Can Matlab 2020 a or b support RTX 30 series GPU for deep leaning?

Can Matlab 2020 a or 2020 b support RTX 30 series GPU for deep leaning? Since the GPUs are newly relased.

Answer 1

[Ameer Hamza]

It is not fully supported yet. These GPUs are based on Ampere Architecture, and documentation: https://www.mathworks.com/help/parallel-computing/gpu-support-by-release.html shows that support for Ampere is still limited. You can expect to see errors and unexpected behaviour.

[Edric Ellis]

More details in this answer: https://www.mathworks.com/matlabcentral/answers/592198-does-matlab-support-nvidia-ampere-cards-for-gpu-computation

[Yuanlong Zhang]

Hi Ellis,

Would it be possible to get 2020b support new RTX30s GPU soon? I think many reaserchers have ordered those new cards but without matlab support is a really a pain.

Best,

Yuanlong

[Walter Roberson]

Would it be possible to get 2020b support new RTX30s GPU soon

That is unlikely. Mathworks historically has provided support for new GPUs in new releases, not as a bug-fix to old releases.

[Jast]

This is indeed somewhat stupid. Come on Mathworks, you should know better that all the DL people are flocking to the Ampere architecture.

[Walter Roberson]

Jast: do you think that Mathworks should be writing high-speed deep learning CUDA libraries itself, in competition with NVIDIA? Or should Mathworks be spending its time pressing NVIDIA to fix the high-speed libraries that NVIDIA provides?

Reminder: writing those kinds of high speed libraries requires very deep knowledge of the board architectures -- knowledge of what the board really does, not just abstractions about it should do. The kind of knowledge to say "Errr... there is a missing gate, so if you execute this particular instruction on the 7th decoder, and the data happens to have this particular bit-pattern, that the third-LSB sticks high unless you happened to have executed this kind of instruction before-hand."