Design Challenges in HPC for AI/ML Applications
Vikram Rajan, Niwaran Chandra Kumar, Nihar Ranjan, Amgothu Murali Krishna · 2024
High-performance computing (HPC) and Artificial Intelligence (AI) are two areas of computing that are seeing fast-paced growth. The HPC systems have evolved to incorporate the capability of running mid-level AI algorithms for C4I (Command, Control, Communications, Computers and Intelligence) networks. HPC solutions can be used at the edge or in the cloud. They are used to solve large-scale computing problems, which makes it suitable for a wide range of problems. They are able to perform this in reasonable cost and time. AI enables HPC systems to more intelligently process workloads and make full use of the processing capability. With a variety of CPUs, GPUs, TPUs, and ASICs, design of the system architecture to keep pace with the advancing technology is an uphill task. In the current paper, we discuss about one such HPC system of CPUs and GPUs and the challenges involved in architecture and design. Specifically, on ensuring the bandwidth between the CPU and GPU and ensuring the thermal performance of the system.