GPU recommendation for Deep Learning and AI

Hi, I would like to invest in a NVIDIA GPU to faciliate the deep learning / AI work with Matlab. I considering Tesla V100 32Gb and Quadro RTX 8000 48Gb. Your feedback would be greatly appreciated. Kind regards,

[Abolfazl Nejatian]

Dear Mariusz,

i have install Tesla V100 in my server last week and run gpuBech.

i hope results are usful for you.

System Information

OSTypeLinux

OSVersionbuildd@lcy01-amd64-023

Name:2 * Intel(R) Xeon(R) Gold 6134 CPU @ 3.20GHz

MATLAB Release: R2019b

GPU

NameTesla V100-PCIE-32GB

Clock1380 MHz

NumProcessors80

ComputeCapability7.0

TotalMemory31.75 GB

CUDAVersion10.1

DriverVersion430.50

[Walter Roberson]

Nice performance !

Answer 1

[Joss Knight]

The Tesla V100 is a passively cooled device only suitable for servers. Is that available to you? The Quadro card you indicate is mainly good for serious workstation graphics and rather overpriced for compute. Generally the Titan RTX would be your best bet for deep learning workflows, or a GeForce GTX 2080 for something slightly cheaper.

[M.G.]

I also need to make a decision on which GPU to buy (for deep learning) and I need an answer to the following question to be able to determine which GPU I should buy.

Does Matlab use single-precision computations for deep learning by default (specifically, the Reinforcement Learning Toolbox) ?

I know that Matlab generally uses double-precision computations, and I noticed that most graphic cards (RTX 2000 and 3000 series for example) that have high single-precision computation speeds have rather slow double-precision speeds (1:32). Only super expensive GPUs (like V100 and A100) maintain high speed computations for double-precision (1:2). So, if Matlab defaults to single-precision, I can just buy a relatively cheap card and get good performance on the Reinforcement Learning Toolbox. Otherwise, I'd have to buy a much more expensive card to be able to get high speed on double-precision.

I'd really appreciate some help here :D

[Joss Knight]

Yes, single precision is the default for deep learning workflows.