engineering.fb.com — Triton Low-Level Extensions (TLX), is an extension to Triton that helps engineers optimize GPU performance while preserving Triton’s productive programming model. TLX introduces explicit war…
hgpu.org — As deep learning (DL) models become increasingly complex and hardware architectures grow more heterogeneous, DL compilers play a critical role in optimizing computational efficiency and enabling sc…
hgpu.org — Low latency is critical to user experience. As NVIDIA Blackwell and Rubin GPUs advance multi-die scaling, data placement and cross-die communication become increasingly important. However, the late…
ROCm Blogs — Profile an eight-layer training proxy derived from the DeepSeek-V4-Flash Primus configuration on eight MI350X GPUs with madengine and TraceLens, then turn operator-to-kernel insights into faster runs.
octomind — Flash Attention explained: what VRAM it saves, what stays in the KV cache, and exact Ollama, llama.cpp, vLLM and PyTorch settings, checked October 2026.
ROCm Blogs — FlyDSL mono kernels cut decode latency in ATOM, with MiniMax-M3 and DeepSeek-V4.1 results.
Aditya Kumar — The XLA review said that when you disagree with XLA’s fusion decisions, your recourse is to restructure your source or leave through custom_call. That is true of fusion. It isn’t true of the pipeline as a whole. XLA is a C++ compiler with an ordinary pass manager, and you can build it from source and add a pass. The pass itself isn’t the interesting part. Writing one forces you to learn how the pipeline decides it is done, and that matters more than any single rewrite.
NVIDIA Technical Blog — As artificial intelligence and accelerated computing continue to push the boundaries of what’s possible, the latest NVIDIA CUDA release arrives to unlock unprecedented performance and scale.
ROCm Blogs — Learn how cache and kernel optimizations raised GLM-5.2 AgentX throughput and interactivity on AMD Instinct MI355X.
The top AI news in your email delivered each morning. Choose categories