AMD Launches ROCm.AI Platform to Automate GPU Kernel Optimization

AMD's new ROCm.AI platform uses frontier AI models to automatically optimize GPU kernels, aiming to close the performance gap with CUDA without manual tuning.

This article was drafted with AI assistance from multiple sources and was reviewed and approved by a human editor before publication.

AMD has unveiled ROCm.AI, a platform that uses frontier AI models to automatically optimize GPU kernels, at its Advancing AI event in San Francisco during the week of 2026-07-24, according to the company.

ROCm.AI plugs into existing code assistants to provide tools and documentation for deploying, debugging, and optimizing models on AMD Instinct hardware, AMD said. The platform includes Hyperloom, an automated workload performance optimization system that can spin up an inference server in Docker, run benchmarks, profile workloads, adjust configurations, or generate custom CPU kernels on the fly.

Anush Elangovan, AMD corporate VP of AI software and solutions, said that AMD publishes a machine-readable ISA for each GPU generation, enabling frontier models to program to AMD hardware. "Frontier models are very capable of programming to AMD's hardware," he added.

In testing on AMD's Helios racks, Hyperloom boosted model performance by 38% over baseline, Elangovan said. AMD is working with AI model houses OpenAI and Anthropic to train models to better understand AMD hardware and software. "We are working deeply with frontier model companies so they natively speak AMD programming," Elangovan stated.

ROCm.AI will be offered as a plug-in for coding assistants.