Intel vs. AMD Performance-Per-Watt On Ubuntu 14.04 Linux

[vick]

Like this:

cpu	i3 2120	i5 3470	i7 3770k	i3 4130	i5 4670	i7 4770k	A10 7850k
P_avg[W]	63,7	76,4	96,8	75,6	91,1	113,8	116,3
time[s]	202,1	118,8	88,92	161,96	94,03	75,43	163,14
Energy[J]	12874	9076	8607	12244	8566	8584	18973

That's total energy which isn't what this article is testing exactly. Increasing time and lowering power would give the same total energy but would you actually prefer that? If so, then there are ARM and Atom chips that might give the same total energy.

I want 1/time/energy which could also be converted to 1/time/(average power) to get performance per Watt.

[jwilliams]

That's total energy which isn't what this article is testing exactly. Increasing time and lowering power would give the same total energy but would you actually prefer that? If so, then there are ARM and Atom chips that might give the same total energy.

Yes, total energy required to perform a computation is a reasonable thing to be concerned about if your computer can be powered off or powered down (i.e., sleep or hibernation) into a low power state after the computation is done. And I doubt there are computers with an ARM CPU that can do the same computation for less total energy than a computer with an i7 haswell CPU (remember that it is not just the CPU that requires power during the computation).

[jwilliams]

I want 1/time/energy which could also be converted to 1/time/(average power) to get performance per Watt.

No. energy is not equal to average power. Power is equal to energy per unit time.

Also, one way of measuring performance is in operations per second. Watts are equal to Joules per second. So we can write:

Performance per Watt = ( operations / second ) / ( Joule / second ) = operations / Joule

Note that by operations I am talking about a generalized, high-level kind of operation, not a CPU instruction. So the number of operations is the same for a given compilation, independent of the CPU. So we can ignore "operations" in the above formula and only look at the energy usage in Joules. Which is just what tuke81 showed us.

Again, good job tuke81 !

[bridgman]

It would be really interesting to run those tests with Kaveri at 45W and 65W, says the AMD guy

If the graphics performance scales comparably to Windows benchmarks then a back-of-envelope calculation suggests that the FPS/W numbers on Tropics would end up around 0.17 (15.7 fps / 94.8 w) at 65W and 0.19 (13.9 fps / 74.8W) at 45W, but I haven't actually run the tests and my 2-core low clock 1-stick-of-RAM Kaveri engineering box would not be a particularly good test bed.

The interesting thing is that in the lower power settings the biggest hit from 95/100W to 45W is only about 25% (GPU shader & texturing), CPU hit is maybe 17%, and AFAICS the ROP-limited GPU functions don't drop at all because the # of CUs is reduced from 8 to 6 but the # ROPs and GPU clock seem to be unchanged.

[vick]

Yes, total energy required to perform a computation is a reasonable thing to be concerned about if your computer can be powered off or powered down (i.e., sleep or hibernation) into a low power state after the computation is done. And I doubt there are computers with an ARM CPU that can do the same computation for less total energy than a computer with an i7 haswell CPU (remember that it is not just the CPU that requires power during the computation).

I have no numbers about my ARM/Atom statement. It was just to emphasize that total energy isn't the whole picture. I concede that knowing what the work load is related to that total energy is useful. Maybe Michael can run ARM and Atom tests to see how much energy is used for the same workload.

No. energy is not equal to average power. Power is equal to energy per unit time.

Also, one way of measuring performance is in operations per second. Watts are equal to Joules per second. So we can write:

Performance per Watt = ( operations / second ) / ( Joule / second ) = operations / Joule

Note that by operations I am talking about a generalized, high-level kind of operation, not a CPU instruction. So the number of operations is the same for a given compilation, independent of the CPU. So we can ignore "operations" in the above formula and only look at the energy usage in Joules. Which is just what tuke81 showed us.

Again, good job tuke81 !

That's why I said converted to get a performance per Watt number. However, I can see how my wording can be confused to be seen as equal.

I still prefer performance/Watt or operations/Joule over total energy per work load.

[Tgui]

Like this:

cpu	i3 2120	i5 3470	i7 3770k	i3 4130	i5 4670	i7 4770k	A10 7850k
P_avg[W]	63,7	76,4	96,8	75,6	91,1	113,8	116,3
time[s]	202,1	118,8	88,92	161,96	94,03	75,43	163,14
Energy[J]	12874	9076	8607	12244	8566	8584	18973

This is very well done. Thank you. I prefer total energy usage as you provided.

[vick]

That's why I said converted to get a performance per Watt number. However, I can see how my wording can be confused to be seen as equal.

I still prefer performance/Watt or operations/Joule over total energy per work load.

I'd like to say that I prefer performance/Watt or operations/Joule from an active consumer use point of view. I can see energy/workload can be useful for servers and automatic background tasks but when it comes to consumer use like browsing the internet, playing a game, writing an office document, etc., the performance/Watt or operations/Joule are a better metric for a sense of efficiency. At least for me who does more regular Joe Schmo tasks.

[jwilliams]

I'd like to say that I prefer performance/Watt or operations/Joule from an active consumer use point of view. I can see energy/workload can be useful for servers and automatic background tasks but when it comes to consumer use like browsing the internet, playing a game, writing an office document, etc., the performance/Watt or operations/Joule are a better metric for a sense of efficiency. At least for me who does more regular Joe Schmo tasks.

I already showed that performance per Watt and operations per Joule are the same as just measuring the total energy (for a given computational task). I can make no sense of what you wrote above. It sounds like you are confused. Note that Joules are a unit of energy, and operations and computational workload are the same thing.

[vick]

I already showed that performance per Watt and operations per Joule are the same as just measuring the total energy (for a given computational task). I can make no sense of what you wrote above. It sounds like you are confused. Note that Joules are a unit of energy, and operations and computational workload are the same thing.

Maybe my English isn't very good. I'm sorry. I was using those terms like they're equal. Should I have said performance/Watt (operations/Joule) or performance per Watt / operations per Joule (this looks more confusing to me though)?

I still say that for some cases energy/work load can mislead people if the energy saved was at a cost of longer than desired time or performance.

Do you have numbers for energy/work load on recent Atom chips? Maybe there's a Phoronix article somewhere.

[jwilliams]

INote that Joules are a unit of energy, and operations and computational workload are the same thing.

I should clarify that "workload" is not a well-defined term. Depending on interpretation, it could mean a fixed amount of work (i.e., 1000 lines of code that perform some operation), or it could be an amount work per unit time (i.e., operations per second). I suggest you not use that word if you want to improve your understanding of Performance per Watt.