Does CUDA PDL support overlapping kernels on consumer GPUs under Windows WDDM?

[IlyaGrebnov]

I’m trying to get PDL working with both the CUDA runtime and driver APIs, but I haven’t been able to get my kernels to overlap. There doesn’t seem to be much information available on PDL’s constraints, so I’m asking here. I’m on Windows 11 with CUDA 13.0 and an RTX 5090, and I suspect WDDM might be the limiting factor.

My question is: does PDL support overlapping kernels execution on consumer-grade GPUs on Windows when using the WDDM driver?

[bernhardmgruber]

@pciolkosz do you happen to know?

[IlyaGrebnov]

@pciolkosz do you have any insights on this?

[pciolkosz]

I asked internally and there are certain limitations of WDDM that make it incompatible with PDL and stream workloads. It should be supported in CUDA graphs

[IlyaGrebnov]

Thank you for confirming! I would appreciate it if these limitations were documented as well. was using this page to find the details: https://docs.nvidia.com/cuda/cuda-programming-guide/03-advanced/advanced-host-programming.html#programmatic-dependent-kernel-launch