AMD ROCm 5.6
Discover ROCm 5.6 latest AI enhancements and optimizations, pushing performance to new heights for machine learning and beyond.
AMD to Add ROCm Support on Select RDNA 3 GPUs this Fall
The key technology influencing the upcoming computing generation is artificial intelligence. Its all witnessed in recent months how the rise of generative AI and LLMs is changing how it use technology and increasing demand for high-performance computing in data centers with GPUs at their core.
ROCm is an open software platform that facilitates cross-platform HPC and AI innovation by enabling researchers to leverage the capabilities of AMD Instinct accelerators. A whole range of optimizations for AI and HPC applications are included in ROCm 5. These include support for new technologies such as the OpenAI Triton programming language, support for new data types, and optimized kernels for large language models.
ROCm validation now encompasses validation across thousands of models and operations as well as hundreds of thousands of framework tests every night. Support for popular frameworks including as PyTorch, TensorFlow, ONNX, and JAX is part of this, allowing for the best possible out-of-the-box developer experience for all AI models developed on top of these frameworks.
New AI software add-ons for large language (and other) models, along with other performance improvements throughout the ROCm library portfolio, have improved the capabilities of ROCm 5.6. Additionally, it shows ongoing advancements assisting the AI community, such as:
- ROCm QA’s incorporation of the Hugging Face unit test suite
- OpenAI Triton is gradually supported in PyTorch 2.0 inductor mode.
- Support for the larger community to use ROCm for PyTorch, TensorFlow, and JAX to enable OpenXLA support
Along with changes to ROCm development and deployment tools including install, ROCgdb (the CPU-GPU integrated debugger), ROCm profiler, and documentations, ROCm 5.6 also brings updates to a number of math libraries, such as FFT, BLAS, and solvers that serve as the foundation for HPC applications.
Underpinning AMD support for OpenAI Triton and OpenXLA compilers, it will keep working to further optimize frameworks and backend compilers for optimal performance, including MLIR infrastructure improvements, as the industry shifts to an open ecosystem that supports a wide range of accelerators. On the Hugging Face AMD portal, they will also keep adding more open-source AI models that are tailored for AMD solutions.
Future ROCm releases will increase the number of HPC applications supported by AMD Instinct products, and it have released many publishing recipes at AMD Infinity Hub to help clients construct the HPC application containers.
Additionally, AMD has heard about the community’s struggles with particular driver difficulties on unsupported GPUs and observed a great deal of interest from developers who wish to run the ROCm open software platform for AI and ML on both Radeon consumer and Radeon Pro workstation GPUs. AMD to report that the issues in ROCm 5.6 have been resolved, and determined to continue growing our support.
With the Radeon Pro v620 and w6800, two AMD RDNA 2 workstation GPUs that are already supported, intend to extend ROCm support to a few AMD RDNA 3 workstation and consumer GPUs. Starting with the 48GB Radeon PRO W7900 and the 24GB Radeon RX 7900 XTX, formal support for RDNA 3-based GPUs on Linux is expected to roll out this fall. More cards and enhanced features will be provided later.
