Tag
作者发布了一个免费、代码优先的 GPU 内核工程系列课程（共 44 节）的介绍与路线图视频，从朴素 GEMM 讲到 FlashAttention，覆盖 Hopper 与 Blackwell 架构，并以 roofline 模型和数据移动为核心教学理念。
A free "GPU Kernel Engineering" course series of 44 lessons has been released, starting from CPU parallel computing and CUDA basics, and progressively diving into memory optimization, GEMM, Tensor Cores, Flash Attention, and all the way to cutting-edge hardware like H100/Blackwell and AI-written kernels.
PyTorch Helion now integrates NVIDIA CompileIQ's Advanced Control Files into its autotuner, enabling joint search over kernel configurations and low-level ptxas compiler controls to extract last-mile performance from already autotuned GPU kernels.
doubleAI 发布 Maxwell，首个“代码执行模型”，可在程序运行时重写二进制以适配特定硬件与数据，并通过形式化证明保证优化等价，带来 38× 数据管线加速与 4.25× AI 推理内核加速。
斯坦福 CS336（2026 春）第 6 讲的笔记，讲解如何通过减少 GPU 内存往返来优化内核：对比 softmax 多次遍历与融合行内核的内存流量，介绍 Triton 的块级编程模型，并强调先用 CUDA event 测量基线、再做内核融合与分块优化。
介绍 SentryWard：一个开源的 GPU kernel 形式化验证工具，可证明同步安全、无数据竞争与资源生命周期正确性，并通过系统性移除冗余同步使 NVIDIA B200 上的 kernel 最快提速 1.4 倍。
David Bellamy shares a concise curriculum for learning core RL and inference systems ideas, citing the HF Ultra-Scale Playbook, Baseten's inference engineering guide, Lambert's RLHF book, and PMPP for kernels, supplemented by blog posts.
DeepSeek 开源了其训练模型所用的 GPU kernel 库 DeepGEMM，将 FP8/FP4/BF16 GEMM、融合 MoE 通信重叠、MQA 打分等核心算子统一到单一 CUDA 代码库中，通过 DeepJIT 运行时编译，官方称性能可达 H800 上 1550 TFLOPS，并已推出昇腾（Ascend）版本。
NVIDIA的Guray Ozen将在PyTorch Conference North America 2026上展示CUTLASS for Python的最新进展，包括CuTe扩展、任务调度器和编译器诊断等面向开发者与AI代理的高效GPU内核编写新功能。
A tweet praising Hugging Face Kernels (hf.co/kernels) as a great product, showcasing its catalog of community GPU kernels filterable by platform (CUDA, ROCm, Metal, WebGPU, CPU) and hardware (H100, A100, MI300, Apple Silicon, etc.), including popular entries like flash-attn2, nvfp4-gemm, and ggml quantization kernels.
MetaKernelBench introduces a benchmark measuring whether experience distilled by GPU kernel optimization agents in one DSL (CuTeDSL) transfers to another (TIRx) on the same problem, finding mixed results with lift ranging from -19% to +29% across six models.
该论文提出用短多项式（degree-3/4 BF16）替代 LLM 中原生的 sigmoid、tanh 和 SiLU 超越函数求值，在 NVIDIA GB200 上将完整训练步吞吐提升最高 8.0%，同时保持相近的训练损失和模型精度。
At PyTorch Conference North America 2026, AMD's Liz Li and Jiahui Cao will present "Extending TorchInductor with FlyDSL," an MLIR-native GPU kernel backend for high-performance GEMMs that integrates with TorchInductor's compilation and autotuning pipeline, with benchmark comparisons against Triton on AMD Instinct GPUs.
A hand-written GEMM kernel for NVIDIA Blackwell (B200), built with CUTLASS's CuTeDSL from raw primitives (tcgen05, TMA, mbarriers, TMEM), goes from 164 to 1401 TFLOP/s — about 97% of cuBLAS — with one key change doing most of the work and prefill vs decode workloads yielding different optimizations.
PyTorch/Meta blog reveals Jagged Flash Attention built with TLX (Triton Low-level Extensions) on NVIDIA Blackwell B200, using ~3.2K lines of Triton code—3x less than FlashAttention-4—while outperforming FA4 by ~13% forward and ~50% backward on jagged attention shapes used in Meta's Generative Ads Model.
An in-depth explainer on how LLM inference really works: the prefill/decode loop, KV caching, how file encodings and quantization labels (e.g. NVFP4) determine which kernels a GPU actually runs, demonstrated with Qwen 3.8 27B on an RTX 5090 — same model, same label, two very different kernel paths and performance.
This paper proposes Population-Curated Policy Optimization (PCPO), a sample-efficient parametric self-evolution method that helps LLM algorithm-design agents escape in-context evolutionary stagnation by training on self-generated algorithms. Trained on only 4 chip cases, PCPO outperforms OpenEvolve and ShinkaEvolve across 16 EDA cases, matches GPT-5.5 with an 8B model, and achieves 8.27× average speedup on GPU kernel design.
MLC AI releases TIRx Harness, an open compiler harness that gives AI agents a structured environment to explore, debug, and optimize GPU kernels. Agent-built Kimi Delta Attention kernels achieved 2.94x geometric-mean speedups over FlashKDA (forward) and 6.84x over FLA (backward).
AttnFuse introduces a composable DSL for compiling custom attention patterns, including RoPE, to fused GPU kernels, achieving speedups over existing methods like PyTorch's flex_attention.
This paper evaluates conformance suites for INT8 quantization in LLM inference, finding that power-of-two scales enable cross-kernel bitwise determinism, while tolerance-based checks are limited to preconditions and boundedness.