ROCm is an open source software platform that allows researchers to harness the potential of AMD Instinct accelerators to facilitate cross-platform high-performance computing and AI innovation.
With more and more developers starting to run ROCm on Radeon and Radeon PRO GPUs for AI and ML work, AMD launched the latest ROCm 5.6 and announced ROCm support for consumer graphics cards.
It is said that AMD will first adapt ROCm for the RX 7900 XTX and Pro W7900 RDNA 3 GPUs. The official version will be launched this fall (ROCm 5.6), and more models will be added later.
This is definitely big news for AMD users, especially those who need AI productivity. Not only that, it is noticed that ROCm 5.6 will also bring a series of new features and enhancements, especially for AI and HPC workloads, including:
Integrate the Hugging Face unit test suite into ROCm QA
Gradually support OpenAI Triton in PyTorch 2.0 inductor mode;
Enable OpenXLA support for PyTorch, TensorFlow, and JAX via ROCm;ROCm 5.6 also introduces improvements to math libraries such as FFT, BLAS, and solvers, which are the foundation of HPC applications, and enhancements to ROCm’s development and deployment tools, including installation, ROCgdb (CPU-GPU integrated debugger ), ROCm analyzer and documentation.
AMD also stated that it will continue to work on further optimizing the framework and back-end compilers for better performance, including MLIR infrastructure improvements, as the basis for AMD’s support for OpenAI Triton and OpenXLA compilers, and will continue to be added on the AMD hub of Hugging Face More open source AI models optimized for AMD solutions.
For HPC users, AMD already offers some solutions on the AMD Infinity Hub to allow customers to better build HPC application containers, and future ROCm releases will increase the number of HPC applications supported on AMD Instinct solutions.
