kernels

在 PyTorch Con 2026 上，Mark Saroufim 探讨了支撑更优优化器的线性代数内核，涵盖 QR、对称特征分解与 Cholesky，并从 GPU MODE 的线性代数竞赛中总结了如何通过暴露并行性、重叠独立阶段和合并更新来减少数据搬运。

DeepSeek released DeepGEMM, an open-source GPU math engine used to train and run its own models, claiming up to 98% token cost savings by squeezing more tokens per second out of existing GPUs.

Tianqi Chen (MLC AI) announced an open-source online book and upcoming CMU ML Systems course module on Agentic GPU Programming for MLSys, covering how coding agents use compiler-driven harnesses, knowledge bases, and benchmarking to write and optimize GPU kernels.

DeepSeek 开源了 DeepGEMM Ascend，将 DeepGEMM 移植到华为昇腾平台，API 完全兼容并支持 BF16/FP8/FP4 GEMM、MQA logits 和 MegaMoE 算子，在 GEMM 上达到硬件性能极限的 99.8%、MegaMoE 达 98%。

This article explains how GPUs organize computational work into kernels, grids, blocks, and warps to achieve massive parallelism and hide memory latency through streaming multiprocessors.

A team announced they achieved first place on NVIDIA's SOL-ExecBench benchmark across all tracks, beating 583 teams with real production kernels on B200 GPUs.

The article encourages support for @ViC305, a key contributor to the DGX community who produces AI-related work such as quants and kernels, despite not owning a DGX Spark.

A tweet explains the core components of inference engineering: loadable encodings, backend operations, and native hardware arithmetic, with a discussion on kernels and GPU efficiency.

This article provides an introduction to GPU kernels in the context of AI inference and performance engineering, explaining their definition, how they are used, and the advantages of custom kernels for optimization.

Hugging Face releases 207 WebGPU kernels for local AI inference in browsers, along with a JavaScript loader library and a benchmarking tool called Fleet.

Introduces v100-skinny, a custom kernel library enabling fast NVFP4 inference on V100 (sm70) GPUs, achieving 366 t/s for Qwen3.6 27B in best-case extraction, with lower speeds for structured generation and code.

Y Combinator hosted a Paper Club where researchers presented innovations in multi-GPU kernel optimization, including ParallelKittens, a CUDA framework that simplifies development of overlapped multi-GPU kernels and achieves significant speedups across workloads.

QuixiAI rebrands ThunderKittens and ThunderMittens into QuixiCore-CUDA and QuixiCore-Metal, creating a unified family of cross-platform kernels for AI workloads.

A highly recommended 4.5-hour GPU programming lesson on CUDA and ThunderKittens by Ben Spector, offering an in-depth, behind-the-scenes look at kernel optimization.

QuixiAI ported ThunderKittens to Metal, enabling kernel support on MPS and MLX for training models on Mac.

TokenSpeed-Kernel is a portable, high-performance kernel system for LLM inference that enables zero vendor-specific model code and supports multiple GPU architectures, achieving up to 3.6x higher throughput on AMD MI355X.

A blog post describing a tiny compiler that demonstrates how to lower data-parallel kernels by converting for loops into vectorized loops with lanes and masks, implemented in ~180 lines of Python.

Downloaded and compiled 16,459 CUDA kernels from SakanaAI's open-source dataset, benchmarking them for performance.

Ahmad Osman announces four upcoming articles covering LLM decoding/prefilling, LLM kernels, and hardware comparisons (CPUs, GPUs, Tenstorrent, Apple Silicon), building on his existing 'Local LLMs From Zero to Hero' series.

The tweet explains that running AI models is really about running optimized kernels, and that inference engines and their kernel implementations are critical for performance, not just the model or hardware.