Anthropic says Claude completed GPU kernel adaptation of a model to AMD accelerators in 3 days
English text is being prepared.
Anthropic says Claude handled almost all of the GPU kernel code porting and writing involved in adapting the model to AMD AI accelerators, with very little human intervention, and the entire process took only 3 days. According to the company's past experience, onboarding a new chip platform using traditional engineering methods typically requires dozens of engineers and takes 6 to 12 months. The chips actually adapted so far include NVIDIA, AMD, Google TPU, and Amazon Trainium. AMD and Anthropic previously announced they will deploy up to 2GW of GPU compute based on AMD's MI450 series, with the first 1GW planned to begin in the first half of 2027; the two sides will also use Claude to optimize AMD GPU workloads and accelerate ROCm software development.
Why read this
AI score 70
The original text provides a time comparison between traditional adaptation and the AI-assisted process, allowing readers to gauge the room for change in the cost of migrating to a chip platform.
Topics
Body
Claude 移植或编写这项工作中几乎全部的 GPU 内核代码,期间人工干预非常少。该案例引起外界对英伟达 CUDA 软件生态和 AI 芯片竞争的讨论,CUDA 是英伟达的护城河,但在 AI 时代这种护城河还能持续多久还需要观察。
A 社称根据公司过往经验,采用传统工程方法接入新芯片平台通常需要数十名工程师花费数月 (实际需要 6 到 12 个月),而在 AI 参与后这个过程可以被缩短到仅仅 3 天。
英伟达的 CUDA 汇集大量开发工具、计算库和针对 GPU 优化的软件,长期以来都是 AI 开发的重要基础,因为将原本针对特定硬件优化的代码适配到其他芯片,需要投入大量工程资源。
而 AI 参与底层代码移植后,可以帮助工程师更快编写和调整 GPU 内核,企业接入不同芯片平台的成本和时间就有机会下降。现阶段 A 社实际适配的芯片包括但不限于英伟达、AMD、Google TPU 和亚马逊 Trainium 等。
AMD 与 Anthropic 此前已经宣布将部署最高 2GW 规模的、基于 AMD MI450 系列的 GPU 算力,首批 1GW 部署计划于 2027 年上半年开始,双方还将利用 Claude 优化 AMD GPU 工作负载、加快 ROCm 软件开发。