AI Briefing#cuda
The latest AI and developer news about #cuda, with the original source and a short summary.
The latest AI and developer news about #cuda, with the original source and a short summary.
A build method for using CUDA and ROCm simultaneously with llama.cpp on Windows was shared.
Built a compiler that lowers PyTorch graphs down to CUDA using 5,000 lines of Python and raw CUDA.
Starting with Ubuntu 26.04, Canonical will let users install CUDA and ROCm with just apt.
With a 2x RTX 3090 Ti setup, applying `-sm tensor` increased llama-bench throughput.
llama.cpp has added DKQ=320/DV=256 flash-attn CUDA support for Mistral Small 4.
Luce DFlash boosted Qwen3.6-27B throughput by up to 2x on a single RTX 3090.
The llama.cpp CUDA MMQ kernel improved speed in some ranges using fastdiv and dynamic tiling.
Hunyuan3D-Paint and Shape were ported to MLX/MPS for Apple Silicon, making texture generation 3x faster.
Moonshot released FlashKDA, which is up to 2.22x faster than Triton on H20.
TRELLIS.2 was ported to Apple Silicon's MPS, enabling 3D conversion without an NVIDIA GPU.
Ported TRELLIS.2 to run 3D generation on Apple Silicon without CUDA.
On the RTX 5060 Ti 16GB, Qwen3.6-35B-A3B ran about 10% faster than the prebuilt version.
Unweight reduces LLM weights by 15~22% on H100 while keeping outputs bit-exact.
llama.cpp's ggml added graph_reused to reduce CUDA's O(n) change checks.
Nvidia's moat lies not in the chip itself but in its power to control the supply chain and ecosystem together.
Cursor's multi-agent system made 235 CUDA kernels 38% faster in just 3 weeks.
AMD is strengthening its strategy to counter the CUDA ecosystem with ROCm.
A 32-bit overflow in vLLM's Mamba CUDA kernel caused logprob mismatches.
At GTC 2026, NVIDIA unveiled CUDA's 20th anniversary and DLSS 5, announcing its growth strategy driven by explosive AI computing demand.
HuggingFace has unveiled a new skill that helps AI agents write CUDA kernels directly.
The cause of rare gibberish was traced to vLLM's Mamba scheduler and cache.
Hugging Face has released the kernel-builder library, which supports everything from developing to deploying custom CUDA kernels.
Hugging Face has unveiled Kernel Hub, which lets developers easily load optimized compute kernels to boost model performance.
Explains how to visualize and analyze GPU memory usage using PyTorch's memory snapshot feature.