AMD ROCm 10 Launches With 3.3x Inference Gains and TheRock Build

AMD has released ROCm 10, a major update to its open-source software platform for AI and GPU compute. This release skips versions 8 and 9 to deliver significant architectural changes for developers. The update matters because it introduces a new build system and general availability for ROCm.AI, which promises faster model training and inference. Developers working with AMD Instinct GPUs can now access these performance gains without waiting for separate SDK updates.

New build system and unified AI platform target faster model training

The core of this release is ROCm.AI, which is now generally available to all users. This platform combines AMD Skills, the ROCm Command Line Interface (CLI), and ROCm Hyperloom into a unified experience. AMD also retired the separate HIP SDK, replacing it with the new ROCm Core SDK. This consolidation simplifies the development environment for those building AI applications on AMD hardware.

AMD claims that a system configured with ROCm.AI achieved an average 3.3x improvement in inference performance. The company also reported a 2.4x improvement in training performance compared to ROCm 7. These tests were conducted using eight Instinct MI355X GPUs. The performance gains stem from the adoption of the new TheRock build system across the entire software stack.

ROCm CLI is now available as a Technology Preview for both Windows and Linux. This cross-platform support allows developers to manage their ROCm environments more flexibly. The release marks a shift in how AMD structures its AI software tools for broader accessibility. We have been tracking ROCm closely and reviewed the AMD ROCm 7.14 Stable Release Adds for context on previous iterations.

ROCm 10 is now available as an open-source platform for AI and GPU compute. The update provides developers with improved performance metrics and a streamlined SDK structure. Users can begin testing the new TheRock build system and ROCm.AI features immediately.