Most exciting addition in CUDA 12.9 for me is CUDA_LOG_FILE. You can finally get error strings to describe the error you received from a CUDA API call in more detail than a generic CUDA_ERROR_INVALID_VALUE! docs.nvidia.com/cuda/cuda-c-pr…

🚨🔥 CUTLASS 4.0 is released 🔥🚨 pip install nvidia-cutlass-dsl 4.0 marks a major shift for CUTLASS: towards native GPU programming in Python slidehelloworld.png docs.nvidia.com/cutlass/media/…

We believe low level access to hardware is extremely important. High level generators rob away the freedom of programmers to experiment with new ideas and kernel designs while C++ is too slow to compile, learn, and debug. CuTe DSL provides the best of both worlds ⚡

I love Cutlass, and this new Python DSL looks very well-designed. Will for sure accelerate kernel dev + exploring new ideas in ML + GPU. I'm already playing with it and having fun

timelapse #58 (14.5 hrs): - used cutlass python DSL to increase elementwise add/mul memory throughput (from pytorch 500GB/s to cutlass 850GB/s) - diving into cutlass 4.0 (minus tile abstractions) - cuda book design decisions with maharshi (महर्षि) - restructure of 5 chapters -

Introducing soarXiv ✈️, the most beautiful way to explore human knowledge Take any paper's URL and replace arxiv with soarxiv (show in video) to teleport to its place in the universe I've embedded all 2.8M papers up until April 2025 Try it at: soarxiv dot org

Every GPU kernel writer in shambles

did i mention that this is totally nuts?

We've been thinking about what the "ideal" architecture should look like in the era where inference is driving AI progress. GTA & GLA are steps in this direction: attention variants tailored for inference: high arithmetic intensity (make GPUs go brr even during decoding), easy to

Another 🔥 blog about CUTLASS from Colfax International, this time focusing on the gory details of block-scaled MXFP and NVFP data types and Blackwell kernels for them. research.colfax-intl.com/cutlass-tutori…