Using RTX 5090 for GPU computing in forward compatibility

Dear users,

I have read a couple of very useful posts

https://www.mathworks.com/matlabcentral/answers/2173867-can-i-use-my-nvidia-blackwell-architecture-gpu-with-matlab-for-gpu-computing

https://www.mathworks.com/help/parallel-computing/gpu-computing-requirements.html

which explain that Matlab currently does not fully support Nvidia gpus with compute capability greater than 9.0.

This means that RTX 5090 (or any GPU based on the new Blackwell architecture) is not fully supported natively, but it can be enabled with forward compatibility as follows:

parallel.gpu.enableCUDAForwardCompatibility(1)

However, Matlab documentation also says that "Enabling forward compatibility can result in wrong answers and unexpected behavior during GPU computations." (please see second link posted above).

So I would like to ask users who own a NVIDIA RTX 50XX what is their experience and how well forward compatibility works in practice. This feedback will be quite useful because I am currently considering upgrading my GPU to a new 5090 but I wonder if I should wait a bit until a future realease provide full native support.

Thanks in advance for comments/feedback etc.!

[JP Norair]

I have not had great success in utilizing my RTX 5080. It works, but the performance isn't what I would expect. Like many, I'm waiting for 2025b. Not even until very recently does Tensorflow support Blackwell (and it still requires a custom build), so you can offload certain workflows to Python from Matlab via the Python interface. I have done this to an extent.

[Alessandro]

Thanks for your feedback, @JP Norair. I had the same impression as you.

[Alessandro]

@JP Norair: have you had a chance of trying R2025b? Any better with Blackwell?

Answer 1

[Walter Roberson]

@Joss Knight who is Mathworks Staff with the Parallel Computing Toolbox, recently wrote https://www.mathworks.com/matlabcentral/answers/2178660-nvidia-gpu-geforce-5060-5090#answer_1568337

"The practical answer is that I haven't yet seen any problems with MATLAB R2024b and R2025a running on Blackwell in forward compatibility mode."

Answer 2

[Eric]

The MathWorks documentation explicitly states that the parallel.gpu.enableCUDAForwardCompatibility(1) function can enable compatibility with new hardware, such as the NVIDIA Blackwell architecture.