CPU/GPU Architecture Recommendation

I'm the IT guy for someone running Matlab. He's primarily focused on the deep learning toolbox as I understand. Unfortunately I'm not sure how well his code is optimized, or really anything about it. I'm trying to build him a much beefier pc as his current model takes 12 hours to run. I put him temporarily on a 5 year old xeon with a GTX 1060 card and it dropped to 8 hours. What I'm trying to figure out generally is what is the best hardware lines to put him on without knowing much about what he's running.

For a CPU i9-9940x or a Intel Xeon Gold 6230

For a video card, a RTX 2080 or a RTX Quadro 5000.

Does either line in each case have something about it that would make it better in general for MATLAB over the other? Sorry for the lack of details, this is more of a general architecture question rather than a specific use case scenario.

Thanks.

[Walter Roberson]

It depends on whether they are using half-precision or single-precision or double precision. The fastest single-precision cards are not always the fastest double-precision cards.

It also depends on how much memory they need.

Also, will the system have another video card that can be used for actual video, or does the video need to be run by the NVIDIA GPU ?

[Chris Mibus]

I'm just going to throw like 128gb in for RAM. I doubt he'll use it all but he's been running out now at 32. I'd prefer to future proof it a bit. The PC will be setup for him to remote into so that the video card isn't used for monitors.

[Walter Roberson]

Extra video is important because there are a lot of Nvidia that do not support remote access computing through a remote desktop connection. The ones that do support remote access support TCC https://docs.nvidia.com/gameworks/content/developertools/desktop/nsight/tesla_compute_cluster.htm but a device in TCC does not support local graphics even for console.

Answer 1

[Joss Knight]

For Deep Learning the RTX 2080 is a good choice.

[Walter Roberson]

The RTX Quadro 5000 has twice the memory though.

[Joss Knight]

Sure, that card seems fine too, from its specs. The 2080 Super and Ti seem to be a bit faster, but the 2080 a bit slower. Extra memory is very useful though.The cost of the RTX 5000 seems reasonable, whereas the RTX 6000 and 8000 are more expensive.

[Chris Mibus]

What about a 4000? Doing a bit of speccing, it seems like the geforce cards aren't really available when you get a workstation with a xeon processor. I'd have to basically build a powerful gaming pc to go the geforce route which means going more with the i9 instead.

[Joss Knight]

At this point I'd just be reading the specs for you and guessing at what's important for your purposes. Basically go for the highest single precision TFLOPS you can afford, the largest memory you can afford.