Anthropic 的 AI 助手 Claude 在仅一名工程师的参与下，于一个周末内独立完成了 AMD 新型 GPU MI355X 的适配和性能优化工作，无需人工代码修改[1]。这一进展标志着英伟达构建了 20 年的 CUDA 生态护城河面临直接挑战。
为支持 AI 模型的自主优化能力，AMD 推出了 ROCm.AI 工具集，其中包含 AMD Skills 知识库和 Hyperloom 推理服务[1]。这套工具使 AI Agent 能够自主调试 GPU 性能。在实际应用中，Hyperloom 将 MiniMax M3 的输出速度提升了 38%，并在一次测试中成功运行了 1.4 万个模型[1]。AMD 还改变了 GPU 芯片说明书的编写方式，提供机器可读的 ISA 指令集[1]，使 AI 模型能够更直接地理解硬件特性。
在部署规模方面，Anthropic 计划在 AMD Helios 系统中部署最高 2GW 的 Instinct GPU，首批 1GW 的部署预计于 2027 年上半年启动[1]。AMD 副总裁 Anush Elangovan 表示，这一方向的目标是让前沿模型"天生会说 AMD 编程"[1]。
Anthropic's Claude AI model has demonstrated a significant breakthrough by independently completing the adaptation and performance optimization of AMD's MI355X GPU over a single weekend, without requiring human engineers to modify code [1]. This achievement marks a direct challenge to NVIDIA's two-decade-old CUDA ecosystem.
The optimization was accomplished with minimal human involvement, leveraging just one Anthropic engineer working alongside Claude [1]. To support such autonomous GPU optimization, AMD has released the ROCm.AI toolset, which includes the AMD Skills knowledge base and Hyperloom inference service [1]. These tools enable AI agents to self-debug GPU performance. According to AMD Vice President Anush Elangovan, the goal is to make cutting-edge models "speak AMD programming natively" [1].
The partnership extends beyond software optimization. Anthropic has committed to deploying up to 2 gigawatts of Instinct GPUs within AMD's Helios systems, with the first phase of 1 gigawatt scheduled to launch in the first half of 2027 [1]. Additionally, AMD has redesigned its GPU instruction set documentation to be machine-readable, enabling AI systems to directly interpret GPU specifications [1]. Early results from the Hyperloom service show a 38% speed improvement in MiniMax M3 inference throughput during testing involving 14,000 models [1].
