Epyc 9334 qs over 7950x for vfx and 3d

[Bartosz Grzelak]

Hello servethehome fans!

I am looking to build vfx workstation this year,
and I came across above epyc in qs version on eBay for roughly the same price 150usd + or - as the 7950x new.

Now, would it be crazy to go for qs sample and risk inconsistent performance?

I actually have messaged couple of those eBay sellers selling 9334's and they have all confirmed a compatibility with supermicro H13 boards.

Maybe someone has some experience with those units?

 

[Patriot]

QS is qualifications sample, if its the same stepping as production its the same silicon. ES is generally earlier samples with errata's and such that may cause inconsistencies or instability or early degradation. You should confirm the stepping before buying.

What is your VFX workload?
I helped a buddy with a Houdini workload among other things.
Built on rome with 2x64c and and a tb of ram. Getting to see 256t hit load is quite satisfying.

The ram bandwidth over consumer is probably going to be quite helpful for your tasks, but you will be much lower clocks so just 32c might not be as much faster as you expect. Your costs will be much larger delta than you expect as well...due to ram costs. You are going to want to feed it atleast 6 dimms... 12 preferable.

 

[RolloZ170]

nd I came across above epyc in qs version

not any is a real QS, even if the seller states it is one.
and there are Genoa B0 and B1 QS. you should think about why AMD made the B1 stepping.

 

[Patriot]

https://www.amd.com/system/files/TechDocs/57095-PUB_1.01.pdf

 

[bayleyw]

They're not really comparable. The 9334 is faster, but because scaling is crap past 16 cores it's not anywhere close to 2x faster than a 7950x. The boards also have long POST times, no audio, and suspect GPU compatibility (my H11SSL is really picky about when it decides to show the BIOS through the dGPU, for example), but on the other hand you get more PCIe lanes (kind of useful for VFX) and virtually unlimited RAM. The latter might be a gating decision for you, based on what I know about production rendering workloads the models have a voracious appetite for RAM - the WDAS Moana Island test scene uses something like 140 GB of memory and even that's from 2016.

I am not sure if Genoa still has the same pathologies as Rome/Milan - the previous gen AMD parts really didn't like partial memory population. I had a bespoke not-supposed-to-be-bandwidth-limited application suffer a 2x loss of performance from leaving 2/8 channels unpopulated.

 

[Bartosz Grzelak]

QS is qualifications sample, if its the same stepping as production its the same silicon. ES is generally earlier samples with errata's and such that may cause inconsistencies or instability or early degradation. You should confirm the stepping before buying.

What is your VFX workload?
I helped a buddy with a Houdini workload among other things.
Built on rome with 2x64c and and a tb of ram. Getting to see 256t hit load is quite satisfying.

The ram bandwidth over consumer is probably going to be quite helpful for your tasks, but you will be much lower clocks so just 32c might not be as much faster as you expect. Your costs will be much larger delta than you expect as well...due to ram costs. You are going to want to feed it atleast 6 dimms... 12 preferable.
View attachment 30784

Sweet setup, it looks very expensive though.

I am trying to transition my career from architectural / construction visualisation to vfx. Tyflow, houdini, nuke, max, maybe some blender, unreal, and the new ambergen / liquidgen . Many of those software's are CUDA heavy though. According to Patrick from STH you can actually run it without all 12 dims populated, but there is performance penalty of course and yeah I have the lower clocks in mind. 3ds max's clunky ui runs on single core only unfortunately .

They're not really comparable. The 9334 is faster, but because scaling is crap past 16 cores it's not anywhere close to 2x faster than a 7950x. The boards also have long POST times, no audio, and suspect GPU compatibility (my H11SSL is really picky about when it decides to show the BIOS through the dGPU, for example), but on the other hand you get more PCIe lanes (kind of useful for VFX) and virtually unlimited RAM. The latter might be a gating decision for you, based on what I know about production rendering workloads the models have a voracious appetite for RAM - the WDAS Moana Island test scene uses something like 140 GB of memory and even that's from 2016.

I am not sure if Genoa still has the same pathologies as Rome/Milan - the previous gen AMD parts really didn't like partial memory population. I had a bespoke not-supposed-to-be-bandwidth-limited application suffer a 2x loss of performance from leaving 2/8 channels unpopulated.

I am currently running 2700x now with only 16gb ram. When I open my scene's 90 percent of ram is taken right away, mind you I've always try to optimize my scene's, so I cant open anything else after that. I dont mind slow post times, as my pc is already taking ages, having 100gb nic and raid card in it. Also my audio goes through my monitor over hdmi anyway as all my gear is tacked away in a rack.

not any is a real QS, even if the seller states it is one.
and there are Genoa B0 and B1 QS. you should think about why AMD made the B1 stepping.

It looks like it is B0 stepping on most of the ebay listings.