华尔街见闻·2026-01-23 09:42Core Viewpoint - The article highlights the potential of generative AI, specifically the Claude Code platform, in facilitating the migration of code from NVIDIA's CUDA to AMD's ROCm platform, which could challenge NVIDIA's long-standing technological moat built around CUDA [1][2]. Group 1: Migration Capabilities - A user successfully migrated an entire CUDA backend to AMD's ROCm platform using Claude Code without needing an intermediate conversion layer, raising market interest [2]. - The migration process reportedly faced minimal issues, primarily related to "data layout" differences, indicating the tool's effectiveness in handling simpler kernel code [3][4]. Group 2: Limitations and Challenges - Industry experts caution that the success of Claude Code may be limited to simpler kernel codes, as more complex codebases requiring deep hardware optimization still pose significant challenges for AI tools [3][6]. - The real difficulty lies in migrating interconnected complex codebases, which necessitates a comprehensive understanding of contextual information for effective conversion to ROCm [7]. Group 3: Technical Framework - Claude Code operates using an intelligent agent framework that smartly replaces CUDA keywords with their ROCm counterparts while maintaining the underlying logic of specific kernels, rather than merely performing keyword substitutions [4]. - The tool simplifies the migration process, allowing developers to execute the transition via a command-line interface without the need for complex configuration environments like Hipify, thus lowering the barriers for platform migration [4].