AMD publishes machine-readable GPU specs so AI models can auto-tune code

At its Advancing AI event in San Francisco this week, AMD corporate VP of AI software and solutions Anush Elangovan announced ROCm.AI, a platform designed to let frontier models automatically write and optimize GPU code for AMD hardware. The core insight is simple: frontier models (large language models) have become surprisingly good at writing low-level GPU kernels—the handwritten code that developers normally have to craft to unlock a chip's true potential. AMD's advantage is that it has already published machine-readable ISA (instruction-set architecture) specifications for every generation of its GPUs, meaning these models have access to the precise hardware details they need to generate correct, efficient code.

ROCm.AI includes Hyperloom, an automated workload optimization tool that orchestrates this process end-to-end. When invoked—for instance, by a developer or code assistant asking to "optimize MiniMax M3 with Hyperloom"—the tool spins up an inference server in a Docker container, runs benchmarks to establish a performance baseline, profiles the workload to identify bottlenecks, and then either adjusts configuration or generates custom kernels on the fly. In testing on AMD's newly launched Helios racks, Elangovan claims this process achieved a 38 percent performance boost over baseline.

To make this work at scale, AMD is not just relying on what frontier models already know. Instead, it is working deeply with model houses like OpenAI and Anthropic to ensure their models are trained to understand AMD hardware and software natively. As Elangovan put it: "We're not just using the frontier model to generate a kernel. We're working deeply with frontier model companies so that they natively speak AMD programming." ROCm.AI will be offered both as a built-in command-line interface and as a plug-in for popular code assistants—including Anthropic's Claude Code, OpenAI's Codex, Google's Antigravity, and Cursor—so developers can trigger optimization directly from their existing workflows.

AMD has long faced a perception problem: while its GPUs are increasingly competitive in raw performance, developers default to Nvidia because CUDA makes it easier to write code that actually runs fast. The so-called CUDA moat has weakened in recent years as frameworks like PyTorch and JAX allow write-once-run-anywhere code, but the gap remains when it comes to optimization. Hand-tuning GPU kernels and matrix-multiplication routines to exploit hardware fully is a specialized skill, and most developers lack it.

AMD's strategy—publishing machine-readable ISA specs and training frontier models to understand its hardware—flips the problem. Instead of requiring developers to learn low-level GPU programming, the AI model itself becomes the optimizer. By working directly with OpenAI and Anthropic to ensure their models "natively speak AMD programming," AMD is baking hardware knowledge into the training itself. This means that when a developer (or a code assistant prompted by a developer) asks for optimization, the model already understands AMD's architecture at a deep level. The 38 percent performance boost in Hyperloom testing, if generalizable, would make AMD's convenience gap vs. Nvidia measurably smaller.

No comments yet. Be the first to share your thoughts!