Distance between memory and processor

[Chemp_93]

Homework Statement

This problem is from Spacetime Physics by Taylor

In one second some desktop computers can carry out one million instructions in sequence. Assume that carrying out one of the instruction requires transmission of data from the memory to the processor and transmission of the result back to the memory for storage.

a) What is the maximum average distance between memory and processor in a "one-megaflop" computer? Is this maximum distance increased or decresed if the signal travels through conductors at one half the speed of light in a vacuum?

Homework Equations

1 mega-flop = 1 sec / 10⁶ instructions

c = 3 * 10⁸ m/s

The Attempt at a Solution

For the first question asked;
With all that is given, I was not sure how to go about the problem so I just did some dimensional analysis to get a unit of meters.

C * 1 sec/ 10^6 instructions ≈ 300 m/ instruction of light travel time.

This turns out to be incorrect, my TA said that I must take into account that light takes a round trip.
My guess would be to multiply by 2, but doesn't the units of m/instruction take account for the round trip (since it is 'per' instruction')?

Very little is given so I am almost positive that the calculations should not be mathematically complicated.

Can someone tell me what is wrong with my logic?

 

[jedishrfu]

one million instructions per second means one instruction takes 1 microsecond that means that 0.5us to get data and 0.5us to put it back. How far does light travel in 0.5us?

 

[Chemp_93]

0.5 μs * c = 150 meters traveled.

But that makes no sense because the distance between memory and processor can't be 150 meters apart

 

[Mark44]

Homework Statement

This problem is from Spacetime Physics by Taylor

In one second some desktop computers can carry out one million instructions in sequence. Assume that carrying out one of the instruction requires transmission of data from the memory to the processor and transmission of the result back to the memory for storage.

a) What is the maximum average distance between memory and processor in a "one-megaflop" computer? Is this maximum distance increased or decresed if the signal travels through conductors at one half the speed of light in a vacuum?

Homework Equations

1 mega-flop = 1 sec / 10⁶ instructions

The units on the right are upside-down.
1 megaflop = 10⁶ instructions/second

BTW, the "flop" part means "floating-point" instructions, which are more processor intensive than integer instructions.

c = 3 * 10⁸ m/s

The Attempt at a Solution

For the first question asked;
With all that is given, I was not sure how to go about the problem so I just did some dimensional analysis to get a unit of meters.

C * 1 sec/ 10^6 instructions ≈ 300 m/ instruction of light travel time.

This turns out to be incorrect, my TA said that I must take into account that light takes a round trip.
My guess would be to multiply by 2, but doesn't the units of m/instruction take account for the round trip (since it is 'per' instruction')?

Very little is given so I am almost positive that the calculations should not be mathematically complicated.

Can someone tell me what is wrong with my logic?

 

[Chemp_93]

The units on the right are upside-down.
1 megaflop = 10⁶ instructions/second

BTW, the "flop" part means "floating-point" instructions, which are more processor intensive than integer instructions.

Okay... that doesn't really help though.

 

[CWatters]

150m is the maximum distance ignoring any time required by the memory or processor to do their thing.

 

[Chemp_93]

150m is the maximum distance ignoring any time required by the memory or processor to do their thing.

I figured that 150 m is more likely the solution. So if the memory and processor took more time 'to do their thing', wouldn't the distance between them increase?

 

[CWatters]

Think about it a bit more...

The system has 1uS to get the data from memory to the CPU and back to the memory. If the memory and CPU "waste" some of that time there will less time will be available to send the data over the wires. Therefore the maximum distance they can be apart will ...?

 

[Chemp_93]

Thanks CWatters I see what you're saying.

But, the math isn't telling me the same.
If L = the length between memory and processor then..

2L = C * t(time it takes for each instruction)

L = C*t/ 2

If I increase the time for each instruction (t) wouldn't that increase L?

 

[phinds]

If I increase the time for each instruction (t) wouldn't that increase L?

It would increase the permissible MAXIMUM length. There is no requirement that computer manufacturers put the memory in the next room, so that just means that as far as the actual length used, the computer could run faster. The slower you make the instruction, the faster it COULD run and the farther apart the memory and CPU COULD be.

 

[Chemp_93]

Alright that makes more sense. Thanks for the help.

 

[thegreenlaser]

0.5 μs * c = 150 meters traveled.

But that makes no sense because the distance between memory and processor can't be 150 meters apart

150 m is indeed huge compared to the size of a typical computer today, but remember that 1 million (mega) operations per second is also REALLY slow compared to a typical computer today. For example, my PC has a 3.47 GHz processor, which means it's doing up to 3.47 billion operations per second. That turns into around a 4 cm maximum limit on the distance that signals can travel. Add in all the other delays, and suddenly you may have to worry quite a lot about speed-of-light propagation delays.

Of course, as you can imagine, dealing with signal delays in a computer is MUCH more complicated than your question makes it out to be. But it's interesting to note that we actually have sort of hit a "maximum frequency" (at around 2-4 GHz) in the last decade or so, and it's become difficult to push past that. I think it's mostly due to heat dissipation issues at high frequency, but speed of light delay might very well play a part in that. The tendency now is to add multiple processors running in parallel ("dual core," "quad core," etc.) rather than boosting the frequency.

 

[Chemp_93]

150 m is indeed huge compared to the size of a typical computer today, but remember that 1 million (mega) operations per second is also REALLY slow compared to a typical computer today. For example, my PC has a 3.47 GHz processor, which means it's doing up to 3.47 billion operations per second. That turns into around a 4 cm maximum limit on the distance that signals can travel. Add in all the other delays, and suddenly you may have to worry quite a lot about speed-of-light propagation delays.

Of course, as you can imagine, dealing with signal delays in a computer is MUCH more complicated than your question makes it out to be. But it's interesting to note that we actually have sort of hit a "maximum frequency" (at around 2-4 GHz) in the last decade or so, and it's become difficult to push past that. I think it's mostly due to heat dissipation issues at high frequency, but speed of light delay might very well play a part in that. The tendency now is to add multiple processors running in parallel ("dual core," "quad core," etc.) rather than boosting the frequency.

What exactly is speed of light propagation delay?

Also, are there limits on physical size of these parallel processors that the speed of light imposes?

 

[D H]

What exactly is speed of light propagation delay?

It's what you've been calculating.

Being pessimistic, here's what could happen inside a poorly designed computer:

1. The CPU starts processing an operation and determines it needs an item from memory.
2. The CPU sends a fetch command to memory to retrieve those items.
3. The fetch command travels at a speed that is at most the speed of light from the CPU to memory.
4. The memory module receives the fetch command,
5. The memory module looks up the requested items.
6. The memory module sends the requested items to the CPU.
7. The retrieved data travels at a speed that is at most the speed of light from memory to the CPU.
8. The CPU computes the result of the operation.
9. The CPU sends the result and the address in which to store the result to memory.
10. The store command travels at a speed that is at most the speed of light from the CPU to memory.
11. The memory module receives the store command.
12. The memory module processes the store command.
13. The memory module sends an acknowledgment to the CPU.
14. The acknowledgment travels at a speed that is at most the speed of light from memory to the CPU.
15. The CPU receives the acknowledgment.
16. The CPU starts processing the next operation (repeat step #1).

In this pessimistic architecture, the speed of light propagation delay rears its ugly head four different times, highlighted in bold. This is pessimistic due to items 10 to 15. The CPU can bypass those last six steps and start with the next command if we can assume that the memory module will store the data as requested. That still leaves two propagation delays. Note well: Omitting those steps might not be a good idea if a computer has multiple CPUs.

The calculation in Taylor only looks at steps 3 and 7. It ignores that the other steps take time, too, and it ignores that the there is a time delay in the signal itself. It implicitly assumes a zero width signal. The amount of information that can be transmitted in a zero width signal is of course zero. The calculation in Taylor gives an upper limit on the distance between CPU and memory.

After taking all those other considerations into account, you'll find that a modern computer is physically impossible. (Steps 5 and 12 in particular are extremely slow.) There is no way to achieve the speeds a modern computer attains with that pessimistic architecture, or even with the optimistic architecture that results from bypassing steps 10 to 15.

Modern computers attain their ridiculously high operating speeds by not going to memory if at all possible. They keep a local copy (cache) of a parts of memory on the CPU, and only go to memory when that local cache does not contain the needed information. Even that isn't good enough. Modern computers have a hierarchy of memory caches.