ROCm 7.1.1 released
AMD has released ROCm 7.1.1, a new update to its open-source software stack designed for graphics processing unit (GPU) computation. This release provides a comprehensive ecosystem for developers to customize their GPU software to meet specific needs, making it an ideal choice for high-performance computing (HPC), artificial intelligence (AI), scientific computing, and computer-aided design (CAD).
This change isn't just about hype but about providing practical tools for developers who need serious performance from AMD's Instinct GPUs. Think HPC, AI research, and scientific modeling; this version makes it easier to tailor your software specifically and get reliable results across these areas.
What’s notably improved: There’s better support underneath the hood now for virtualization environments like KVM SR-IOV with some specific new guest OS options (Ubuntu 24.04 being one of them). This includes backing for RHEL too, both versions 9.7 and 10.1, using their relevant kernels.
Digging into features: On the monitoring side, AMD has really focused on making things smoother for developers who use MI350X or MI355X GPUs. Their SMI tool now does less work repeatedly (thanks to caching) and can actually pull more detailed performance stats per partition, which is useful stuff.
Also addressed are some older issues with the MI325X cards when running in SR-IOV Mode 1, which is probably good news if you're still using those configurations. And there’s a specific improvement for GEMM kernel selection efficiency; developers working on matrix operations should find this helpful.
For deep learning folks: If you work with PyTorch, you'll want to know it's now compatible with version 2.9. The ROCm ecosystem itself is getting stronger, too. Plus, Hugging Face Transformers models can now be built using gfx1201 hardware (the specific GPU architecture designation).
Biggest news for data science: The ROCm Data Science Toolkit has officially moved into the general availability stage. This means hipDF and hipMM are now considered fully production-ready components.
But there’s more! Two entirely new tools have made the leap to full release: hipRAFT and hipVS. And performance enhancements aren't limited just to the core stack; they've boosted things in DGL (used for graph neural networks) and llama.cpp, plus added improvements related to creating offline installers.
On a side note about fixes: They've addressed several annoying issues in different areas of ROCm, including AMD SMI, Composable Kernel setup, the high-level abstraction layer called HIP, machine learning tools like MIGraphX or RCCL (used for distributed deep learning), and important performance libraries like rocBLAS.
Finally, practical considerations: The Composable Kernel component is pushing ahead to adopt C++20 features in a future update. This means developers using it will need their build environments ready with those newer tools installed; otherwise, they might encounter bumps during the upgrade path later on if they stick with ROCm 7.1.1.
It feels like another mature step forward for AMD’s open-source GPU platform.
Release ROCm 7.1.1 Release
ROCm 7.1.1 release notes The release notes provide a summary of notable changes since the previous ROCm release.
