ptx

PTXBench is introduced as a benchmark to evaluate and adapt large language models for optimizing GPU kernels using architecture-specific PTX, showing uneven performance and fine-tuning insights.

An educational tweet explaining the three levels of NVIDIA's software stack (CUDA C++, PTX, SASS) and how CUDA's abstraction creates a moat, while mentioning Luminal's automatic compiler search.

Claude Fable 5 achieves top results on KernelBench-Hard by hand-writing PTX code for B200 fp8 GEMM, outperforming other models and reaching 44-59% of peak performance on compute-bound shapes.

This article explains the entire process of compiling and launching a CUDA kernel, from source code to hardware execution, using a simple vector addition example and detailing the role of nvcc, PTX, SASS, and ioctls.

INT21 announced PTX Kernel Factory, a self-improving agent swarm that autonomously generates expert-level PTX GPU kernels, with open-source proof-of-concept implementations and beta access.

CUDA 13.3 introduces significant enhancements including Tile C++ support, C++23 standard, improved NVRTC, stable CUDA Python 1.0 APIs, and PTX 9.3 with new fabric instructions and async multimem operations, targeting kernel developers and runtime engineers.