Democratizing AI Accelerators for HPC Applications: Challenges, Success, and Support

Recently, there has been a surge in the number of novel AI hardware architectures, such as Cerebras CS-2, GroqRack, SambaNova DataScale, Graphcore IPUs, and Habana Gaudi. Some of these systems are accessible to the research community through testbeds in Europe (e.g. EPCC at the University of Edinburgh) and in the US (e.g. ALCF AI Testbed). Beyond their primary use in AI, these accelerators have sparked intense interest and rapid development in their capacity for handling HPC workloads and driving algorithmic research. Their distinctive architectural designs, often with significant raw compute capability and high bandwidth, have proven capable of significantly enhancing real-world performance for specific workloads, often outperforming GPUs. As these technologies are still emerging, the maturity levels of associated tooling and software stacks vary considerably, and there are gaps in documentation and support. This BoF aims to delve into several pertinent questions: What are the prevailing trends in AI accelerators for HPC applications? How do the different architectures better suit different workloads? Where can one find robust training resources? What are the common challenges users encounter with these accelerators? How can vendors enhance support for developers? Where can interested users access and experiment with these hardware resources? Pooling the extensive knowledge within the community, this BoF intends to assemble developers experienced in, or are interested in using AI accelerators for HPC applications. The aim is to foster discussion and share challenges, experiences, success stories, and insights gained from using these accelerators.

---

Contributors:

• Juan J. Durillo