逆向工程:逐位精确模拟英伟达 Blackwell 张量核心
ycombinator · x · 2026-08-05
A development team has successfully built a software model of NVIDIA Blackwell's Tensor Core MMA instruction that matches the output of real B200 hardware bit-for-bit.
This is not just a numerical approximation, but an exact replication of the underlying logic. The author shares the process of how they reverse-engineered the hardware to achieve this.
「Infra」频道最新
- AI 编码瓶颈不在模型，而在测试与 CI 流水线 — hichaelmart · 2026-08-05
- xAI 宣布建设第四座数据中心,配备 22 万块 GB300 — chrisgrayson · 2026-08-05
- 告别云端重配环境：本地大模型持久化 GPU 桌面算力经济账分析 — ievseev · 2026-08-05
- 图分析基准 Graph500 入选 SPEC CPU 2026 处理器测试套件 — Prof_DavidBader · 2026-08-05
- Groq 推出全新推理架构 LPX：结合自研 LPU 与英伟达 GPU — GroqInc · 2026-08-05
- 车库太阳能供电：384GB至强主机远程跑AI编程 — angadsg · 2026-08-05