Using external GPU with MatLab 2020a

Hello, I am using an external TITAN V GPU housed in an AKiTiO brand expansion chassis that connects to the laptop through Thunderbolt 3. I am using MatLab 2020a and I have installed the CUDA versions and the recommended driver for the TITAN V GPU. , matlab if it recognizes the gpu, but when simulating a CNN, the training time is very slow and is almost the same as the one generated with the internal GPU (GTX 980M) that my laptop has.

Why is this happening? How could I solve it?

See if someone from Matlab Staff can help me.

Thanks for your attention.

Cheers.

Answer 1

[Joss Knight]

There may be nothing you can do. Data transfer times to an external GPU are much higher. It might help to increase the MiniBatchSize training option, or enable DispatchInBackground.

Of course, it's plausible that your network is very small and not using the GPU efficiently, so all you're seeing is the same overhead and the device makes no difference. What is the input size to your network?