03-18, 14:00–14:25 (US/Eastern), Tsai Auditorium (CGIS S010)
The growing demands for accelerated computing in batch HPC and AI training call for innovative strategies to enhance infrastructure utilization. GPU fractionalization enables dynamic allocation and sharing of GPU resources, resulting in cost savings and improved efficiency. This talk will discuss key approaches, including NVIDIA Multi-Instance GPU (MIG) and JuiceLabs' dynamic GPU sharing software, highlighting their features, impact on system design, and interaction with the Open OnDemand ecosystem. We will also present a new initiative between Cambridge Computer and JuiceLabs to develop the integration of their GPU-sharing technology with Open OnDemand and discuss how institutions can start using the product today and contribute to the design and vision of the product.
Dr. Christopher S. Simmons is Cambridge Computer Services’ Scientist in Residence responsible for providing thought leadership for clients and cultivating relationships in our industry. He has over 25 years of experience in all facets of computational science and high performance computing. Dr. Simmons earned his Ph.D. from the University of Texas at Austin in Physical Chemistry with an emphasis in Computational Quantum Mechanics and is the Project Lead for OpenHPC.