Recommendation for multi-GPU setup

[harshitaneja]

Building a workstation with threadripper 2950X(need the cores) and 3 GPU- 1x 2080Ti and 2x 2080.

Mostly for data science work. NVLink is not needed as work will be parallelized across multiple GPUs. 

 

Need help with GPU recommendations(CUDA support required) and motherboard. I need to live in the same room as this device which will run 24x7 so would need to really quiet. Water or air cooling? I live in India where it is tough to find people to help with water cooling(at least that I know of) so would be doing it myself. Is IChill Black 2080Ti a good option?

 

Any suggestions or recommendations or pointers to places I could ask or read at. Thanks a lot for the help.

[Princess Luna]

What workload are you going to do exactly? you might be better served with a single TITAN V due to the increased Tensor and Cuda Core count alongside driver specific optimizations and certification.

 

It is not uncommon to see the TITAN V sweeping the floor with a 2080 Ti on hardware acceleration workloads and there's water blocks for it to make up the loud FE cooler.

 

The thing is, depending on what you'll do a single TITAN V may perform on pair with a 2080 Ti + 2x 2080 and if that's the case you might not even need the extra PCI-e Lanes meaning you could further save up money on an i9 9900K instead of a TR 2950X build.

[GoldenLag]

11 minutes ago, Princess Cadence said:

TR 2950X build.

Or just get a 1920x as it seems he might only need PCIe lanes. 

[harshitaneja]

7 minutes ago, Princess Cadence said:

What workload are you going to do exactly? you might be better served with a single TITAN V due to the increased Tensor and Cuda Core count alongside driver specific optimizations and certification.

 

It is not uncommon to see the TITAN V sweeping the floor with a 2080 Ti on hardware acceleration workloads and there's water blocks for it to make up the loud FE cooler.

 

The thing is, depending on what you'll do a single TITAN V may perform on pair with a 2080 Ti + 2x 2080 and if that's the case you might not even need the extra PCI-e Lanes meaning you could further save up money on an i9 9900K instead of a TR 2950X build.

Thanks for the response. Titan V is an older card which has similar if not worse performance in certain cases in comparison to 2080Ti in FP16 workloads. Even in FP32 difference isn't large enough to warrant paying around 2.5x for it. I think you are confusing it with Titan RTX. Even though Titan cards have 24GB of memory in comparison to 11GB of a single 2080Ti. But total memory of all three planned GPUs 11+8+8 (GB) should be sufficient to counter that(now I know I am comparing a single memory with sum of multiple memories which is not a straightforward comparison) without much of an increase in price. With this given that my work mostly is parallelizable I can run it across all the three at the same time which should give higher performance and if not, it offers the benefit of running different models on all the three GPUs. 

I need the 16 cores offered by 2950X(well I wanted the 32 core 2990WX but couldn't justify the cost) as I need to run a lot of virtual machines and some of statistical work is done using libraries which don't yet have GPU options and thus need the cores for that. So PCIe lanes are not an issue.

Thanks again for your input.

[harshitaneja]

6 minutes ago, GoldenLag said:

Or just get a 1920x as it seems he might only need PCIe lanes. 

Maybe I should have mentioned it in the answer, 2950X is for the extra cores as some of my work involves running statistical libraries which are currently CPU only. Extra PCIe lanes are cherry on top.

[GoldenLag]

5 minutes ago, harshitaneja said:

-snip-.

What workload specifically?

 

If it can use OpenCL you will have much more success with tripple Radeon 7 cards.

[harshitaneja]

1 minute ago, GoldenLag said:

What workload specifically?

 

If it can use OpenCL you will have much more success with tripple Radeon 7 cards.

CUDA is sadly required as tensorflow and certain others only support it which is bad as I really don't want Nvidia being a linux user and their abysmal driver support. I have edited the question to include CUDA support.

[__orange]

For graphics cards, AIO water cooling cards are recommended if you have enough space to mount radiators. Radial fan cooling card is second choise. Never use axial fan colling card in multi-card setup.