AMD publishes machine-readable ISA so frontier models can write its GPU kernels

AMD has published a machine-readable ISA for its Instinct GPUs and partnered with Anthropic and OpenAI to let frontier models natively write and optimize low-level kernels, claiming a 38% inference speedup over baseline via automated profiling tools. The move directly attacks the CUDA performance moat by enabling PyTorch and JAX workloads on AMD hardware to approach hand-tuned throughput through prompts rather than kernel engineers.

Hacker News https://www.theregister.com/ai-and-ml/2026/07/24/amd-vibe-codes-its-way-past-the-cuda-moat-with-rocmai/5278580 AMD publishes machine-readable ISA so frontier models can write its GPU kernels Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. AMD is now shipping machine-readable ISA specs plus an agentic optimization tool Hyperloom that plugs into Claude Code, Codex, Cursor, etc., and claims a 38% inference speedup over baseline by auto-generating and tuning kernels on Instinct hardware. If it holds up, this directly attacks the "CUDA moat" performance gap—meaning your existing PyTorch/JAX workloads on AMD could get close to hand-tuned throughput via prompts rather than kernel engineers, changing the cost calculus of moving inference off NVIDIA. Treat the 38% as vendor-claimed and benchmark your own model/rack config before betting deployments on it. AMD has published its machine-readable GPU ISA and partnered with Anthropic and OpenAI to let frontier models natively write and optimize low-level AMD Instinct kernels, yielding a 38 percent performance boost over baseline using automated profiling tools. In production, this allows your deployment agents to bypass manual CUDA conversion Bottlenecks and auto-tune non-Nvidia hardware on the fly. This significantly lowers the engineering barrier to scaling down inferencing costs on AMD hardware without sacrificing custom kernel-level performance optimizations.