Workstation with CPU of 4.5+ Ghz base frequency that supports 256GB RAM + RTX 5090 or 2X RTXXXX?

[Vizioen]

Hi all, currently I have two AMD TR workstations: 3970x and 3990x, and I'm a bit tired of the lower frequencies (3.7 - 3.9Ghz) for working and test rendering, so I'm looking for a new CPU with higher frequencies. I want to replace my slowest threadripper workstation and turn that one in a rendering node, remove the GPU (RTX2080ti).

I was looking for CPUs with a base frequency of say at least 4.5 GHz, but also support of 256GB of RAM. Unfortunately I found that the Ryzen 9 9950X and 7950X don't support 256GB RAM (192GB Max), although I found this article HERE but couldn't find anything else that supports these claims or people that build these configurations, except for this old thread on Reddit: HERE but that's also inconclusive.

Then I found the Threadripper PRO 7945WX and Threadripper PRO 7955WX, both have quite high frequencies and the prices are reasonable (1200-1500$), BUT I can't find any benchmarks on the V-Ray Benchmark website nor find a lot about them in general? It seems they aren't widely available to buy either, which I find weird. Is there something special about these CPUs? They are released over a year ago.

Rendering performance wise, I was looking for something that matches more or less the V-Ray benchmark of the 3990x (around 61000 points), but as I don't know the benchmarks of the 7945WX or 7955WX, I have no idea if it does, I would guess they should be in the same ballpark. I don't have a problem if it's a bit lower, as long as the base frequency is at least 4.5Ghz, preferably even higher.

So in short, what would you guys recommend for a CPU with at least 4.5Ghz (and supports at least 256GB RAM). As I can't find anything else besides the aforementioned 7945WX and 7955WX. Also is it actually worth it doing this upgrade 3.7Ghz to 4.5 Ghz, will I notice a lot of difference working in 3dsMax, Photoshop, etc, for instance? I'm looking for a quite big performance boost while working in general, not just a little 'snappier'.

Edit: Actually I'm not sure I should look at the base frequency or the boost frequency for having a smoother work experience (viewport, materials, etc) then there might be better options probably than the proposed ones, for instance the TR 7960X?

Another thing I'm searching for is a GPU as I want to experiment more with GPU rendering (Vantage or Envision), running local LLMs or generative AIs (Comfy, etc) and real-time gaming engine solutions (Unreal, Twinmotion,..). Currently I have a 2080Ti and 3080Ti, but they have quite low VRAM and for GPU rendering, real-time and AI they are really struggling.

Nvidia just announced their new RTX50XX series, releasing end of January, and the RTX5090 looks nice because of the 32GB VRAM, although the price tag is quite insane.

So I was wondering wouldn't it be better to maybe buy two 16GB RTX40XX series, not sure which one though and if it will be cheaper or if the performance will be more or less the same, can . Also Unreal and Twinmotion only support multiGPU with NVLink/ SLI so for raster rendering the second won't do anything, unless I'm looking at 2 x 3090, which is the last to support that. So I guess dual GPUs + NVLink will be more expensive and running on lower technology. Is there any benefit in Vray GPU / Vantage / ... to run multiple GPUs instead of a beefy one like that RTX5090? I read this article: HERE and it seems running multiple GPUs is actually not that beneficial in terms of VRAM not stacking, so might not be such a good choice to run multiple GPUs

Which leads me to an additional question, if I would decide on the new system, I'm left with the RTX 2080Ti and RTX 3080Ti. I was thinking to put these together with my AMD 3990x for my workstation at home, but seeing as it's two different cards and most of my projects still require a lot of memory, I probably wouldn't benefit from it that, using it that way? Is there anything else I could use the 2080ti for?

Thanks!

[Nejc Kilar]

I reckon the 7945WX and 7955WX are OEM only CPUs which means you can only get them in a pre-built system. IIRC the 7960X (or 7965WX if you want the more PCI-E lanes PRO platform) is the lowest grade Threadripper CPU you can buy directly from a shop. That excludes Ebay sellers of course

I think, although I could be wrong, you can basically consider these "lower end Threadrippers" (still magnificent CPUs but just lower end in the Threadripper lineup) as ~their Ryzen counterparts. So a 7955WX is kind of a 7950x-ish CPU. The Threadrippers are a bit slower but broadly speaking I don't recall seeing benchmarks that would put them in a different class, its just that they probably boost a couple of 100mhz less than the Ryzen parts which technically is slower sure but not by a large margin. You do get more PCI-E lanes in return though

IMHO the closest competitor to the 3990x would be the 7970x / 7975WX which if I'm not mistaken is generally just a hair bit slower than the 3990x. Puget Systems did some testing (https://www.pugetsystems.com/labs/ar...s-generations/) and I'm sure you can find similar data elsewhere online including on the V-Ray's benchmark website (https://benchmark.chaos.com/vray/v6/...?id=816&id=810 - DO NOTE I'm not 100% sure if any of these are overclocked so treat the comparison as an estimate).

As for GPUs, depends on the types of projects / things you want to do imho. From a usability standpoint an important thing to consider is the VRAM buffer size. If your projects (or if you like having say Resolve open and rendering at the same time) take up lots of VRAM then it probably makes sense to go for the card that has a big enough VRAM buffer otherwise you'll be in for quite a rough ride. Broadly speaking, the 5xxx stack is imho kind of designed to push you to the more expensive 32GB GPU as 16GB can be a little on the low end for a lot of the CG work (and Nvidia probably wants to sell more 5090s and 5080s lol) although that might not be at all true for the types of projects you want to do.

I mean if you got by with a 2080ti then 16GB will probably be enough.

Hope my reply helps!

[Vizioen]

I reckon the 7945WX and 7955WX are OEM only CPUs which means you can only get them in a pre-built system. IIRC the 7960X (or 7965WX if you want the more PCI-E lanes PRO platform) is the lowest grade Threadripper CPU you can buy directly from a shop. That excludes Ebay sellers of course

I think, although I could be wrong, you can basically consider these "lower end Threadrippers" (still magnificent CPUs but just lower end in the Threadripper lineup) as ~their Ryzen counterparts. So a 7955WX is kind of a 7950x-ish CPU. The Threadrippers are a bit slower but broadly speaking I don't recall seeing benchmarks that would put them in a different class, its just that they probably boost a couple of 100mhz less than the Ryzen parts which technically is slower sure but not by a large margin. You do get more PCI-E lanes in return though

IMHO the closest competitor to the 3990x would be the 7970x / 7975WX which if I'm not mistaken is generally just a hair bit slower than the 3990x. Puget Systems did some testing (https://www.pugetsystems.com/labs/ar...s-generations/) and I'm sure you can find similar data elsewhere online including on the V-Ray's benchmark website (https://benchmark.chaos.com/vray/v6/...?id=816&id=810 - DO NOTE I'm not 100% sure if any of these are overclocked so treat the comparison as an estimate).

As for GPUs, depends on the types of projects / things you want to do imho. From a usability standpoint an important thing to consider is the VRAM buffer size. If your projects (or if you like having say Resolve open and rendering at the same time) take up lots of VRAM then it probably makes sense to go for the card that has a big enough VRAM buffer otherwise you'll be in for quite a rough ride. Broadly speaking, the 5xxx stack is imho kind of designed to push you to the more expensive 32GB GPU as 16GB can be a little on the low end for a lot of the CG work (and Nvidia probably wants to sell more 5090s and 5080s lol) although that might not be at all true for the types of projects you want to do.

I mean if you got by with a 2080ti then 16GB will probably be enough.

Hope my reply helps!

Hi,

Thanks for the extensive reply. I kind of bit the bullet on the processor more or less and I'll probably go for a 7970X, not sure why I would go with a 7955WX as the benchmarks are roughly the same.

As for the GPU, I'd like to go as high with VRAM as possible (within budget ofcourse), the 2080/3080Ti, weren't nearly enough for any of my projects and my experimentation with AI (LLM, ComfyUI), so I'll probably go for a 5090 if I can find a founders edition. It's a shame that Nvidia abandoned NVLINK for the regular consumer though so we can't stack memory anymore.

[Nejc Kilar]

Oh yeah, sounds like a smart choice to me if you're mostly concerned with rendering performance.

The WX parts are typically used when you need the PCI-E lanes and or the memory bandwidth. PCI-E lanes do come in handy if you have say 2 or 3 GPUs plus 10 NVME drives or whatever but if you don't then imho its just money wasted.
And the memory bandwidth doesn't really come into play for when it comes to rendering although I am not sure if that is the case for LLM training - I kind of recall hearing AI people opting to go with WX versions precisely because of that but I could also be misremembering things. I don't really do any AI dev to know better so... Don't listen to me haha

The 5090 has 32GB which is more than we ever head on consumer gear so I guess it should solid at the least. NVLink would be nice but personally I suspect it might be a casualty of product segmentation from Nvidia. Who knows for real though

[Vizioen]

Oh yeah, sounds like a smart choice to me if you're mostly concerned with rendering performance.

The WX parts are typically used when you need the PCI-E lanes and or the memory bandwidth. PCI-E lanes do come in handy if you have say 2 or 3 GPUs plus 10 NVME drives or whatever but if you don't then imho its just money wasted.
And the memory bandwidth doesn't really come into play for when it comes to rendering although I am not sure if that is the case for LLM training - I kind of recall hearing AI people opting to go with WX versions precisely because of that but I could also be misremembering things. I don't really do any AI dev to know better so... Don't listen to me haha

The 5090 has 32GB which is more than we ever head on consumer gear so I guess it should solid at the least. NVLink would be nice but personally I suspect it might be a casualty of product segmentation from Nvidia. Who knows for real though

I'm concerned about both actually, I want to have at least the rendering performance of the 3990x but with a smoother experience in my software for single threaded processes as well. In general I have:

1 GPU
1 x 2-4 TB NVME SSD for windows and software
1 x enclosure (high point SSD7505) with 2 X 4TB NVME SSD in RAID for all my working files (cloud synced)

So not sure if I need more PCI E lanes, my 2 workstations are doing quite well with this setup. I'm also not interested in this moment to train my own AI rather just use what's out there. But you never know it becomes more accessible.

As for NVLink, yeah they killed it so they can shift this to the more business side ($$$$$) of things, unfortunately.

[Nejc Kilar]

Hmm, I would say that with that kind of a setup you do not need a WS part.
To me it sounds like you're looking for a 7970x or a 7980x.

The 7970x will be ~3990x (sometimes a bit faster, sometimes a bit slower for multi threaded stuff but always faster in single threaded stuff, it really isn't a slouch) and the only reason I'm mentioning the 7980x is if you feel like spending more for another sizeable boost in multi threaded performance - single shouldn't suffer all that much going the 7980x route imho.

Then again there is quite a price difference so based on what you are saying to me it sounds like the 7970x would be a better fit.

The non WS Threadripper platform typically gives you 3x full PCI-E 5.0 slots plus 3x direct NVME slots. So I suppose technically with something like an ASUS Pro WS TRX50-SAGE WIFI you could have all three drives mounted directly on the motherboard (although the 3rd drive would be linked via chipset PCI-E 4.0 x4, so might be a tiny bit slower) or you could use the expansion card on the bottom most slot - that one would then run in PCI-E 4.0 mode but then again your expansion card is 4.0 so you don't really need 5.0 anyway. In this case the first slot would be used for the GPU.

The only thing to keep in mind with the 7970x is that it is a bit more difficult to cool. It's got 350W in a much smaller area than the 64+ core parts.

[Vizioen]

Hi thanks for the reply. Yeah indeed, if I had the budget I'd probably go for a 7980x as from the benchmarks I can see it performance quite well in single threaded situations too, but the price is quite steep and I'd rather invest in more GPU power as I already have an 3990x and 3970x so I'd have more than enough CPU power for my projects.

Both my threadrippers are cooled with a corsair h150i pro, not sure if that's the best idea, but so far I had no real issues besides when I tried overclocking with PBO, it's a little unstable and runs quite hot when doing so. (+/- 95°). If there would be a better way to cool maybe with PBO, I'd love to hear it.

[oscar lorenzo]

Check out this AIO, it covers 100% of the processor, and the radiator is made of copper, all other brands on the market are made of aluminum, which makes it cool a little more.
I have a 3970X and rendering in summer with an ambient temperature of 28º it doesn't go over 75

I have it mounted in push-pull with 3 noctua fans (https://noctua.at/en/nf-a14-industrialppc-2000-pwm) in push, and 3 Fractal Design fans (https://www.fractal-design.com/produ...i-hf-14/black/) in pull

https://www.aquatuning.com/es/refrig...ra-420-cpu-aio