28
\$\begingroup\$

For someone who writes software but doesn't have a computer engineering background, the dozens of abstraction layers have always bewildered me.

I know all source code in their most fundamental level is broken down into 1s and 0s. I also know all CPUs at their most fundamental level are broken down into billions of transistors, and then logic gates & ALUs.

But this is where my knowledge gap stops:

1s and 0s -> ???? -> Transistors & Logic Gates & ALUs exist

All holistic answers filling that gap would be appreciated, especially if you can describe the process of deciding which specific transistors/units are chosen to do what work.

P.S. The answer shouldn't be a dissertation, as I realize this can go on forever. Just a succinct explanation/overview with keywords.

\$\endgroup\$
  • \$\begingroup\$ This conversation has been moved to chat. The purpose of comments is to request clarification or point out errors so that the question can be improved. They are not for extended discussion, nor are they the place to write answers. Any conclusions reached should be edited back into the question and/or any answer(s). \$\endgroup\$ – Dave Tweed Nov 4 at 16:18
24
\$\begingroup\$

Ah, you're missing the STATE MACHINE concept. That's where we can "write code" made out of TTL hardware chips: data-selectors, 4-bit counters, gangs of parallel flipflops. (But all those are the complicated parts, while the idea behind "state machines" is fairly simple.)

"State-machine" is also commonly called "micro-code." Also called "bit-slice" or "microsequencer." It's also labeled as the instruction-decoder inside the processor chip. (So, it's the "tiny person" inside the CPU-chip who reads the opcodes and actually performs the listed actions.)

In all the many intro/popular explanations of computers, they'll teach us all about logic gates, and about full blown embedded processors, but never about the Abstraction Layer that's sandwiched invisibly in between the two. They don't try to explain the tiny man in there.

The simplest state-machine is a ROM chip with its address lines connected to a many-bits digital up-counter. Then, the ROM output bytes are treated as individual wires or control-lines. (It's like a motorized washing-machine timer, stepping between N different settings in succession.) As the binary address counts up, the output-word's eight wires (or sixteen, 24, 32, etc.,) they can produce any timed pattern we want. Just pre-write the desired pattern in the ROM.

This is much like a mechanical music box. Or a controller for light-bulb patterns on 1950s advertising signs: a bunch of rotating disk-cams with leaf-switches on the edges. Carve some hills and valleys into the edge of the bakelite disks, and you can produce any timed light-patterns you desire.

But the true power of the idea will only arise if we connect our binary counter to three or four of the ROM's address-input lines. Then use the remaining extra ROM address-lines as inputs! So for example, if the ROM has 8 addr lines, we can connect our 3-bit up-counter to three of the lines. That way the counter will create a stepped sequence of eight bit-patterns on the ROM output. And, the ROM then stores thirty-two different versions of these, selected by the remaining five addr lines.

A 6502 MICROPROCESSOR, FOUND INSIDE APPLE-II AND OTHERS
enter image description here

Next, use the five ROM address-inputs ...TO SELECT A MACHINE-LANGUAGE OPCODE! Different bits placed on those five extra addr-lines will trigger the different 8-pulse sequences stored in the ROM. Each of 32 possible opcode instructions will be made of 8 steps (or fewer.) Finally, use all the ROM output-bits as control lines.

This output-line here, when high, routes two registers together into the adder, for adding two numbers. This other one is pulse-incrementing the main CPU address register, for stepping through the machine code stored in RAM. This other wire latches the Adder's output, so it can be dumped into one of the CPU registers during ADD instruction. Another one will dump some register into the main address-register, for performing JMP instruction.

In other words, the CPU itself is made of software. But it's a bit-pattern stored permanently in a few words of ROM memory. Change the bit patterns and you alter what the opcodes do. The state-machine is where the deepest level of software is physically made of hardware. Think of the state-machine as the "little man" inside the computer who reads each machine-code instruction and sends pulses to the control-lines which manipulate registers and performs each opcode's little sequence of steps. (And, the microsequencer ROM is the little man's simple brain.)

A very cool animation is the Visual 6502 chip, in javascript for browsers (6502 is the cpu used in the old APPLE-II computers.) Sample image above.

http://www.visual6502.org/JSSim/

It can actually execute a machine-code program, while you watch all the internal conductors and logic gates changing color inside the IC. The chip shows all the registers, memory-addr counters, adder, shifter, etc. But it also has a massive, random-looking patch of checkerboard along the top edge. That's the bitslice, the ROM, the permanent pattern for its opcodes all created by the running state-machine.


Steampunk: if you're going to build a Babbage thinking-engine, you'll want your rotating music-box cylinder to spin at about 20,000RPM, and the little studs on the cylinder should be made of tungsten-iridium alloy, since that cylinder contains the opcodes, in patterns of little bumps, and its immense rate of rotation determines the CPU speed. (Maybe use little silver dots on a glass cylinder, and some of those new-fangled Selenium Photocells, rather than tiny leaf-switches to read the cylinder pattern. Steam-punk optical computing!!!!!) Back during WWII the germans had 15KHz mechanical television with the line-scanning performed by a rotating quartz octagon mirror, air-levitated and spinning at something like 100,000RPM. Use one of those, and rate your computer power in terms of horsepower of air-compressor power supply.

Or this: computers are the app software, which is made of high-level language, which is made of interpreter code, which is made of assembler, which is made of machine-code opcodes... which are made of state-machine ROM hardware sequences, which are made of registers and data-selectors and counters and flipflops, which are all made of logic gates, which are made of individual transistors, which are made of impure silicon, which is made of atoms of Si with a very few of phosphorus, boron. WHICH are made of nuclei and electrons, which are made of protons and neutrons, which are made of quarks riding upon the boiling Fermi-level sea.

Or this: at the center of your computer is a mechanical laundry-timer knob. But it's spinning at over a GHz rate.

\$\endgroup\$
  • 4
    \$\begingroup\$ FSM and microcode are two different things. Microcode is no longer used in modern processors except as a fall back mechanism in certain CISC devices. Modern processors use direct hardware control, i.e. an FSM, due to its many advantages. \$\endgroup\$ – user110971 Nov 3 at 13:05
  • 1
    \$\begingroup\$ @user110971 How is that so if microcode updates are released frequently for modern processors, such as Intel x86/64 CPUs? Particularly after the Spectre fiasco. \$\endgroup\$ – Coldblackice Nov 3 at 21:09
  • 1
    \$\begingroup\$ "...which are made of quarks riding upon the boiling Fermi-level sea" This was straight poetry. Beautiful. Well done. \$\endgroup\$ – Coldblackice Nov 3 at 21:12
  • 2
    \$\begingroup\$ @Coldblackice I should have picked my words more carefully. Simple instructions, like mult and mov, are implemented using hardware control. There are certain complex instructions, like fsin (floating point sine), that are probably implemented in microcode. I say probably because Intel is not exactly open about their architecture. You are also not certain that fsin is going to be used by the compilers. In fact, on modern processor the implementation of sin using said simple instructions appears to be faster than fsin. In addition, the compiler will try to target the maximum range of CPUs. \$\endgroup\$ – user110971 Nov 3 at 21:31
  • \$\begingroup\$ @user110971 Ah, got it. Interesting info, thanks for expounding. \$\endgroup\$ – Coldblackice Nov 3 at 21:45
7
\$\begingroup\$

A processor is really a finite state machine (FSM) for implementing the machine code instructions. It reads the instructions from memory and uses the required hardware, such as the ALU, to implement them.

Here is the data path of the MIPS architecture.

MIPS

source

You have a control unit implementing said FSM and is responsible for ensuring the data is directed to the correct logic circuitry. A program counter (PC) points to the next instruction to be fetched. After the instruction is fetched, the PC is incremented so it points to the next instruction, except if there is a branching instruction, which overwrites the PC. The instruction is then implemented by the control unit.

As an example, consider the instruction to add 37 to the value stored in memory location 74. These values are provided in the opcode of the instruction. Firstly, the control unit fetches the data at address 74. Secondly, said data is supplied to one of the inputs of the ALU. Thirdly, 37 is supplied to the other input of the ALU. Fourthly, the addition operation is selected for the ALU. Finally, the result of the ALU is written back to address 74.

\$\endgroup\$
  • \$\begingroup\$ This is the real answer, imo, for regular computers. Basically CPU and GPU. \$\endgroup\$ – Andrew Nov 4 at 16:22
  • \$\begingroup\$ I would also recommend OP (or others) try doing some actual coding in assembly. Utilizing registers and labels and jump (JMP) commands and such will give one a better idea of CPU's. Writing GPU shader code will give one a better idea of how GPU's work. \$\endgroup\$ – Andrew Nov 4 at 16:27
3
\$\begingroup\$

Transistors are used to build logic gates. Gates are used to build logic circuits and memories. A modern CPU is, loosely speaking, built as two parts (1) a datapath that does maths and loads and stores values in memory (also made of transistors), and (2) control circuits that configure that datapath based on machine instructions (which are the 1s and 0s the OP alludes to in the question). That fills the knowledge gap without requiring a complete course in computer architecture.

\$\endgroup\$
3
\$\begingroup\$

Yes, machine code is 0's and 1's. The step from here to electrical signals is small; each 0 may be 0 volt and 1 may be 1.8 volt (or the opposite!) They are so similar that no "translation" is needed here.

In a 32-bit architecture the instruction bus is 32 bits wide; in other words there are 32 individual lines, each with one of these voltage levels. The magic comes from the clock. Every time the clock flips its line, this activates basic circuits like latches (https://en.wikipedia.org/wiki/Flip-flop_(electronics)) to act on the signal levels present on the bus. The transistors will take a certain time to act and the signal needs time to propagate, on the scale of a nanosecond. The clock must run slow enough for all signals to stabilize at their new value before flipping again.

The transistors are used to build up logical elements like NOR gates (https://en.wikipedia.org/wiki/NOR_gate), adders (https://en.wikipedia.org/wiki/Adder_(electronics)), etc, which form a fetch-execute sequence in the CPU https://en.wikipedia.org/wiki/Instruction_cycle

From there the abstraction continues through the intricacies of modern digital integrated circuits towards software. It's an amazing achievement.

\$\endgroup\$
3
\$\begingroup\$

the souce code is translated into machine code using a compiler or an assembler (or some combination)

this machine code is a different pattern of ones and zeroes

In the tha CPU the the instriction decoder marshalls the machine code into actions takes by the ALU registers, i/o etc. The instriction decoder it typically a bunch of gates.

\$\endgroup\$
2
\$\begingroup\$

Here's something that the other answers haven't touched on yet... there's also something called synthesis which is one way to code the actual transistors and their physical layouts. You write the design in a Hardware Description Language (usually VHDL or Verilog). Then either you configure a Field Programmable Gate Array (FPGA) or you have actual hardware built where they map your code into transistors, known as an Application-Specific Integrated Circuit (ASICs). This is how the CPUs and GPUs themselves are created.

This isn't exactly what you asked, but might interest you!

(Note, for the pedantic, there are other ways to configure silicon, e.g. schematic capture, but that doesn't answer the question of "code to transistors."

\$\endgroup\$
2
\$\begingroup\$

As others have already said, the key is probably to understand what a logic gate is, and how transistors are used to build one. Logic gates are the fundamental building blocks of anything digital. A logic gate simply takes one, two, or occasionally more binary input values, and produces an output according to a predefined rule (called a truth table). The most common one is called NAND (or not-and), and I use it as an example. The rule for a AND is that it takes takes both inputs, and only produces a 1 output if both input 1 AND input 2 are 1. A NAND inverts that: it only produces a 0 if both inputs are 1. The common chip used is called 7400. You can actually still buy it today, in only slightly modified form from what was used in the Apple II.

NAND operation

Expressed as a truth table:

0 NAND 0 = 1
0 NAND 1 = 1
1 NAND 0 = 1
1 NAND 1 = 0

Wikipedia also includes a schematic down to several versions of how it is implemented at the transistor level. https://en.wikipedia.org/wiki/NAND_gate

There are various other types of gates, with different truth tables.

Now the next thing is how you get from that to the computer. For that, it may help to ignore today's CPUs with the billions of transistors, and look at the far simpler older computers, such as the Apple II. Since that computer is now more than 40 years old, a lot of that old information is freely available - for instance, here: https://mirrors.apple2.org.za/Apple%20II%20Documentation%20Project/Books/W.%20Gayler%20-%20The%20Apple%20II%20Circuit%20Description.pdf

The Apple II (and the original IBM PC) were actually built from mostly individual gates (except for the CPU itself). There were also a few timer chips (Apple famously used the 555 timer for all kinds of fascinating stuff). The 555 is not a logic gate, but it is similarly simple in construction.

And of course there was memory, and a few other specialty items.

So with that information, you can actually follow the electrical signals as they travel through your computer (either mentally, or with something like an oscilloscope or logic analyzer).

That leaves the CPU itself as a black box. Fortunately, the 6502 (used in Apple II, as well as Commodore, Amiga and other computers of that era) is again very simple. Somebody went to the trouble of building one from individual transistors (and 7000 times the size of the integrated circuit): https://monster6502.com/ According to that site, the 6502 contains only roughly 4200 transistors. And with this replica, you can follow, at the transistor level, exactly what happens when you try to execute a machine-language instruction such as "ASL" (arithmetic shift left).

Another site you may take a look at is the virtual 6502. It doesn't quite go down to the transistor level, though. You still have to understand that the registers are made up of gates (probably flip-flops with about 6 transistors per bit), and that the arithmetic-logic-unit (ALU) is also made up of gates.

\$\endgroup\$
1
\$\begingroup\$

For single cycle instructions, it’s basically lookup tables, from instruction machine code as the inputs, to lookup table outputs that are all the enables (etc.) needed to select and configure the various chunks of logic (registers, selectors, ALUs, buffers, etc.) required to accomplish the assigned task of the instruction. For multi-cycle instructions, one of the lookup table outputs is the next state for a state machine to step through the required steps. For pipelined implementations, a mix of the two.

As an optimization, the lookup table is reduced to an smaller functionally equivalent pile of logic gates, often called the instruction (or micro-instruction) decoder.

The 1’s and 0’s are already output from memory (or memory controller or subsystem), and/or registers, as high and low voltage levels suitable for driving inputs to a bunch of transistor logic gates, such as an instruction decoder.

\$\endgroup\$
0
\$\begingroup\$

Short answer in addition to mbeaty and others:

There is actually NO other abstraction layer. 1's and 0's are themselves open or closed transistors and high/low voltages somewhere in the computer.

On every clock cycle (itself a transistor transition between open and closed) the next memory cell (8 or 16 or 32 transistors in the RAM) is connected (by some more transistors) to the transistors in the CPU that "decode" the instruction - switch some more transistors in a way to transform the input of the instruction into the result (simplified, of course).

\$\endgroup\$
  • 1
    \$\begingroup\$ This isn't to say that an additional abstraction layer or two are not useful for one's understanding. Many modern CPUs include microcode, and it's certainly correct to interpret the filling of microcode buffers on a transistor level. However, it's also helpful to consider microcode as an additional form of instructions beneath assembly, with some abstract circuit performing microcode decoding. \$\endgroup\$ – Reinstate Monica - ζ-- Nov 4 at 0:38

Not the answer you're looking for? Browse other questions tagged or ask your own question.