Role of GPU nodes

Existing and emerging applications ranging from scientific computing, to big data, to context-aware computing will continue to drive microprocessor designers to improve performance to new heights. While performance is crucial for these applications, power-efficiency has become a primary microprocessor design goal in the many-core era. Therefore, current and future generation microprocessors have abandoned frequency scaling, and instead leverage parallelism, heterogeneity, reconfigurability, and domain-specific enhancement to provide power-efficient performance.

IDRE has increasing interest in Graphics Processing Unit (GPU) and many cores based computing architectures. This is because the overwhelming majority of microprocessor designs these days leverage parallelism in some way, whether it be the thread-level parallelism seen in many-core and/or hyperthreaded designs or the instruction-level parallelism exploited in dynamically scheduled processors. Heterogeneous designs feature different types of cores and other on-chip components that enable application writers to map their software to the appropriate hardware for optimal power efficiency. Even today’s top supercomputers in the world are increasingly using heterogeneous and power efficient processor arrangements such as Graphics Processing Unit (GPU) and/or Many integrated cores (MIC)-Intel Phi. Per the current trend it is most likely that future systems will use such heterogeneous architectures to make Exascale supercomputer systems.

IDRE’s GPU/Many Core program was established to connect domain experts across campus with technology experts in GPUs and many-core computing. In particular it is about discussing and sharing expertise and thus helping researchers to port and design algorithms to run existing or new codes effectively on a variety of new computer architectures, including multi-core processors, many integrated cores (MIC), and graphical processing units (GPUs).

The current projects under GPU/Many Core program are:

Development of Particle-in-Cell Algorithms for Advanced Computer Architectures MRI Grant from NSF for acquisition of a GPU-based cluster for Plasma, High-Energy Density and Computational Science at UCLA-Dawson2 cluster