Performance drop on mobile RTX4080

I have MATLAB 2023b with Parallel Computing Toolbox for running GPU-optimized (more or less) code with some FFTs and other operations on 5000sh x 1000sh matrices.

Typically, it takes <3s to run an iteration on Tesla V100 (or even RTX 3060). However, recently I migrated to a workstation with RTX4080 on it and found significant drop in performance of the same code (<30s).

I suspect that it has to do with the native support of the CUDA toolkit, I've noticed that gpuDevice says that Cuda Toolkit version is 11.8 however 40series might be natively support 12 and higher.

I can provide more details, but I wanted to check first if other RTX 40-series users faced similar performance drop.

Answer 1

[Joss Knight]

The 4080 is a good 10x slower than the V100 in double precision so this doesn't surprise me - it is designed for workstation graphics not HPC. If you're sure that the 3060 also massively outperforms it then you might have something - share some code and we can take a look.

The 4080 is fully supported by CUDA 11.8. Optimizations specific to Ada hadn't been completed by the time that toolkit came out so it's possible it might be executing some sub-optimal code, but usually that's more about it failing to reach its full potential rather than a significant under-performance.