eGPU set up for deep learning and compute tasks on a small laptop

Gaming and Visualization Technologies General Topics and Other SDKs GPU - Hardware
, ,
1

Hi,

I was wondering what experience anyone has setting up an eGPU for use with Nvidia graphics cards under various scenarios. My laptop (Dell Precision 3470 32 GB RAM Intel i5-1250P integrated graphics only) requires an external GPU for DL training hardware acceleration. It has Thunderbolt 4 and would suggest an external GPU could be possible on Windows 10 (or above). Various enclosures are plug-and-play if Nvidia graphics drivers (and CUDA/CuDNN for DL/machine learning) are set up properly. An example is Razer Core X - Mercury that has sufficient power requirements for many GPUs (usually used for better gaming graphics).

Also, could an older (cheaper) computer Tesla card be used for accelerated (scientific) computing, but I expect that is unlikely due to software/driver configuration/thermal requirements (usually only used in Linux HPC servers)? It appears that the eGPU enclosure is only certified for GTX/RTX/some Quadro cards.

Finally the latest releases of Tensorflow for GPU requires Linux so as the Mercury is only certified for Windows or Mac, it may be more tricky to set up in Ubuntu (e.g. 20.04 LTS). A NGC docker container is likely to be used. It is preferable to visualise output on the laptop rather than an external monitor (via the eGPU enclosure).

I know many use GPU cloud computing instances for DL training or compute nodes (and may ultimately be the better/easier option here), but it is interesting to find out if it can be done locally using an eGPU.

2

Hi there @DL_Chem_GPU and welcome to the NVIDIA developer forums.

I am using a Razor Core (some older version) myself with an RTX 3080. So far I only used Desktop grade PCIe GPUs in the enclosure and it is indeed recommended. But I did use it for DL tasks as well. And if i remember correctly it was also recognized at least for compute in an old Ubuntu 18.04 setup I had. But I don’t recall if i needed a special installation process for that. I think I did not since TB should map it as another PCIe device for the OS.

Memory transfer might seem a bottle-neck, but the max PCIe 4x you reach through eGPU should be enough if your DL models don’t depend on data streaming.

But I would be very interested to hear your own experiments and experience with this! Please share!

Thanks!