductance, capacitance, and external interference plaguing electrical signals. Optical fibers keep the light beam contained within the fiber core by a phenomenon known as total internal reflectance.

An optical fiber is composed of two layers of ultra-pure glass, each layer made of glass with a slightly different refractive index, or capacity to "bend" light. With one type of glass concentrically layered around a central glass core, light introduced into the central core cannot escape outside the fiber, but is confined to travel within the core:

These layers of glass are very thin, the outer "cladding" typically 125 microns (1 micron = 1 millionth of a meter, or 10^-6 meter) in diameter. This thinness gives the fiber considerable flexibility. To protect the fiber from physical damage, it is usually given a thin plastic coating, placed inside of a plastic tube, wrapped with kevlar fibers for tensile strength, and given an outer sheath of plastic similar to electrical wire insulation. Like electrical wires, optical fibers are often bundled together within the same sheath to form a single cable.

Optical fibers exceed the data-handling performance of copper wire in almost every regard. They are totally immune to electromagnetic interference and have very high bandwidths. However, they are not without certain weaknesses.

One weakness of optical fiber is a phenomenon known as microbending. This is where the fiber is bend around too small of a radius, causing light to escape the inner core, through the cladding:

Not only does microbending lead to diminished signal strength due to the lost light, but it also constitutes a security weakness in that a light sensor intentionally placed on the outside of a sharp bend could intercept digital data transmitted over the fiber.

Another problem unique to optical fiber is signal distortion due to multiple light paths, or modes, having different distances over the length of the fiber. When light is emitted by a source, the photons (light particles) do not all travel the exact same path. This fact is patently obvious in any source of light not conforming to a straight beam, but is true even in devices such as lasers. If the optical fiber core is large enough in diameter, it will support multiple pathways for photons to travel, each of these pathways having a slightly different length from one end of the fiber to the other. This type of optical fiber is called multimode fiber:

A light pulse emitted by the LED taking a shorter path through the fiber will arrive at the detector sooner than light pulses taking longer paths. The result is distortion of the square-wave's rising and falling edges, called pulse stretching. This problem becomes worse as the overall fiber length is increased:

However, if the fiber core is made small enough (around 5 microns in diameter), light modes are restricted to a single pathway with one length. Fiber so designed to permit only a single mode of light is known as single-mode fiber. Because single-mode fiber escapes the problem of pulse stretching experienced in long cables, it is the fiber of choice for long-distance (several miles or more) networks. The drawback, of course, is that with only one mode of light, single-mode fibers do not conduct as as much light as multimode fibers. Over long distances, this exacerbates the need for "repeater" units to boost light power.

Network topology

If we want to connect two digital devices with a network, we would have a kind of network known as "point-to-point:"

For the sake of simplicity, the network wiring is symbolized as a single line between the two devices. In actuality, it may be a twisted pair of wires, a coaxial cable, an optical fiber, or even a seven-conductor BogusBus. Right now, we're merely focusing on the "shape" of the network, technically known as its topology.

If we want to include more devices (sometimes called nodes) on this network, we have several options of network configuration to choose from:

Many network standards dictate the type of topology which is used, while others are more versatile. Ethernet, for example, is commonly implemented in a "bus" topology but can also be implemented in a "star" or "ring" topology with the appropriate interconnecting equipment. Other networks, such as RS-232C, are almost exclusively point-to-point; and token ring (as you might have guessed) is implemented solely in a ring topology.

Different topologies have different pros and cons associated with them:

Point-to-point

Quite obviously the only choice for two nodes.

Bus

Very simple to install and maintain. Nodes can be easily added or removed with minimal wiring changes. On the other hand, the one bus network must handle all communication signals from all nodes. This is known as broadcast networking, and is analogous to a group of people talking to each other over a single telephone connection, where only one person can talk at a time (limiting data exchange rates), and everyone can hear everyone else when they talk (which can be a data security issue). Also, a break in the bus wiring can lead to nodes being isolated in groups.

Star

With devices known as "gateways" at branching points in the network, data flow can be restricted between nodes, allowing for private communication between specific groups of nodes. This addresses some of the speed and security issues of the simple bus topology. However, those branches could easily be cut off from the rest of the "star" network if one of the gateways were to fail. Can also be implemented with "switches" to connect individual nodes to a larger network on demand. Such a switched network is similar to the standard telephone system.

Ring

This topology provides the best reliability with the least amount of wiring. Since each node has two connection points to the ring, a single break in any part of the ring doesn't affect the integrity of the network. The devices, however, must be designed with this topology in mind. Also, the network must be interrupted to install or remove nodes. As with bus topology, ring networks are broadcast by nature.

As you might suspect, two or more ring topologies may be combined to give the "best of both worlds" in a particular application. Quite often, industrial networks end up in this fashion over time, simply from engineers and technicians joining multiple networks together for the benefit of plant-wide information access.

Network protocols

Aside from the issues of the physical network (signal types and voltage levels, connector pinouts, cabling, topology, etc.), there needs to be a standardized way in which communication is arbitrated between multiple nodes in a network, even if its as simple as a two-node, point-to-point system. When a node "talks" on the network, it is generating a signal on the network wiring, be it high and low DC voltage levels, some kind of modulated AC carrier wave signal, or even pulses of light in a fiber. Nodes that "listen" are simply measuring that applied signal on the network (from the transmitting node) and passively monitoring it. If two or more nodes "talk" at the same time, however, their output signals may clash (imagine two logic gates trying to apply opposite signal voltages to a single line on a bus!), corrupting the transmitted data.

The standardized method by which nodes are allowed to transmit to the bus or network wiring is called a protocol. There are many different protocols for arbitrating the use of a common network between multiple nodes, and I'll cover just a few here. However, its good to be aware of these few, and to understand why some work better for some purposes than others. Usually, a specific protocol is associated with a standardized type of network. This is merely another "layer" to the set of standards which are specified under the titles of various networks.

The International Standards Organization (ISO) has specified a general architecture of network specifications in their DIS7498 model (applicable to most any digital network). Consisting of seven "layers," this outline attempts to categorize all levels of abstraction necessary to communicate digital data.

• Level 1: Physical Specifies electrical and mechanical details of communication: wire type, connector design, signal types and levels.
• Level 2: Data link Defines formats of messages, how data is to be addressed, and error detection/correction techniques.
• Level 3: Network Establishes procedures for encapsulation of data into "packets" for transmission and reception.
• Level 4: Transport Among other things, the transport layer defines how complete data files are to be handled over a network.
• Level 5: Session Organizes data transfer in terms of beginning and end of a specific transmission. Analogous to job control on a multitasking computer operating system.
• Level 6: Presentation Includes definitions for character sets, terminal control, and graphics commands so that abstract data can be readily encoded and decoded between communicating devices.
• Level 7: Application The end-user standards for generating and/or interpreting communicated data in its final form. In other words, the actual computer programs using the communicated data.

Some established network protocols only cover one or a few of the DIS7498 levels. For example, the widely used RS-232C serial communications protocol really only addresses the first ("physical") layer of this seven-layer model. Other protocols, such as the X-windows graphical client/server system developed at MIT for distributed graphic-user-interface computer systems, cover all seven layers.

Different protocols may use the same physical layer standard. An example of this is the RS-422A and RS-485 protocols, both of which use the same differential-voltage transmitter and receiver circuitry, using the same voltage levels to denote binary 1's and 0's. On a physical level, these two communication protocols are identical. However, on a more abstract level the protocols are different: RS-422A is point-to-point only, while RS-485 supports a bus topology "multidrop" with up to 32 addressable nodes.

Perhaps the simplest type of protocol is the one where there is only one transmitter, and all the other nodes are merely receivers. Such is the case for BogusBus, where a single transmitter generates the voltage signals impressed on the network wiring, and one or more receiver units (with 5 lamps each) light up in accord with the transmitter's output. This is always the case with a simplex network: there's only one talker, and everyone else listens!

When we have multiple transmitting nodes, we must orchestrate their transmissions in such a way that they don't conflict with one another. Nodes shouldn't be allowed to talk when another node is talking, so we give each node the ability to "listen" and to refrain from talking until the network is silent. This basic approach is called Carrier Sense Multiple Access (CSMA), and there exists a few variations on this theme. Please note that CSMA is not a standardized protocol in itself, but rather a methodology that certain protocols follow.

One variation is to simply let any node begin to talk as soon as the network is silent. This is analogous to a group of people meeting at a round table: anyone has the ability to start talking, so long as they don't interrupt anyone else. As soon as the last person stops talking, the next person waiting to talk will begin. So, what happens when two or more people start talking at once? In a network, the simultaneous transmission of two or more nodes is called a collision. With CSMA/CD (CSMA/Collision Detection), the nodes that collide simply reset themselves with a random delay timer circuit, and the first one to finish its time delay tries to talk again. This is the basic protocol for the popular Ethernet network.

Another variation of CSMA is CSMA/BA (CSMA/Bitwise Arbitration), where colliding nodes refer to pre-set priority numbers which dictate which one has permission to speak first. In other words, each node has a "rank" which settles any dispute over who gets to start talking first after a collision occurs, much like a group of people where dignitaries and common citizens are mixed. If a collision occurs, the dignitary is generally allowed to speak first and the common person waits afterward.

In either of the two examples above (CSMA/CD and CSMA/BA), we assumed that any node could initiate a conversation so long as the network was silent. This is referred to as the "unsolicited" mode of communication. There is a variation called "solicited" mode for either CSMA/CD or CSMA/BA where the initial transmission is only allowed to occur when a designated master node requests (solicits) a reply. Collision detection (CD) or bitwise arbitration (BA) applies only to post-collision arbitration as multiple nodes respond to the master device's request.

An entirely different strategy for node communication is the Master/Slave protocol, where a single master device allots time slots for all the other nodes on the network to transmit, and schedules these time slots so that multiple nodes cannot collide. The master device addresses each node by name, one at a time, letting that node talk for a certain amount of time. When it is finished, the master addresses the next node, and so on, and so on.

Yet another strategy is the Token-Passing protocol, where each node gets a turn to talk (one at a time), and then grants permission for the next node to talk when its done. Permission to talk is passed around from node to node as each one hands off the "token" to the next in sequential order. The token itself is not a physical thing: it is a series of binary 1's and 0's broadcast on the network, carrying a specific address of the next node permitted to talk. Although token-passing protocol is often associated with ring-topology networks, it is not restricted to any topology in particular. And when this protocol is implemented in a ring network, the sequence of token passing does not have to follow the physical connection sequence of the ring.

Just as with topologies, multiple protocols may be joined together over different segments of a heterogeneous network, for maximum benefit. For instance, a dedicated Master/Slave network connecting instruments together on the manufacturing plant floor may be linked through a gateway device to an Ethernet network which links multiple desktop computer workstations together, one of those computer workstations acting as a gateway to link the data to an FDDI fiber network back to the plant's mainframe computer. Each network type, topology, and protocol serves different needs and applications best, but through gateway devices, they can all share the same data.

It is also possible to blend multiple protocol strategies into a new hybrid within a single network type. Such is the case for Foundation Fieldbus, which combines Master/Slave with a form of token-passing. A Link Active Scheduler (LAS) device sends scheduled "Compel Data" (CD) commands to query slave devices on the Fieldbus for time-critical information. In this regard, Fieldbus is a Master/Slave protocol. However, when there's time between CD queries, the LAS sends out "tokens" to each of the other devices on the Fieldbus, one at a time, giving them opportunity to transmit any unscheduled data. When those devices are done transmitting their information, they return the token back to the LAS. The LAS also probes for new devices on the Fieldbus with a "Probe Node" (PN) message, which is expected to produce a "Probe Response" (PR) back to the LAS. The responses of devices back to the LAS, whether by PR message or returned token, dictate their standing on a "Live List" database which the LAS maintains. Proper operation of the LAS device is absolutely critical to the functioning of the Fieldbus, so there are provisions for redundant LAS operation by assigning "Link Master" status to some of the nodes, empowering them to become alternate Link Active Schedulers if the operating LAS fails.

Other data communications protocols exist, but these are the most popular. I had the opportunity to work on an old (circa 1975) industrial control system made by Honeywell where a master device called the Highway Traffic Director, or HTD, arbitrated all network communications. What made this network interesting is that the signal sent from the HTD to all slave devices for permitting transmission was not communicated on the network wiring itself, but rather on sets of individual twisted-pair cables connecting the HTD with each slave device. Devices on the network were then divided into two categories: those nodes connected to the HTD which were allowed to initiate transmission, and those nodes not connected to the HTD which could only transmit in response to a query sent by one of the former nodes. Primitive and slow are the only fitting adjectives for this communication network scheme, but it functioned adequately for its time.

Practical considerations

A principal consideration for industrial control networks, where the monitoring and control of real-life processes must often occur quickly and at set times, is the guaranteed maximum communication time from one node to another. If you're controlling the position of a nuclear reactor coolant valve with a digital network, you need to be able to guarantee that the valve's network node will receive the proper positioning signals from the control computer at the right times. If not, very bad things could happen!

The ability for a network to guarantee data "throughput" is called determinism. A deterministic network has a guaranteed maximum time delay for data transfer from node to node, whereas a non-deterministic network does not. The preeminent example of a non-deterministic network is Ethernet, where the nodes rely on random time-delay circuits to reset and re-attempt transmission after a collision. Being that a node's transmission of data could be delayed indefinitely from a long series of re-sets and re-tries after repeated collisions, there is no guarantee that its data will ever get sent out to the network. Realistically though, the odds are so astronomically great that such a thing would happen that it is of little practical concern in a lightly-loaded network.

Another important consideration, especially for industrial control networks, is network fault tolerance: that is, how susceptible is a particular network's signaling, topology, and/or protocol to failures? We've already briefly discussed some of the issues surrounding topology, but protocol impacts reliability just as much. For example, a Master/Slave network, while being extremely deterministic (a good thing for critical controls), is entirely dependent upon the master node to keep everything going (generally a bad thing for critical controls). If the master node fails for any reason, none of the other nodes will be able to transmit any data at all, because they'll never receive their alloted time slot permissions to do so, and the whole system will fail.

A similar issue surrounds token-passing systems: what happens if the node holding the token were to fail before passing the token on to the next node? Some token-passing systems address this possibility by having a few designated nodes generate a new token if the network is silent for too long. This works fine if a node holding the token dies, but it causes problems if part of a network falls silent because a cable connection comes undone: the portion of the network that falls silent generates its own token after awhile, and you essentially are left with two smaller networks with one token that's getting passed around each of them to sustain communication. Trouble occurs, however, if that cable connection gets plugged back in: those two segmented networks are joined in to one again, and now there's two tokens being passed around one network, resulting in nodes' transmissions colliding!

There is no "perfect network" for all applications. The task of the engineer and technician is to know the application and know the operations of the network(s) available. Only then can efficient system design and maintenance become a reality.

Read More......