Dual EPYC 64cores ES: 17.05 (& 1014FPS) [UPDATED with KeyShot 10]

ohashi3d · March 28, 2020, 10:53:10 AM

I've modified the subject of topic and I'll post the results with KeyShot10 soon later.
And I'll keep the old post below just for the records.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

I've just tested AMD EPYC 64cores Engineering Sample CPU '100-000000053-04', got at eBay.

Motherboard: Supermicro H11DSi-NT Rev. 1.01 (with self-modified R2.1 BIOS for Zen2 Rome)
CPU: Dual AMD Eng Sample 100-000000053-04, boosted ~2.7GHz(SMT-on), ~3.0GHz(SMT-off), 64 Core
RAM: DDR4-2666 2R(Dual Rank) 8GB x16
GPU: Quadro RTX 4000
Camera Benchmark: ~720FPS(SMT-on) ~998FPS(SMT-off) [UPDATE: I've replaced the sceenshots with the one by KeyShot 9.3.14]
CAUTION: H11DSi-NT(=dual CPU motherboard) with single '100-000000053-04' could not boot up, In my experience.

Motherboard: Supermicro H11SSL-i Rev. 2.0
CPU: Single AMD 64cores Eng Sample 100-000000053-04, boosted ~2.7GHz(SMT-on)
RAM: DDR4-3200 2R(Dual Rank) 16GB x8
GPU: GeForce GT710
Camera Benchmark: ~740FPS

I had expected that dual ES '100-000000053-04' would exceed 1000FPS easily, but no.
Disabling SMT(Simultaneous Multithreading) is a workaround, but it affects V-ray bench and CINEBENCH negatively.
So, KeyShot may have some issue about thread control over 128 threads...
[UPDATE: KeyShot 9.3.14 realtime mode may still have issue, but now it seems that KeyShot Vierwer 9.3.14 can handle 256 threads properly... I'll post about it later.]

designgestalt · March 28, 2020, 02:08:37 PM

holy moly ...

DMerz III · March 28, 2020, 04:52:39 PM

I also can't believe this didn't crack 1000. Has anyone been able to achieve that yet?

Eugen Fetsch · March 29, 2020, 04:02:08 AM

Quote from: kyosen on March 28, 2020, 10:53:10 AM
So, KeyShot may have some issue about thread control over 128 threads...

Or maybe Windows does? https://www.anandtech.com/show/15483/amd-threadripper-3990x-review/3

ohashi3d · March 29, 2020, 06:49:25 AM

Quote from: Eugen Fetsch on March 29, 2020, 04:02:08 AM
Quote from: kyosen on March 28, 2020, 10:53:10 AM
So, KeyShot may have some issue about thread control over 128 threads...
Or maybe Windows does? https://www.anandtech.com/show/15483/amd-threadripper-3990x-review/3

Thanks...hmm, Win10 Pro for Workstations (or Enterprise)... I still hope the solution built-in KeyShot under Win10 Pro(not "for Workstations")!

ohashi3d · April 05, 2020, 07:10:48 PM

In the KeyShot 9.2 Beta Discussion section, I've just post the results of the benchmark built in KeyShot Viewer 9.2.71:
https://forum.keyshot.com/index.php?topic=25963.msg109455#msg109455

ohashi3d · May 04, 2020, 07:11:43 PM

I've modified the first post on this thread and as described there, now It looks that KeyShot Viwer 9.3.14 Benchmark can handle 256 threads properly...better than 9.2.86, at least.

Motherboard: Supermicro H11DSi-NT Rev. 1.01 (with self-modified R2.1 BIOS for Zen2 Rome)
CPU: Dual AMD Eng Sample 100-000000053-04, boosted ~2.7GHz(SMT-on), ~3.0GHz(SMT-off), 64 Core
RAM: DDR4-2666 2R(Dual Rank) 8GB x16
GPU: Quadro RTX 4000

CPU Benchmark: 13.66(SMT-off) -> 15.84(SMT-on)

DerekCicero · May 12, 2020, 10:15:33 AM

Yes the Benchmark Tool was updated for 9.3 for more accurate, consistent results. Thanks for using it!

kreita · July 18, 2020, 04:31:21 AM

Wow, thank you for all the benchmark work on the 128 core dual EPYC. As mentioned it would be great if KS could do further testing with dual EPYC using Win 10 Pro vs Win 10 Pro Workstation or Enterprise? Just wondering if even higher performance levels can be achieved. My interests are to maximize the speed of KS's live preview. If fast enough, I could change out materials and furniture in my interior scenes live in front of clients to view (basically have a configurator). How cool would that be. Right now with a 28 core machine It takes 30 - 40 minutes for good clarity (1gig+ aircraft interior models with high polygon and many lights). Future looks good for high core count machines. To get over 100 cores 3 yrs ago I'd guess price at 40K+, now with dual EPYC 17-20K.

mafrieger · July 27, 2020, 04:06:36 AM

yes would be awesome to examine the borders of keyshot and windows ;-)
didn't need to be 256 threads, but 128 real cores..

mafrieger · August 05, 2020, 01:47:53 AM

in comparison:

1 x AMD RYZEN THREADRIPPER 3970X (3.70 GHz) 32 cores / 64T with SMT-on is showing around 8,7 in Benchmark
see https://forum.keyshot.com/index.php?topic=26615.0

mafrieger · August 05, 2020, 01:59:57 AM

having this in mind,

15.84(SMT-on) using 128core/256T @ 2,7Ghz

which is about 128cores/32cores * 2,7GHz/3,7GHz = 2,9 fold of raw power

for 15,84points / 8,7point = 1,8 fold scaling

is hmmm not the thing one is really looking for when investing that much money in a 2xEpyc 64core platform...

if this really solely hanging on windows scheduler, MS has to do some homework... it's 2020 and 128core Workstations are here to stay.. and even more cores are not that far ahead...

(hopefully you can prove this caculation wrong or leverage some parts of this not ideal scaling without having to wait for MS's Windows 11..)

ohashi3d · August 06, 2020, 03:12:18 AM

Hi mafrieger,

Basically your calculation looks valid, I think.
Ryzen Threadripper 3970X may be running at around boosted 4GHz, so the difference of raw power may be a bit smaller/calm though.

Over 128 threads issue is a fundamental problem of MS's Windows, I suspect.
If Luxion disclose a KeyShot for Linux edition, we may see another scenario...

Anyway, I recommend ThreadRipper system than dual EPYC system, for KeyShot, at this moment;)

mafrieger · August 06, 2020, 12:17:45 PM

Quote from: ohashi3d on August 06, 2020, 03:12:18 AM

Over 128 threads issue is a fundamental problem of MS's Windows, I suspect.
If Luxion disclose a KeyShot for Linux edition, we may see another scenario...

yes that's sadly know, but for me it looks like that the problems do not start at around 128threads but much earlier...
Otherwise there should be a pretty linear scaling when going from 64cores/64threads (one processor, SMT disabled) to 128cores/128threads (two processors, SMT disabled) - can you proof this?
Would be pretty interesting to learn when linear scaling begins to end...

ohashi3d · August 06, 2020, 03:31:00 PM

Here is my post about my dual EPYC 7402 system: https://forum.keyshot.com/index.php?topic=25190.msg105438#msg105438
EPYC 7402 is a physical 24 cores / logical 48 threads CPU.

In the post above, dual EPYC 7402 (SMT-on) system shows 732.2 FPS for the good old Camera benchmark.
And I've attached the screenshot of a single EPYC 7402 (SMT-on) system...it shows 457.4.
732.2 / 457.4 = 1.6
This is the scaling factor from 48 threads (1 socket) to 96 threads (2 sockets), for KeyShot realtime mode...KeyShot Viewer Bench may show different trend.
It's not linear, but still acceptable for me, as from the very beginning I think that we can't expect linear scaling from the 1 socket system to 2 sockets system...