Dual Epyc 9654 experience?

[hrvojezg00]

Hi, Is anyone here running a dual Epyc Genoa system, preferably 9654 or 9754`s? I wonder if the scale is better than with EPYC 7702`s 1 CPU vs. 2 CPU systems.

[andrew1988]

dual cpu system is not optimized with corona 10 and above so far
maybe windows process group is the culprit  i am not sure

Hi, Is anyone here running a dual Epyc Genoa system, preferably 9654 or 9754`s? I wonder if the scale is better than with EPYC 7702`s 1 CPU vs. 2 CPU systems.

[webuilddreams]

Hi, we have a single processor AMD epyc 9754 with a supermicro mobo and 384 GB memory DDR5, it is fast as hell BUT the geometry fase is 10x slower than my old pc (dual intel xeon E5-2699V4 arround 8 years old!!), we are still trying to figure out why, and we render a lot of animation so this phase comes back in every frame, we loose between 60 and 120 seconds per frame!

[Nejc Kilar]

Hi, we have a single processor AMD epyc 9754 with a supermicro mobo and 384 GB memory DDR5, it is fast as hell BUT the geometry fase is 10x slower than my old pc (dual intel xeon E5-2699V4 arround 8 years old!!), we are still trying to figure out why, and we render a lot of animation so this phase comes back in every frame, we loose between 60 and 120 seconds per frame!

Hmm that really shouldn't be the case imho - yes the 2699v4 has a 500mhz higher boost clock but the IPC has come a long way in the 10 years since it was released.

Could you please double check what frequency is being reported (and core utilization) during scene parsing? Not all the cores should be busy but you should be close to at least 3ghz and that would indicate it is not the hardware at fault.

[webuilddreams]

Hi Nejc Kilar, we checked it and in de screenshot that was taking during the geomtery phase u see that the core speed is arround 3 GHz and it utilizes 70% of the cores plus the next time we render the same it gets worse, from 0:39 min first time to a stunning 3:05 seconds (after 5 renders same cam) only for the geometry phase!

[maru]

Hi, I'm not sure what is going on here, but it doesn't look like something expected. Can you please share a sample scene where this can be reproduced here: https://support.chaos.com/hc/en-us/requests/new

[webuilddreams]

Hi maru, yes we can share this scene, but it happens in all the project i am working on, tested this on 20+ different projects/3dsmax files and in every case it is the same: the geometry phase takes 10 x longer? also compared this with a friend (ryzen 5900x) and his geometry phase was arround 12 x faster (same scenes/projects, tested this with over 10 diff max files) so share?

[maru]

Yes please, even a simple scene is better than no scene.

[hrvojezg00]

Yes please, even a simple scene is better than no scene.

Maru, any news on this?

[maru]

Yes please, even a simple scene is better than no scene.

Maru, any news on this?

There is no change - I still believe a simple scene is better than no scene. :)

On a serious note: once we have any test results from a user scene or our internal scenes, we will surely share them.

[hrvojezg00]

Hi maru, yes we can share this scene, but it happens in all the project i am working on, tested this on 20+ different projects/3dsmax files and in every case it is the same: the geometry phase takes 10 x longer? also compared this with a friend (ryzen 5900x) and his geometry phase was arround 12 x faster (same scenes/projects, tested this with over 10 diff max files) so share?

It would be great to sort it out, and it would help others decide whether to go with 9654/9754. Can you please share the scene with Maru to sort it out?

[webuilddreams]

Yes, we just did via the ticket.

[webuilddreams]

Hi maru, yes we can share this scene, but it happens in all the project i am working on, tested this on 20+ different projects/3dsmax files and in every case it is the same: the geometry phase takes 10 x longer? also compared this with a friend (ryzen 5900x) and his geometry phase was arround 12 x faster (same scenes/projects, tested this with over 10 diff max files) so share?

It would be great to sort it out, and it would help others decide whether to go with 9654/9754. Can you please share the scene with Maru to sort it out?

Yes very happy with it, we use a singel 9754, we did some tests with a double 9654 EPYC, and it turns out that we have an overall speed of 140% compared to a single 9654. With the 9754 we have a speed of 130% compared to 2 x 9654.