AMD FX-8350 Linux Performance-Per-Watt

[curaga]

Very curious results, having the older cpu beat the newer one in perf/W in a number of scenarios. And also having lower idle and max power use!

That's a different result from Windows reviews, which is why it's good we have Phoronix around.

[schmalzler]

All those recent tests on FX-8350 had mtune=generic (at least that's what has been written under the config table). So they don't show the real potential (if any). No AVX, no AVX2, no AES, ...
All those gathered numbers don't have any value...

[mcgreg]

Average watt usage of fx 8150 @ 4,6ghz is 297,8
time to compile is 79,25
297,8 / 79,25 = 3,76 - which is correct

Average watt usage of the fx 8350 @ 4,6ghz is 264,3
time to compile is 72,98
264,3 / 72,98 = 3,62

-> the calculation is correct but I dont see any sense in it.
shoulnd't it be more like:

fx 8150@4,6ghz: 79,25 / 297,8 = 0,266
fx 8350@4,6ghz: 72,98 / 264,3 = 0,276

=> the fx 8350@4,6ghz gives 3,8% more performance per watt compared to the fx 8150@4,6ghz


also I'd calculate
297,8 * 79,25 = 23601 kJ for the fx 8150 @ 4,6ghz
264,3 * 72,98 = 19288 kJ for the fx 8350 @ 4,6 ghz

19288 / 23601 = 0,817 .. means that the fx 8350 almost needs 18% less power than the fx 8150 @ 4,6ghz - for compiling the kernel - which is definitely a better job.

correct me if I am wrong of course



mcgreg

[curaga]

-> the calculation is correct but I dont see any sense in it.
shoulnd't it be more like:

fx 8150@4,6ghz: 79,25 / 297,8 = 0,266
fx 8350@4,6ghz: 72,98 / 264,3 = 0,276

=> the fx 8350@4,6ghz gives 3,8% more performance per watt compared to the fx 8150@4,6ghz

I think you got that wrong - that's seconds per watt, and the 8350 is taking _longer_ per watt. This would mean it's worse, not better?

[droste]

I think you got that wrong - that's seconds per watt, and the 8350 is taking _longer_ per watt. This would mean it's worse, not better?

No, it means that if the 8350 would only get 1 watt it can "compute" for 0,276s and the 8150 would only "compute" for 0,266s. So more would be better. But this number alone says nothing because longer "computing" does not generally mean it does more things.

/edit:
The only meaningful number is indeed this:

297,8 * 79,25 = 23601 kJ for the fx 8150 @ 4,6ghz
264,3 * 72,98 = 19288 kJ for the fx 8350 @ 4,6 ghz

[mcgreg]

No, it means that if the 8350 would only get 1 watt it can "compute" for 0,276s and the 8150 would only "compute" for 0,266s. So more would be better. But this number alone says nothing because longer "computing" does not generally mean it does more things.

Well, in that case it actually does. Because

The info says, you need 264,3s/297,8s to finish compiling the kernel.
fx 8350@4,6ghz 0,276s / 264,3s = 0,104 % of the kernel finished = 1 watt
fx 8150@4,6ghz 0,266 / 297,8s = 0,089 % of the kernel finished = 1 watt

correct me if I'm wrong

[bug77]

Here's another interesting review: performance per clock. http://hardocp.com/article/2012/10/2...c_overclocking

[droste]

Well yes, if calculate the number by doing what is done here (full run of the same workload) it can be compared.

but it's
8350@4,6ghz 0,276s / 72,89s = 0,379 % of the kernel finished = 1 watt
8150@4,6ghz 0,266s / 79,25s = 0,336 % of the kernel finished = 1 watt

If you calculate this by actually give the CPU only 1 watt and say compile the kernel and see what is done, you can not be sure that the same things were computed in this time, because they may start compiling different files. Not that someone would do this :-D. But still, context is needed (as in how did you come up with the number).

[mcgreg]

Ahh yes, I used the wrong numbers, and you corrected it, thx.

So, the results is clear: the FX 8350 is more efficient in performance / watt and michaels results is wrong then - this is actually what most people were already thinking, we just proved it.

[abacaba]

also I'd calculate
297,8 * 79,25 = 23601 kJ for the fx 8150 @ 4,6ghz
264,3 * 72,98 = 19288 kJ for the fx 8350 @ 4,6 ghz

mcgreg

Correct, except "J", not "kJ".

Or, one can get inverse: 1/23601 = 0.000042371, which means ~42 microlinuxkernels (are compiled) per joule. Here, the more the number is the better.