Post your Ryzen Blender Demo Scores! (AMD clarifies Blender Benchmark Confusion, Run @ 150 Samples)

[Excessi0n]

i7-6700K @ 4.9GHz core, 4.7GHz cache, DDR4-3200 CL16:

100 Samples: 30.87 seconds
150 Samples: 46.23 seconds
200 Samples: 61.49 seconds

 

[ctk1981]

I7-5775c @4ghz

100 samples: 41.83

150 samples: 1.02.50


Sent from my iPhone using Tapatalk

 

[bbhaag]

Here are the results from my i7-6700k oc to 4.4 ram is clocked at 3000 cl15

100 samples 34.75

150 samples 51.52

 

[chuy409]

1 min 06 sec on 100 scale.
1 min 39 sec at 150 scale
Using a Phenom 960t unlocked to x6@ 4.1ghz w/ 1777mhz ddr3.

Is that good?

 

[siriq]

Here are my results with stock Blender, for reference:

Skylake 6700K, 2.78a

100 samples: 0:36.0
150 samples: 0:53.7
200 samples: 1:11.7

Your FX result is significantly better at 4.8 than mine was at 5. I guess I'll need to retest at 4.8. The 5 GHz testing wasn't on a tested-as-Prime-stable clock because my board's VRMs run too hot for Prime testing at that level of voltage so it's possible that the scores are lower than they should be because of less than optimal voltage. It should be able to handle 4.8 much easier. I can run Prime at 4.7 on the most strenuous settings.

update.... It does look like the 5 GHz results are bogged down by inadequate voltage and/or excessive VRM heat. My guess is that the CPU NB voltage was too low. I managed to get a much better CB multi score when I bumped that up when I did 4.6 testing. Here are the results at 4.8:

4.8 GHz

100: 1:09.5
150: 144:5

4.8 GHz “SIMD”

100: 36.8
150: 55

CB 15 multi: 767 (same score, two runs)

CPU-Z 1.78

single: 1414
multi: 9212

If someone with a high-grade board could test at 5 GHz that would be good.

According to SIMD scores, FX really holding up. I am actually happy for that

My CB 15 multi is 758

Now, just want to see the ZEN SIMD score.

Side note: i have rendered an animation , which took around 5-6 hours with SIMD build. I recommend to use this version to everyone!
Here is the animation of the beginning of the video: https://www.youtube.com/watch?v=iLnu5GObf2g
The normal 720p takes only 30 min or so.

 

[siriq]

1 min 06 sec on 100 scale.
1 min 39 sec at 150 scale
Using a Phenom 960t unlocked to x6@ 4.1ghz w/ 1777mhz ddr3.

Is that good?

Very good result.

 

[rvborgh]

Rather interesting... it seems with improved LLano K10 core it gives 99% the performance of K10 (on this benchmark) per core at same frequency as K10 but without L3 cache. i really wonder how LLano would have done with L3 (39037 vs 39500).

Just finished up testing on my A8 3850 based HTPC (W7 12 GB RAM) and saw 3:32. Not a bad showing for old Llano, 18.75% lower clocks 19.12% slower render. Basically lines up with the core being tweaked just enough to reach equal performance per clock even without the L3 cache.

 

[rvborgh]

here's a relative Blender Ryzen benchmark (150 samples) i put together from a whole bunch of sources.

#s on left are samples per second per core per GHz (so you can compare processors by core throughput)

As you can see... Ryzen does rather well

19,815 25:14 min : AMD E-350 „Zacate” (2C/2T, 1.6 GHz, kein Turbo, DDR3-1066 SC) von deppjones
20,222 08:36 min : AMD A8-4500M „Trinity” (2C/4T, 2.3 GHz, kein Turbo, DDR3-1333 DC) von Fab8
21,428 00:50 min : 2x [AMD Opteron 6380 „Abu Dhabi” (8M/16T, 2.5 GHz, 2.8 GHz Turbo, DDR3-1333 ECC QC NUMA)] von Raspo
23,255 02:09 min : AMD FX-8350 „Vishera” (4M/8T, 4.0 GHz, 4.2 GHz Turbo, DDR3-1866 DC) von sompe
23,437 02:08 min : AMD FX-8350 „Vishera” (4M/8T, 4.0 GHz, 4.2 GHz Turbo, DDR3-1866 DC) von Nero24
23,478 04:29 min : AMD A10-7860K „Godavari” (2M/4T, 3.8 GHz, kein Turbo, DDR3-1600 DC) von eratte
23,504 01:46 min : AMD FX-8350 „Vishera” (4M/8T, 4.8 GHz, kein Turbo, DDR3-2133 DC) von Atombossler
23,645.32 FX-8370E (@ 5GHz)
23,668.64 FX-8350 stock
23,677 02:39 min : AMD FX-6350 „Vishera” (3M/6T, 4.25 GHz, kein Turbo, DDR3-1866 DC) von erde-m
23,767 01:39 min : AMD FX-8350 „Vishera” (4M/8T, 5.1 GHz, kein Turbo, DDR3-2133 DC) von Atombossler
23,809 01:45 min : AMD FX-9590 „Vishera” (4M/8T, 4.8 GHz, kein Turbo, DDR3-1333 DC) von Spacecake
23,861 01:47 min : AMD FX-9590 „Vishera” (4M/8T, 4.7 GHz, kein Turbo, DDR3-1333 DC) von Spacecake
23,973 03:33 min : AMD Opteron 3365 „Delhi” (4C/8T 2.35 GHz, kein Turbo, DDR3-1600 DC) von Fab8
24,188 04:02 min : AMD A10-6800K „Richland” (2M/4T, 4.1 GHz, 4.4 GHz Turbo, DDR3-2133 DC) von sompe
24,479 04:18 min : AMD A10-7860K „Godavari” (2M/4T, 3.8 GHz, kein Turbo, DDR3-2133 DC) von eratte
25,483 04:37 min : AMD A10-7850K „Kaveri” (2M/4T, 3.4 GHz, kein Turbo) von Casi030
25,518 09:30 min : AMD Athlon II X3 400e „Rena” (3C/3T, 2.2 GHz, kein Turbo, DDR3-1600 DC) von donmartin3000
25,805.06 FX-8370E (@ 5Ghz) single thread per module 25,805.06
26,402 03:22 min : AMD Athlon X4 860K „Kaveri” (2M/4T, 4,5 GHz, kein Turbo, DDR3-1600 DC) von Onkel_Dithmeyer
26,737 03:24 min : AMD Athlon X4 860K „Kaveri” (2M/4T, 4.4 GHz, kein Turbo, DDR3-1600 DC) von David64
28,261 06:26 min : AMD Athlon 5370 „Kabini” (4C/4T, 2.2 GHz, kein Turbo, DDR3-1600 SC) von Nero24
28,419 09:23 min : AMD A4-5000 „Kabini” (4C/4T, 1.5 GHz, kein Turbo, DDR3L-1600 SC) von genervt
28,571 09:20 min : AMD A4-5000 „Kabini” (4C/4T, 1.5 GHz, kein Turbo, DDR3-1600 SC) von Nero24
28,811.52 Phenom 965 BE@ 3.4
31,968 11:55 min : AMD Athlon II M320 „Caspian” (2C/2T, 2.1 GHz, kein Turbo, DDR3-1066 DC) von Nero24
32,051 00:26 min : 8x [AMD Opteron 8431 „Istanbul” (6C/6T, 2.4 GHz, kein Turbo, DDR2-667 ECC DC NUMA)] von bschicht86
32,345.01 A8 3850
34,873.58 Athlon X4 845
36,038 Phenom 2 X6 1100 @ 3.4
37,137 08:37 min : AMD Athlon X2 7550 „Kuma” (2C/2T, 2.5 GHz, kein Turbo, DDR2-800 SC) von Nero24
38,784 03:58 min : AMD Phenom II X4 810 „Deneb” (4C/4T, 2.6 GHz, kein Turbo, DDR3-1066 DC) von Nero24
39,215 02:00 min : AMD Phenom II X6 1055T „Thuban” (6C/6T, 3.4 GHz, kein Turbo) von HITCHER
39,434 02:59 min : AMD Phenom II X4 965 „Deneb” (4C/4T, 3.4 GHz, kein Turbo, DDR3-1333) von enigmation
39,608 (first core) quad Opteron 61xx, "Magny Cours", 48 K10 cores @3.0 GHz, 35,461 (48th core), DDR3-1333
40,920 01:55 min : AMD Phenom II X6 1100T „Thuban” (6C/6T, 3.3 GHz, 3.7 GHz Turbo, DDR3-1600 DC) von C4rp3di3m
51,282.05 Sandy Bridge 2500K @ 4.5 Ghz
51,306.17 i5 750 (“Lynnfield” 4C/4T, 3.8 GHz)
51,724 01:56 min : Intel Core i5 2500K „Sandy Bridge” (4C/4T, 4.0 GHz, kein Turbo, DDR3-2133 DC) von eratte
52,287.58 Sandy Bridge I5 @ 3 Ghz
54,421 01:45 min : Intel Core i5 3750K „Ivy Bridge” (4C/4T, 4.2 GHz, kein Turbo, DDR3-1333 DC) von mibo
54,770 i5-3570K (3.6 GHz), "Ivy Bridge"
62,500 00:32 min : 2x [Intel Xeon E5-2670 „Sandy-Bridge-EP” (8C/16T, 2.6 GHz, 3.3 GHz Turbo, DDR3-1600 DC)] von Sabroe SMC
64,205.46 4770K 64,205.46
67,838.8 SkyLake i5@3.63GHz
68,356.78 Xeon 5660
68,982.76 3770k
69,135 (first core) 63,694 (16th core) Xeon E5 2670 x2 (3.0 GHz turbo)
70,175 01:00 min : Intel Core i7 3930K „Sandy Bridge-E” (6C/12T, 3.8 GHz, kein Turbo, DDR3-1866 TC) von eratte
70,838 01:27 min : Intel Core i5 4690K „Haswell” (4C/4T, 4.4 GHz, kein Turbo, DDR3-1600 DC) von Stromae
70,937 01:05 min : Intel Core i7 980X „Westmere” (6C/12T, 3.47 GHz, kein Turbo, DDR3-1600 TC) von eratte
71,556 01:26 min : Intel Core i7 3820 „Sandy Bridge-EP” (4C/8T, 3.9 GHz, kein Turbo, DDR3-1600 QC) von wolle23
75,093 01:34 min : Intel Xeon E3-1270 „Sandy Bridge” (4C/8T 3.4 GHz, 3.8 GHz Turbo, DDR3-1333 DC) von sjrothe
75,650 00:47 min : Intel Xeon E5 V2 ES „Ivy Bridge-EP” (10C/20T, 2.7 GHz, kein Turbo, DDR3-1866 QC) von Sabroe SMC
76,628 04:21 min : Intel Core i5 520M „Arrandale” (2C/4T, 2.4 GHz, 2.9 GHz Turbo) von David64
77,896 01:19 min : Intel Core i7 3770K „Ivy Bridge” (4C/8T, 3.9 GHz, kein Turbo) von Nightshift
78,431 01:30 min : Intel Core i7 3770K „Ivy Bridge” (4C/8T, 3.4 GHz, kein Turbo) von Nightshift
78,431 00:51 min : Intel Core i7 4930K „Ivy Bridge-E” (6C/12T, 4.0 GHz, kein Turbo, DDR3-1600 QC) von eratte
78,947 01:16 min : Intel Core i7 3770K „Ivy Bridge” (4C/8T, 4.0 GHz, kein Turbo) von Nightshift
78,947.37 SkyLake
81,081 01:14 min : Intel Core i7 2600K „Sandy Bridge” (4C/8T, 4.0 GHz, kein Turbo, DDR3-1600 DC) von Sabroe SMC
82,644 01:28 min : Intel Xeon E3-1230 v2 „Ivy Bridge” (4C/8T, 3.3 GHz, 3.7 Turbo, DDR3-1600 DC) von genervt
83,457.36 3770 (non K) at 3.4 GHz 83,457.36
86,642 04:37 min : Intel Core i3 5005U „Broadwell” (2C/4T, 2.0 GHz, kein Turbo, DDR3-1600 SC) von eratte
89,921 02:37 min : Intel Core i3 4130 „Haswell” (2C/4T, 3.4 GHz, kein Turbo, DDR3-1600 DC) von eratte
91,185 00:28 min : Intel Core i7 5960X „Haswell-E” (8C/16T, 4.7 GHz, kein Turbo, DDR4-3000 QC) von MusicIsMyLife
91,242 Intel Xeon 1240v3 "Haswell" 3,8Ghz 4C/8T 1333er kein Turbo
91,427.18 5960X
91,954 04:21 min : Intel Core i3 5005U „Broadwell” (2C/4T, 2.0 GHz, kein Turbo, DDR3-1600 DC) von Nero24
92,307 01:05 min : Intel Core i7 4790K „Haswell” (4C/8T, 4.0 GHz, 4.4 GHz Turbo, DDR3-1600 DC) von Nero24
92,528.34 5820K @ 4.4GHz
93,389.76 6950X
93,811.38 5930K
94,581 00:29 min : Intel Core i7 6950X „Broadwell-E” (10C/20T, 3.5 GHz, kein Turbo, DDR4-2800 QC) von thorsam
94,861.66 4790K
95,238 00:42 min : Intel Core i7 6850K „Broadwell-E” (6C/12T, 4.0 GHz, kein Turbo, DDR4-2400 QC) von eratte
96,774 01:02 min : Intel Core i7 4790K „Haswell” (4C/8T, 4.0 GHz, 4.4 GHz Turbo, DDR3-1600 DC) von eratte
97,595 (95,238?) 00:35 min : Intel Core i7 6900K „Broadwell-E” (8C/16T, 3.2 GHz, 3.7 GHz Turbo, DDR4-2400 DC) von AMD
98,039.22 AMD Ryzen (went with 36s), 8C/16T, 3.4GHz no turbo
104,166 i7-7700k "KabyLake" @ 5Ghz / DDR4-2800
105,207.79 i7-6700K "SkyLake"
111,731.84 i7-6700K "SkyLake"

 

[siriq]

Who the hell wants to use an unoptimized Blender? No matter if intel or amd cpu has been used with SIMD edition, you can clearly see the benefit on both vendors.

Best thing to my opinion, make ZEN and rest of the CPU's the same SIMD Blender bench .

 

[rvborgh]

for comparison reasons (to give performance relative to Blender 2.78a running the Ryzen file on Ryzen).

 

[superstition]

for comparison reasons (to give performance relative to Blender 2.78a running the Ryzen file on Ryzen).

Outdated comparison, though.

What's that formula again? I'd like to see, anyway, how the 4.4 GHz and 4.8 GHz results compare, including the 4.4 CMT off results. Also, it would be good to label the results for the people here since there are two entries for the Skylake 6700K.

 

[rvborgh]

this thread is for Blender 2.78a comparisons so for the purposes and in the context of this thread... the SIMD version of Blender is just a curiosity really.

800*800*150/seconds/cores/Ghz

4.8 and 4.4 GHz results should be almost identical since the results are per Ghz

 

[superstition]

this thread is for Blender 2.78a comparisons so for the purposes and in the context of this thread... the SIMD version of Blender is just a curiosity really.

800*800*150/seconds/cores/Ghz

Thanks for posting the formula. My opinion is that the SIMD build is less of a curiosity case than the stock build since it's much more practical to render using a renderer that runs a lot faster but we'll have to agree to disagree.

4.8 and 4.4 GHz results should be almost identical since the results are per Ghz

Normally yes but I'm not convinced my 5 GHz CMT-enabled results haven't been tainted by a weak motherboard so I want to verify.

That formula is a bit of a question itself since it uses "cores". That brings up the whole core/module issue and the way CMT chips have up to 8 integer cores and 4 FPU cores.

I'd say that an integer test and a pure FPU test would be a better benchmark for comparing designs with a formula like this given the existence of CMT designs. That way they could be parsed as 8 core for the integer test and 4 for the FPU test.

 

[superstition]

That formula is a bit of a question itself since it uses "cores". That brings up the whole core/module issue and the way CMT chips have up to 8 integer cores and 4 FPU cores.

SIMD results

168067 — Skylake i7 6700K, 3200 CAS15, stock, static undervolt, 0:35.7
53952 — Piledriver 4.4 GHz CMT off 1:41.1
53935 — i5 750 Lynnfield, 1600 RAM 1:57.1
53752 — Piledriver 5 GHz CMT off, 1:29.36
45683 — Piledriver 4.4 GHz CMT on, 0:59.7
45455 — Piledriver 4.8 GHz CMT on, 55

There is no way to really deal with the threads/cores issue. Changing the equation to threads drops the Skylake result to 84034 but the FPU-to-integer ratio issue is a problem. If, for instance, we count Piledriver as a 4 core part we get these results:

168067 — 6700K, 0:35.7
91366 — PD 4.4 GHz CMT on, 0:59.7
90909 — PD 4.8 GHz CMT on, 55
53952 — PD 4.4 GHz CMT off, 1:41.1
53935 — Lynnfield, 1:57.1
53752 — PD 5 GHz CMT off, 1:29.36

If Blender is using FPU then the reality is somewhere in-between counting as 8 core and counting as 4.

 

[superstition]

Note to siriq...

Your 4.8 overclock doesn't seem optimal after all. I guess I mistook your SIMD results for your stock Blender results or something because your Blender results are lower than my Piledriver results as is your Cinebench multi. In fact, I was able to get 770 as a new high with Cinebench multi at 4.8 (up from two runs at 767).

Using the formula with your SIMD result gets only this score: 37314. Mine was 45455 at 4.8.

With the stock build the result is only 20000. Something is off with your system. Must be throttling.

You said 2:05 (125 seconds) for stock Blender at 150 samples. That's 96000000/125/8/4.8 = 20000. By contrast, I got 23923.44

 

[superstition]

stock Blender, my results put into the equation

113777.51 (@3.8) to 166,290 (@2.6) — i7 4960HQ Macbook Pro, 0:55.51
111731.84 — 6700K in Windows 10
51306.17 — Lynnfield 3.8 4/4
25806.45 — PD 5 GHz CMT off
25680.53 — PD 4.4 GHz CMT off
23944.45 — PD 4.4 GHz CMT on
23923.44 — PD 4.8 GHz CMT on
23645.32 — PD 5 GHz CMT on (overclock not optimized)

Funny how a mobile Haswell part tops the chart. The following possibilities:

Mac build is compiled better.
OS X is more efficient.
128 MB L4 cache was enabled as a victim cache (prior to Broadwell-C) and is a big help.

I'm going for the first one as being the main factor but maybe someone can tell me if the OS X stock 2.78a build indeed has things like AVX turned on or not.

 

[bjt2]

Note to siriq...

Your 4.8 overclock doesn't seem optimal after all. I guess I mistook your SIMD results for your stock Blender results or something because your Blender results are lower than my Piledriver results as is your Cinebench multi. In fact, I was able to get 770 as a new high with Cinebench multi at 4.8 (up from two runs at 767).

Using the formula with your SIMD result gets only this score: 37314. Mine was 45455 at 4.8.

With the stock build the result is only 20000. Something is off with your system. Must be throttling.

You said 2:05 (125 seconds) for stock Blender at 150 samples. That's 96000000/125/8/4.8 = 20000. By contrast, I got 23923.44

In OC there could be also L1/L2/L3 parity/ecc issues, with retransmissions or reloading from RAM...

 

[superstition]

In OC there could be also L1/L2/L3 parity/ecc issues, with retransmissions or reloading from RAM...

I found a big boost in CB R15 multi when I bumped up CPU NB when I did some quick 4.6 testing recently. Even though everything seemed fine with the lower voltage the performance wasn't as good.* This is why the quest for the lowest possible stable voltage doesn't seem like it's always the best one. However, I have seen lower GFLOPs in Linpack (LinX) when voltages are too high, too. That could be the board being weak but my guess is that the CPU runs a little faster when it has the most optimal voltages, although the settings have to be a bit too high to deal with inverse spikes.

*I think this is related to the board because on the higher-end CFZ CPU NB didn't seem to make much of a difference, probably because of the fact that it can have LLC set just for it. Of course, people have also said that board under-reports voltages, too.

Someone said PD, when overclocking via BCLK, likes a 250 setting because of strapping but my board can't do 250. It's happy at 237 and 238 but no higher. Perhaps the additional latency of the less optimal setting makes it easier to have stability. I wonder if the latest BIOS for my board finally fixed the multiplier boot bug. I haven't bothered with my gaming system much since I updated. I have been too busy with my Macbook working on projects.

Anyway... I'm wondering if reducing the core count by 25% when calculating CMT-enabled CPUs would make that equation a bit less silly. Clearly, counting an 8 core PD as 4 is silly, too, since this test isn't just FPU. However, when looking at 4C Sandy's result with the stock Blender versus PD's it makes counting PD as 4C seem more reasonable, particularly given the way PD has been able to hold its own in a bunch of recent games when compared with Sandy.

For instance, if we look at this:

52482.42 — Shivansps’s 2500K @4.2, stock Blender 118.9s
47888.90* — my 8370E @4.4, stock Blender 113.9s

*calculated as 4C

Sure, it looks like the extra integer cores of PD help to close the gap with Sandy but do we really believe that Sandy merits a score of 52482.42 while PD merits a score of only 23944.45?

Consider this result, and that it's only loading 6 cores of the 8C PD, at least with one 980 Ti:

 

[Atari2600]

Thanks for posting the formula. My opinion is that the SIMD build is less of a curiosity case than the stock build since it's much more practical to render using a renderer that runs a lot faster but we'll have to agree to disagree.

I agree with you that a SIMD build would be more useful in reflecting optimised performance - but AMD did not bench Zen with a custom compiled version - and for sound, obvious reasons.

So, without Zen data on a custom compilation, any comparisons across different CPUs on a SIMD build are pretty useless to Zen... thus its a curiosity case as pertains to this thread.

 

[superstition]

I agree with you that a SIMD build would be more useful in reflecting optimised performance - but AMD did not bench Zen with a custom compiled version - and for sound, obvious reasons.

So, without Zen data on a custom compilation, any comparisons across different CPUs on a SIMD build are pretty useless to Zen... thus its a curiosity case as pertains to this thread.

I don't agree. We can just keep re-posting the same opinions but it won't go anywhere without new information.

 

[Atari2600]

I don't agree.

So what Zen score are you going to compare to?

 

[bjt2]

I found a big boost in CB R15 multi when I bumped up CPU NB when I did some quick 4.6 testing recently. Even though everything seemed fine with the lower voltage the performance wasn't as good.* This is why the quest for the lowest possible stable voltage doesn't seem like it's always the best one. However, I have seen lower GFLOPs in Linpack (LinX) when voltages are too high, too. That could be the board being weak but my guess is that the CPU runs a little faster when it has the most optimal voltages, although the settings have to be a bit too high to deal with inverse spikes.

*I think this is related to the board because on the higher-end CFZ CPU NB didn't seem to make much of a difference, probably because of the fact that it can have LLC set just for it. Of course, people have also said that board under-reports voltages, too.

Someone said PD, when overclocking via BCLK, likes a 250 setting because of strapping but my board can't do 250. It's happy at 237 and 238 but no higher. Perhaps the additional latency of the less optimal setting makes it easier to have stability. I wonder if the latest BIOS for my board finally fixed the multiplier boot bug. I haven't bothered with my gaming system much since I updated. I have been too busy with my Macbook working on projects.

Anyway... I'm wondering if reducing the core count by 25% when calculating CMT-enabled CPUs would make that equation a bit less silly. Clearly, counting an 8 core PD as 4 is silly, too, since this test isn't just FPU. However, when looking at 4C Sandy's result with the stock Blender versus PD's it makes counting PD as 4C seem more reasonable, particularly given the way PD has been able to hold its own in a bunch of recent games when compared with Sandy.

For instance, if we look at this:

52482.42 — Shivansps’s 2500K @4.2, stock Blender 118.9s
47888.90* — my 8370E @4.4, stock Blender 113.9s

*calculated as 4C

Sure, it looks like the extra integer cores of PD help to close the gap with Sandy but do we really believe that Sandy merits a score of 52482.42 while PD merits a score of only 23944.45?

Consider this result, and that it's only loading 6 cores of the 8C PD, at least with one 980 Ti:

On CON core the L3 goes at NB frequency. So no wonder that you obtain a boost especially on memory hungry tasks. Moreover if the clock difference between NB and core is some odd ratio (not 1:2, 1:3 etc) there are some CPU cycle additional latencies to syncronize the clocks: you can't transmit data on both rising and falling edges of both clocks and you must use a buffer. So for clock slightly different this can mean that you lose many cycles...

EDIT: for instances, worst case of prime multipliers, e.g. 11 or 13 for NB, they incur in a big penality. This is also a reason they use always 12 (2400MHz) for NB frequency even on higher models. Next good multiplier is 15 (3GHz) and 16 is the best and i think that this is the NB clock on Zen, because L3 is said 5x the bandwidth: 4x the bus and 1.25x the clock. 2.4x1.25=3... Maybe they use 16x, that is better (2^4, 3.2Ghz) and rounded to integer number the speedup...

 

[superstition]

Sorry I wasn't clear. CPU NB referred to the voltage going to that rather than the clock. However, I do have mine at 2400 rather than 2200 which is the standard for the lower-spec 970 chipset (which is what my board has). Overclocking guides for Piledriver typically say it's not necessary to raise CPU NB voltage (because they were written by people using high-end ASUS boards that have special LLC for the CPU NB itself and which reportedly under-report that voltage) while others have people raise the voltage to 1.3V immediately.

I try to avoid raising the BCLK but had to (if going over 4.4) because the board had a BIOS bug that prevented it from booting with a multi over 22. I think that may be solved now with the latest FC BIOS but I'm not sure. I've been too busy with projects on my Macbook to mess around with my gaming system.

Thanks for that northbridge information. I'm glad that my trial and error ended up with the right setting (2400). As long as I can avoid the boot bug (or just stick with 4.4 GHz) I can use that.

So what Zen score are you going to compare to?

That was AMD's decision not mine. However, I hope that at least one person who gets an early sample to review will take the time to test a proper Blender build. It takes very very little time to run 150 samples twice and keep the best result, after unzipping the program and loading the file.

 

[Atari2600]

That was AMD's decision not mine.

Yes, but could you imagine the outcry there would be if AMD used their own compilation as a benchmark?

It would forever be accused of bias against the Intel CPUs.

 

[superstition]

Yes, but could you imagine the outcry there would be if AMD used their own compilation as a benchmark?

It would forever be accused of bias against the Intel CPUs.

The Linux build is already reportedly using modern instructions I've read (both here and on OCN).

It would also be difficult to argue that a binary compiled with Intel's compiler would be horribly biased against Intel.

As far as I know, The Stilt's SIMD build is compiled with Intel's compiler. It has better performance on both PD and Intel CPUs.

The AVX2 build was made with Microsoft's compiler and also posts better results with everything except Lynnfield (where it's a tiny bit slower than stock Blender), although not as much of a gain as the SIMD build.

 

[bjt2]

Sorry I wasn't clear. CPU NB referred to the voltage going to that rather than the clock. However, I do have mine at 2400 rather than 2200 which is the standard for the lower-spec 970 chipset (which is what my board has). Overclocking guides for Piledriver typically say it's not necessary to raise CPU NB voltage (because they were written by people using high-end ASUS boards that have special LLC for the CPU NB itself and which reportedly under-report that voltage) while others have people raise the voltage to 1.3V immediately.

I try to avoid raising the BCLK but had to (if going over 4.4) because the board had a BIOS bug that prevented it from booting with a multi over 22. I think that may be solved now with the latest FC BIOS but I'm not sure. I've been too busy with projects on my Macbook to mess around with my gaming system.

Thanks for that northbridge information. I'm glad that my trial and error ended up with the right setting (2400). As long as I can avoid the boot bug (or just stick with 4.4 GHz) I can use that.

That was AMD's decision not mine. However, I hope that at least one person who gets an early sample to review will take the time to test a proper Blender build. It takes very very little time to run 150 samples twice and keep the best result, after unzipping the program and loading the file.

Probabily if the NB does not reach 3 or 3.2GHz stable, it's better to keep x12 multiplier and raise bclk... Or stick to 2400. But I remember that in the bios guide of CON cores, for performance reasons the multiplier of the CPU should be <= 2*NB. So over 4800 either you raise also the NB multiplier or raise BCLK...

EDIT: this probabily stems from the internal bus width of the NB. if it's clocked too low it can't cope with the data flux...