Of course they are, Broadwell-E has 5-10% higher IPC than Haswell-E.
I'm pretty sure we've come to the conclusion at this point that the 25 second press demo was at 100 samples, and the livestreamed 36 second New Horizons demo was at 150.
Of course they are, Broadwell-E has 5-10% higher IPC than Haswell-E.I ran my 5960x at 3.375 core/ 3.375 cache and 3.5 core/ 3.5 cache at 150 samples for three runs each.
This is a brand new Windows 10 install with zero background programs. I even waited for all background processes to stop so CPU usage was at 0%.
3.375: 38.81 / 38.15 / 38.15
3.500: 37.41 / 36.68 / 36.61
This still doesn't quite seem right compared to AMD's results, especially given my cache is technically overclocked at both speeds and I have fast memory (3000 MHz).
If the 6900k was running at 3.5 during the test, it sorta seems reasonable it might score 0.8 seconds faster. But this would make it seem AMD is claiming to have higher IPC than even Haswell, at least in this test.
I meant the scores don't correspond to what AMD_james said, then again he contradicted himself with the spreadsheet (unconfirmed) scores.I don't see any other 10 core 6950X CPUs in that list. I ran the benchmark a few more times. The lowest score I get is 24.48 seconds at 150 samples.
It seems pretty cut and dry from the demos that AMD is aiming for/claiming (in a vague sense, as no claims are very specific) IPC parity with BDW-E. Matching/beating a 3,2-3,7GHz chip with a 3,4GHz chip pretty much makes that a given. Unless the i7 was significantly thermally limited (which seems very unlikely, given that it would be an extremely unpredictable way of messing up a demo), it would run at maximum all-core turbo (although I don't know the exact frequency for that, I'd wager 3,4-3,5GHz).Of course they are, Broadwell-E has 5-10% higher IPC than Haswell-E.
I'm pretty sure we've come to the conclusion at this point that the 25 second press demo was at 100 samples, and the livestreamed 36 second New Horizons demo was at 150.
Maybe it was 1m and 04s = 64s? Happens to me sometimes.So, I did some math (check it) after changing my clockspeeds to 3.4 across all cores and reran the test at 150 samples.
Zen's 8 cores@3.4 = 36 seconds
36/8 = 4.5 seconds of work per core
Remove the work of two cores and you get 45 seconds for a hypothetical 6 core zen
Sandy's 6 cores@3.4 = 104 seconds
45s vs 104s = Zen is 57% faster per core than Sandy? I know I MUST have screwed something way the hell up. So what was it?
Maybe it was 1m and 04s = 64s? Happens to me sometimes.
It's not how it works. because there's SMT involved.
if you get same improvements 6c vs 6c when smt is disabled on both. then maybe yes 57% improvement for zen on single core.
unfortunately we don't have results without smt. so we can't really compare it like that.
Of course they are, Broadwell-E has 5-10% higher IPC than Haswell-E.
I'm pretty sure we've come to the conclusion at this point that the 25 second press demo was at 100 samples, and the livestreamed 36 second New Horizons demo was at 150.
Blender, just like about every real life load, does not scale linearly with core/thread count. So you need a result for 2670 to make any meaningful comparisons. And even then, that's a stretch.So, I did some math (check it) after changing my clockspeeds to 3.4 across all cores and reran the test at 150 samples.
Zen's 8 cores@3.4 = 36 seconds
36/8 = 4.5 seconds of work per core
Remove the work of two cores and you get 45 seconds for a hypothetical 6 core zen
Sandy's 6 cores@3.4 = 64 seconds
45s vs 64s = Zen is 30% faster per core than Sandy in this test at the same clocks. If Zen is consistently 30% faster than Sandy, then Zen will be a knockout and I might actually buy one!
Yes that's what I meant, except the 36s score is supposed to be with 200 samples (according to the spreadsheet linked by AMD_james) & then there is a new file with samples prefixed at 150?
www.reddit.com/r/Amd/comments/5idz87/ryzen_blender_benchmark_comparison_spreadsheet/
I've seen a number of blender results, IIRC in this very thread, scale linearly with core count & clock speeds so extrapolating the score from 100 -> 150 samples is fairly easy. The spreadsheet, in the reddit thread, however shows the SR7 score as 36s (unconfirmed according to AMD_james) with 200 samples & then he goes on to say that they'll upload a new file with samples set to 150.It is not that hard to figure out.
We can solve the mystery easily by using AdamK47 result above with his 6950x.
He scores 16.95s at 100 samples and 24.92 at 150 samples using a 10C/20T Broadwell-E at 4.12Ghz.
Since blender scales great with cores and clocks, 8C/16T broadwell-e running at multicore turbo boost clock of 3.5Ghz(I assume that chip AMD used managed to boost to this clock during tests) should score:
100 sample run -> 16.95s x 10/8 x 4.12/3.5=24.9s
150 sample run -> 24.92s x 10/8 x 4.12/3.5=36.66s
Scores above are perfectly in line with what AMD showed as both runs were completed within a fraction of a second from the projected run times I got in the formulas above.
I agree that there was unnecessary confusion but it is obvious now that the scores they got are in line with reality and nothing was "rigged".I've seen a number of blender results, IIRC in this very thread, scale linearly with core count & clock speeds so extrapolating the score from 100 -> 150 samples is fairly easy. The spreadsheet, in the reddit thread, however shows the SR7 score as 36s (unconfirmed according to AMD_james) with 200 samples & then he goes on to say that they'll upload a new file with samples set to 150.
At this point in time I do believe AMD's scores but there shouldn't have been this level of confusion, especially coming from an AMD employee.
If that proportionate relationship holds we are talking about a chip that does about 19.5 on Cinebench 11.5. Doggone impressive if you ask me.
If there's a thing that scales very well with cores, it's rendering, and blender likely is no exception.Blender, just like about every real life load, does not scale linearly with core/thread count. So you need a result for 2670 to make any meaningful comparisons. And even then, that's a stretch.
I think there is basically no difference in results between those two versions, somebody has already tested it in this thread (fraction of a point worth of difference). The sample size is the key difference between the live stream run and the press demo run.I think the Sample sizes were chosen to keep the benchmark to a reasonable length for the given presentation.. To match the dialog and not keep people waiting too long.
I still believe 2.77 was used for the press demo and maybe 2.78a for the other as claimed , not sure why the difference.
I Queried James / Robert in that Reddit thread , but they insist it was 2.78a for both.. I'm not convinced. Even though it's hard to read whether it definatly says 2.77 in the video, there's definitely no 'a' on the end of the version No.
I saw the pics on sweclockers as well & it's definitely version 2.77 rendering the logo at least once, I guess we'll just have to wait for AMD's official word on this.I think the Sample sizes were chosen to keep the benchmark to a reasonable length for the given presentation.. To match the dialog and not keep people waiting too long.
I still believe 2.77 was used for the press demo and maybe 2.78a for the other as claimed , not sure why the difference.
I Queried James / Robert in that Reddit thread , but they insist it was 2.78a for both.. I'm not convinced. Even though it's hard to read whether it definatly says 2.77 in the video, there's definitely no 'a' on the end of the version No.
Yes not much between them, but again unnecessary confusion it seems.I think there is basically no difference in results between those two versions, somebody has already tested it in this thread (fraction of a point worth of difference). The sample size is the key difference between the live stream run and the press demo run.