1
2
Show HN: Fluxnation, the Fastest FLUX.1 Execution Kernel
GitHub 上号称最快的 FLUX.1 执行内核，通过自定义 C++/CUDA 引擎和 Triton 脉冲注意力彻底消除开销。
My kernel FLUXNATION is the fastest FLUX.1 kernel currently in existence. I engineered this from the ground up to replace standard ComfyUI SingleStrea…
3
法国对英伟达反垄断调查接近尾声，聚焦 CUDA 生态和产业投资
提供免费GPU算力，支持CUDA生态，快速上手AI开发与实验。
IT之家 7 月 9 日消息，据路透社报道，法国竞争管理局官员对外确认，针对 AI 芯片巨头英伟达的反垄断调查已接近尾声，监管机构大概率很快出具正式异议声明，指控企业存在多项限制市场竞争行为。 这起案件可追溯至 2023 年 9 月。当时，法国调查人员突击检查了英伟达法国办公室，作为针对云计算市场竞…
4
BluTrain: A C++/CUDA Framework for AI Systems
揭秘深度学习规模化的系统工程关键，C++/CUDA框架如何重塑训练性能与数值精度，值得技术人深读。
arXiv:2606.24780v1 Announce Type: new Abstract: Progress in deep learning is, at scale, more a matter of systems engineering than of modelling: the be…
5
Local LLM Inference Optimization: The Complete Guide
手把手教你在 RTX 4070 上榨干本地大模型推理性能，实测数据说话，硬件党必看。
Article URL: https://carteakey.dev/blog/local-inference/local-llm-optimization/ Comments URL: https://news.ycombinator.com/item?id=48629858 Points: 2 …
6
Show HN: NanoEuler – GPT-2 scale model in pure C/CUDA from scratch
从零用C/CUDA实现GPT-2，硬核源码教学，深入理解Transformer底层原理
Article URL: https://github.com/JustVugg/nanoeuler Comments URL: https://news.ycombinator.com/item?id=48601472 Points: 2 # Comments: 0
7
DICE: Diffusion Large Language Models Excel at Generating CUDA Kernels
扩散大语言模型在CUDA内核生成任务上超越自回归模型，新论文DICE展示显著优势。
arXiv:2602.11715v2 Announce Type: replace Abstract: Diffusion large language models (dLLMs) have emerged as a compelling alternative to autoregressive…
8
Show HN: cuTile Rust: Safe, data-race-free GPU kernels in Rust
NVIDIA 开源 cuTile Rust，在 GPU 内核中实现安全无数据竞争，基于 CUDA 13.1+ 支持张量分块计算。
Article URL: https://github.com/nvlabs/cutile-rs Comments URL: https://news.ycombinator.com/item?id=48561410 Points: 1 # Comments: 1
9
Show HN: FlashQwen – A from-scratch CUDA inference engine for Qwen3
FlashQwen 从零构建的 CUDA 推理引擎，专为 Qwen3-8B 优化，开源可部署。
Article URL: https://github.com/frankkk96 Comments URL: https://news.ycombinator.com/item?id=48551028 Points: 2 # Comments: 0
10
Learn CUDA and GPU programming without owning a GPU
无需GPU，通过软件模拟CUDA执行模型，掌握map/reduce/scan并行编程核心。
The usual blocker to learning GPU programming is hardware. The official guides assume you have a CUDA-capable NVIDIA card, a working toolchain, and a …
11
AutoMegaKernel: Compile an LLM into one provably-correct CUDA megakernel
将LLM编译成可验证正确性的CUDA megakernel，大幅提升推理性能与可靠性。
Article URL: https://github.com/RightNow-AI/AutoMegaKernel Comments URL: https://news.ycombinator.com/item?id=48447329 Points: 2 # Comments: 0
12
Rl.cu: Training LLM RL with Pure CUDA
用纯CUDA从头实现LLM强化学习训练，极致的性能与底层控制探索。
Article URL: https://github.com/KJLdefeated/RL.cu Comments URL: https://news.ycombinator.com/item?id=48431886 Points: 1 # Comments: 0
13
英伟达未发售的桌面端 RTX 3050 Ti 显卡现身：3328CUDA + 6GB
英伟达未公开的桌面端RTX 3050 Ti显卡规格意外曝光，3328 CUDA+6GB显存配置或填补中低端空白
IT之家 6 月 5 日消息，消息人士 @白给的盖欧卡 今日在社交平台发布了 NVIDIA（英伟达）未曾推出过的桌面端 GeForce RTX 3050 Ti 显卡 的实物照片、软件信息识别结果、基准测试成绩。 ▲ 图源：@白给的盖欧卡 可以看到该显卡的 PCB 代号为 "PG190 SKU 40"…
14
Use your Nvidia GPU's VRAM as swap space on Linux
将NVIDIA显存当交换空间用，巧用CUDA驱动和NBD协议实现系统资源榨干。
Article URL: https://github.com/c0dejedi/nbd-vram Comments URL: https://news.ycombinator.com/item?id=48377404 Points: 269 # Comments: 70
15
Show HN: Tiny-vLLM – high performance LLM inference engine in C++ and CUDA
用C++和CUDA打造的高性能轻量LLM推理引擎，代码开源，适合自部署优化。
Article URL: https://github.com/jmaczan/tiny-vllm Comments URL: https://news.ycombinator.com/item?id=48328184 Points: 6 # Comments: 0
16
Towards Feedback-to-Plan Decisions for Self-Evolving LLM Agents in CUDA Kernel Generation
ICML 2026 论文：让 LLM 智能体利用反馈自主规划，实现更高效的 CUDA 核生成。
arXiv:2605.26720v1 Announce Type: new Abstract: Large language models (LLMs) have shown strong empirical gains as self-evolving agents for CUDA kernel…
17
CUDABeaver: Benchmarking LLM-Based Automated CUDA Debugging
全新基准测试CUDABeaver，系统评估大语言模型在CUDA自动调试中的表现与能力边界
arXiv:2605.08455v2 Announce Type: replace Abstract: Debugging CUDA programs has long been challenging because failures often arise from subtle interac…
18
DeepSeek V4 on Huawei's Ascend 950: A Real Stress Test for China's AI Chip Ecosystem
首款顶级模型在华为Ascend 950上完成推理验证，200+ CUDA算子改写直击中国AI芯片生态真实压力
In April 2026, DeepSeek released V4 — a 1.6 trillion parameter MoE model — and for the first time, the technical report listed Huawei's Ascend NPU alo…
19
Raising the baseline for the `nvptx64-nvidia-cuda` target
Rust编译器将提升NVIDIA GPU编译目标的最低PTX ISA和架构要求，影响旧硬件兼容性。
The nvptx64-nvidia-cuda target is a compilation target for NVIDIA GPUs. When using this target, the final output is PTX . Two version choices shape th…
20
CUDA Books
精选CUDA编程必备书籍，从入门到进阶，涵盖最新2024-2026年新书，适合所有GPU并行计算学习者。
Article URL: https://github.com/alternbits/awesome-cuda-books Comments URL: https://news.ycombinator.com/item?id=48168485 Points: 219 # Comments: 56