ROCm 7.2.3 Brings Minor Updates, ROCm XIO Documentation
Less than one month after releasing ROCm 7.2.2, the ROCm 7.2.3 is now available with some minor improvements to this open-source AMD GPU compute and AI stack.
ROCm 7.2.3 doesn't come with any hardware/GPU support changes nor on the operating system front. A bit surprising is still not seeing official support yet for Ubuntu 26.04 LTS that released in April, but presumably that will come for ROCm 7.3.
ROCm 7.2.3 does bring improved profiling accuracy for vLLM workloads. Previously there were "large, sporadic idle gaps" when trying to profile with vLLM workloads but now that should be better handled.
MIGraphX has also seen improvements with ROCm 7.2.3 such as for better performance when using the gather operator and better ONNX Runtime reliability.
AMD also improved the documentation with ROCm 7.2.3. One of the key documentation improvements is now documenting ROCm XIO as the API for accelerator-initiated "XIO" for AMD GPU device code. With ROCm XIO there can be direct I/O from the accelerator to NVMe SSDs, RDMA NICs, and SDMA engines from device code without involving the host processor. ROCm XIO quietly released in April as an early-access software technology preview and not yet production rated but continues to be improved upon.
Downloads and more details on today's ROCm 7.2.3 update via the ROCm documentation.
ROCm 7.2.3 doesn't come with any hardware/GPU support changes nor on the operating system front. A bit surprising is still not seeing official support yet for Ubuntu 26.04 LTS that released in April, but presumably that will come for ROCm 7.3.
ROCm 7.2.3 does bring improved profiling accuracy for vLLM workloads. Previously there were "large, sporadic idle gaps" when trying to profile with vLLM workloads but now that should be better handled.
MIGraphX has also seen improvements with ROCm 7.2.3 such as for better performance when using the gather operator and better ONNX Runtime reliability.
AMD also improved the documentation with ROCm 7.2.3. One of the key documentation improvements is now documenting ROCm XIO as the API for accelerator-initiated "XIO" for AMD GPU device code. With ROCm XIO there can be direct I/O from the accelerator to NVMe SSDs, RDMA NICs, and SDMA engines from device code without involving the host processor. ROCm XIO quietly released in April as an early-access software technology preview and not yet production rated but continues to be improved upon.
Downloads and more details on today's ROCm 7.2.3 update via the ROCm documentation.