StarNet++ GPU Acceleration for Non-NVIDIA Devices

[wliphotos]

As far as I know, hardware acceleration of StarNet++ has only widely been done with CUDA-enabled GPU's, i.e. NVIDIA video cards. After sharing my updated take on DarkArchon's classic StarNet CUDA tutorial today on the Astrobiscuit Discord Server, one user lamented not being able to use their AMD Radeon 5700XT to accelerate StarNet. I did a bit of digging, and figured out that there is indeed a way to accelerate StarNet with non-CUDA cards — through Microsoft DirectX 12 and Microsoft DirectML. In fact, any graphics processor that supports DirectX 12, including integrated graphics, can theoretically be used to accelerate StarNet, as basically every graphics card for at least the last five years supports DX12 — AMD, Intel HD, even Qualcomm. I've tested it with my NVIDIA card in non-CUDA mode and it does indeed offer a significant performance improvement over stock CPU processing, but still falls short of CUDA. In addition, DirectML doesn't yet support TF2.x, so it unfortunately doesn't work with StarNet V2. However, given the performance benefits, I think that it's worth sticking with StarNet V1 while we wait for a new version of DirectML that supports TF2.x (I'll be sure to update my tutorial when that comes out). I've written a short tutorial which can be found here:


https://www.williamliphotos.com/gpu-starnet-dml


Thanks for reading, and clear skies,

WL