ROCm 10 turns AI optimization into a developer workflow
AMD is pairing its updated compute stack with agent-assisted profiling, kernel optimization, diagnostics, and model-serving tools.
AMD has released ROCm 10 and made its ROCm.AI development experience generally available. The update combines a more modular compute stack with tools for serving, profiling, diagnostics, memory analysis, and AI-assisted performance work.
One notable component is Hyperloom, which AMD describes as an agentic optimization system that can profile inference workloads, identify bottlenecks, propose changes, benchmark them, and validate correctness. The wider tooling also includes a unified CLI and reusable AMD-specific skills for coding agents.
AMD reports substantial average gains in selected inference and training tests on its own hardware. Those figures come from AMD’s configurations and model mix, so they should be treated as a starting point for workload-specific testing rather than a universal upgrade expectation.
Why it matters: Accelerator competition is increasingly decided by software. Better compilers, kernels, observability, and deployment workflows can matter as much as peak hardware specifications when teams move models into production.
Source attribution
Original reporting and release information: AMD Newsroom.