AMD released ROCm 10.1 on October 5, 2026, saying it improves data movement between storage and GPUs for AI and HPC workloads. It is the company's first software update in the AI/ML category since ROCm 7.14.
AMD reported a significant improvement in data delivery efficiency, measured by the reduction in latency and increased sustained throughput for checkpoint engines and KV-cache offloading systems. That compares with the previous version, which used a traditional host-memory staging step.
ROCm 10.1 is built on AMD Infinity Storage technology and targets AI and HPC workloads where data movement between storage and accelerators is a bottleneck. Availability begins with the release of the update, initially for developers and researchers using AMD GPUs.
"Asynchronous Fast-Path Backend allows read and write requests to run directly on a HIP stream, skipping the host-memory staging step through which data traditionally passes," said Adam Chan, lead engineer at AMD. This change reduces latency and improves sustained throughput for data-intensive workloads.
The announcement follows the launch of AMD Infinity Storage in ROCm 7.14. AMD said the update reflects the growing need to address data delivery bottlenecks as AI models and HPC datasets scale.
AMD did not say how widely the improvements will be adopted, and the company raised the question of whether data movement will remain a bottleneck as models continue to grow. The update is expected to be integrated into future ROCm versions.
Source: amd
