P40 in a workstation computer

[Provable7965]

Hey guys, I posted a few months back about using those cheap used Nvidia server class GPUs in a workstation computer. I finally completed my build, and I am proud to announce that I have managed to use an Nvidia P40 for my workstation-oriented PC.

 

Other than that, I used:

 

• CPU: Ryzen 7700X
• Mainboard: Asus Prime B650 Plus
• RAM: 32GB DDR5 Crucial @4800
• GPU Cooler: Morpheus 2 core edition
• SSD: 2TB Kingston M.2 SSD
• Power supply: 700W BeQuiet! System Power 9
• PCIE power to CPU adapter for the GPU
• Anti-sag support for the GPU
• Zalman S2 case
• Alpenföhn Ben Nevis CPU cooler
• 3 extra case fans + 2 for the GPU cooler (P12 Slim)

I also had to buy some small heatsinks for the VRAMs on the back of the GPU.

 

So far, the system is very performant (I can post some benchmarks if needed later, but 60+ FPS for Minecraft with RTX shaders, max render distance, all settings ultra, speaks for itself, especially since it's a 200$ GPU. Of course, the total price of the GPU is closer to 300$, since the cooling solution, as well as the adapter cable, totals around 100$.

 

I have connected the GPU fans directly to the mainboard, where I use FanControl to set fan curves - even under a lot of stress (Furmark, Blender render, Stable diffusion) the card never exceeds 55°C with around 55% fan speed.

 

Please let me know if you have further questions regarding my build or any complaints about my configuration.

[igormp]

One thing that I'm curious about: why a P40? Is a 1080ti more expensive than it where you live? Otherwise it should net you more raw performance, albeit with less vram (which I believe you're not really making full use of with only 32gb of ram).

 

What's going to be your main use case for this desktop, other than the games you listed?

[Provable7965]

17 minutes ago, igormp said:

One thing that I'm curious about: why a P40? Is a 1080ti more expensive than it where you live? Otherwise it should net you more raw performance, albeit with less vram (which I believe you're not really making full use of with only 32gb of ram).

Yeah, a 1080ti is more expensive, but also does not come with the 24GB of ECC VRAM, which I ...

18 minutes ago, igormp said:

What's going to be your main use case for this desktop, other than the games you listed?

... need for running scientific computations and training ML models, as well as for inference using stable diffusion.

[igormp]

1 hour ago, Provable7965 said:

... need for running scientific computations and training ML models, as well as for inference using stable diffusion.

Wouldn't a 3060 or even 2060 12gb (or the Super model) be better, or are those also more expensive?

 

Even though they only have 12gb of vram, they are also way faster than the p40, specially if you're using FP16, which easily makes them twice as fast and only uses half of the vram.

[Provable7965]

On 9/1/2023 at 7:05 PM, igormp said:

Wouldn't a 3060 or even 2060 12gb (or the Super model) be better, or are those also more expensive?

 

Even though they only have 12gb of vram, they are also way faster than the p40, specially if you're using FP16, which easily makes them twice as fast and only uses half of the vram.

They don't have 24GB of VRAM, which I need, or do they (are there frankenstein'd cards that one can buy)?

[Provable7965]

On 9/1/2023 at 7:05 PM, igormp said:

Wouldn't a 3060 or even 2060 12gb (or the Super model) be better, or are those also more expensive?

 

Even though they only have 12gb of vram, they are also way faster than the p40, specially if you're using FP16, which easily makes them twice as fast and only uses half of the vram.

Also they don't support ECC in their VRAM.

[igormp]

50 minutes ago, Provable7965 said:

They don't have 24GB of VRAM, which I need, or do they (are there frankenstein'd cards that one can buy)?

Sadly no Frankensteins unless you ask to build your own, which by the cost would be way higher

 

As I said, those 12gb work as good as 24gb for Fp16 training/inference while also being way faster than that p40. I can't comment on your non-ML needs tho, so for that it might make more sense. 

51 minutes ago, Provable7965 said:

Also they don't support ECC in their VRAM.

I mean, is that really relevant for training sessions? Most people I know just turn it off for that small speed bump.

[Provable7965]

On 9/3/2023 at 3:17 PM, igormp said:

Sadly no Frankensteins unless you ask to build your own, which by the cost would be way higher

 

As I said, those 12gb work as good as 24gb for Fp16 training/inference while also being way faster than that p40. I can't comment on your non-ML needs tho, so for that it might make more sense.

Only if the software supports Fp16, which some models (even for stable diffusion) don't.

 

On 9/3/2023 at 3:17 PM, igormp said:

I mean, is that really relevant for training sessions? Most people I know just turn it off for that small speed bump.

Not for the training sessions, but for simulation, definitely.

[igormp]

3 minutes ago, Provable7965 said:

Only if the software supports Fp16, which some models (even for stable diffusion) don't.

 

You can always make them support it haha

3 minutes ago, Provable7965 said:

Not for the training sessions, but for simulation, definitely

Yeah, fair enough, I'm not really knowledgeable when it comes to other simulation and HPC stuff.

[santasan]

Thanks for sharing. I was in the process of building a server with Nvidia P40. I have mostly used desktops or laptops at home. Do these special builds work fine with US 110V power plug outlets, or do you need 240V power outlet.  I was eyeing something like the following: Super Micro 1019GP-TT with Nvidia Tesla P40 for GPUdirect using CUDA.