Excited to announce!

@gaunernst - this one in case you want to beat apple's Accelerate

@__tensorcore__ agree with this, but heavily templated code further complicates debugging. I remember one of your devs mentioning that this is (one of) the main reason why using cuda-gdb with CUTLASS is not recommended

The modern NVCC compiler is so advanced that handwritten SASS code will likely be on par with or slower than the code generated by the compiler (assuming well-written CUDA/PTX code)

@dhtikna @typedfemale

RT @awnihannun: Reminder, many institutions outside of the US and China are building amazing foundation models. many-polar world (aka even…

@awnihannun @MistralAI @MiniMax__AI @kyutai_labs @SakanaAILabs @nx_ai_com in Austria

@AnushElangovan 👍👍👍👍

@__ReJ__ @AgileJebrim

@telmin_orca

NVIDIA RTX Blackwell whitepaper is out!

@PytorchToAtoms @giffmana @cHHillee @PytorchToAtoms btw. it depends on how you perform benchmarks: with locked stable clock or unlocked. in the latter case, sure, your results will be affected by the number of iterations, matrix size, etc. as your clock speed varies due to power limits

@PytorchToAtoms @giffmana @cHHillee yes, but it gives you relative performance among the generated kernels. you can then pick the best cutlass kernel and test it however you like

@giffmana @PytorchToAtoms @cHHillee I believe it is what @cHHillee meant by 'autotuning'