GPU Support for RTX 4090

The GPU Computing Requirements documentations states that 3.5 to 8.x CUDA capability is supported.

Since the recently released RTX 4090 has a CUDA capability of 8.9, does that mean it is supported by default?

If not, when do we get support for the Ada Lovelace architecture?

Answer 1

[Joss Knight]

Forgive me for needing to correct Walter, but the last three versions of MATLAB will natively support the 4000 series because, as you say, they are within the 8.x architecture series. Hopper (H100) will not be supported natively.

Answer 2

[Walter Roberson]

No, the last few versions of MATLAB will not support the new RTX 40x0 architecture -- not by default.

You need to use parallel.gpu.enableCUDAForwardCompatibility to give MATLAB permission to use NVIDIA's backwards-compatibility feature. When enabled, each GPU kernel will be automatically recompiled for use with the new device. Sometimes the operations work fine, just not as fast as they might have executed if up-to-date code had been used. Other times, however, the recompiling will introduce bugs... in which case you are fortunate if the code bombs, as at least that tells you your code failed (worse is if the bugs give you wrong answers without telling you something went wrong.)

[Abdalla Rashed]

So, when does the support usually become available after a new architecture release? Like a rough estimation based on previous trends?

[Walter Roberson]

Typically 3rd to 4th MATLAB release after the architecture release, but sometimes longer.

R2022a did not yet support the RTX 30x0 series; I have not been able to find out yet whether R2022b added that support. The RTX 30x0 series was released September 17, 2020, which was the day after R2020b was released, so the third release after, R2022a, did not yet have support.

There were some NVIDIA bugs in the libraries for the 30x0 series, that would have taken extra time to debug and fix, so there was some delay relative to the typical schedule. It is too early yet to know if there are RTX 40x0 related bugs in NVIDIA's libraries.

[Abdalla Rashed]

RTX 30x0 is already supported since R2021a though

[Walter Roberson]

See https://www.mathworks.com/matlabcentral/answers/1814255-gpu-coder-basic-code-generation-failed?s_tid=srchtitle for someone with an RTX 3090 who could not use R2021b.

[Joss Knight]

The Ampere series including the 30x0 have indeed been supported since R2021a, see https://uk.mathworks.com/matlabcentral/answers/592198-does-matlab-support-nvidia-ampere-cards-for-gpu-computation .

That GPU Coder problem is I believe because by default Coder tries to compile native binary for your card's specified compute capability, e.g. sm_86, which isn't supported by the compiler shipped with MATLAB. The user would either need to change their settings to compile for sm_80, or install a newer CUDA toolkit.