Search - metacurate.io

CUDA Oxide 0.3 Vastly Improves Rust Support For NVIDIA CUDA Programming - Phoronix

CUDA Oxide 0.3 updates its Rust-to-CUDA compiler, adopting a stable Rust, unifying runtimes, introducing CUTLASS CuTe via cute-rs, enabling end-to-end models on Blackwell GPUs, and boosting debugging, tracing, testing, and GPU-specific optimizations.

1 min read Listed 14 days ago in Bluesky

www.franksworld.com Blog post

How CUDA Became the Unbreakable Lock on AI Progress – Frank's World of Data Science & AI

CUDA began as a costly gaming-card feature in 2006, later becoming essential to AI via easier GPU programming and widespread framework support, powered by strategic design and persistence. The story also notes ongoing efforts to bypass CUDA and the shift by major players toward alternative, integrated systems.

3 mins read Listed 30 days ago in RSS: Frank's World of Data Science & AI

github.com Documentation

GitHub - NVIDIA/cuda-rust: NVIDIA's CUDA platform for Rust. Host runtime crates plus Tile (cutile-rs) and SIMT (cuda-oxide) kernel programming models in idiomatic Rust. · GitHub

CUDA Rust provides Rust-based tools to program NVIDIA GPUs, offering two kernel models (Tile and SIMT) with accompanying runtime and book-backed documentation; it is an early-stage project with ongoing development and potential API changes.

4 mins read Listed 15 days ago in X

gazewire.com News

Together AI's Compiler Automates GPU Kernel Tuning — GazeWire

Together AI unveils a compiler that uses AI coding agents to auto-tune GPU kernels, with a human-in-the-loop for final release. It claims dramatic reductions in initialization time (467s to 9s) and substantial but unverified throughput gains, while noting ongoing model turnover requires continuous tuning.

3 mins read Listed 9 days ago in Bluesky

github.com Tools and repositories

Release b12284 · ggml-org/llama.cpp · GitHub

This llama.cpp release adds OCP FP8 support to its CUDA backend, including FP8 matrix operations, attention, KV-cache handling, and data copies. It also reports Qwen 27B benchmark results on an NVIDIA RTX PRO 4500 Blackwell GPU.

3 mins read Listed 3 days ago in RSS: Release notes from llama.cpp

arxiv.org Scientific research

[2611.09280] An Empirical Study of Issues in GPU Kernel Libraries for LLM Systems

An empirical study of 300 GitHub issues across seven GPU kernel libraries finds that most reported problems involve build, compiler, or runtime environments. It also identifies hidden constraints in kernel specialization and gaps in testing for numerical and hardware-specific failures.

2 mins read Listed 5 days ago in Bluesky

autonainews.com News

GPU Kernel Optimization Gets 3.7x Boost via Together AI

Together AI launches the Together Compiler to automatically rewrite GPU kernel code using AI agents, claiming large throughput gains on AMD hardware but with an enforced human sign-off before production; the project inaugurates the Polymorphic Inference initiative aimed at continuous, agent-driven optimization.

3 mins read Listed 9 days ago in Bluesky

github.com Tools and repositories

GitHub - theIvanR/llama-on-legacy-gpu: High performance llama cpp patches and fork for legacy gpus · GitHub

A patch set for llama.cpp aims to extend CUDA support and improve inference on legacy NVIDIA GPUs, with Windows NCCL experiments and multi-GPU tests showing preliminary throughput gains on Kepler-era hardware.

6 mins read Listed 21 days ago in reddit/LocalLLaMA/High-Performance llama.cpp fork for Legacy GPUs: Kepler Support, CUDA Graphs & Windows NCCL!

thelastone07.github.io Tutorial

Journal

The piece outlines a practical, stepwise approach to learning CUDA, emphasizing hardware and compiler concepts, PTX/SASS, and NVIDIA H100 architecture. It blends theory with plans for hands-on topics like writing kernels, CUTLASS, inference engines, and open-source work, framed by the goal of making GPU programming more approachable.

35 mins read Listed 18 days ago in reddit/learnmachinelearning/How I would get started with CUDA

github.com Tools and repositories

Release b11879 · ggml-org/llama.cpp · GitHub

The document details CUDA optimizations for fusing FFN gate/up matmuls and GLU in MMQ, describing interleaved row tiling, loader data layout, and write-back, with compatibility notes for specific NVIDIA architectures.

1 min read Listed 20 days ago in RSS: Release notes from llama.cpp