There will be a special NCSA lecture on Monday, December 10 at 2:00 p.m. in 1030 NCSA. Presenting "Future of Systems: AI Hardware" will be Rama Divakaruni, Distinguished Engineer in Process Technology Research and Fabless Interface for IBM Research.
Abstract: In the AI era, workloads are strongly driving systems to become much more architecturally diverse. Today's systems are composed of a mix of heterogeneous components that include CPUs, GPUs, specialized accelerators, and memories. A prime example of this in the HPC space is SUMMIT: the world's most powerful supercomputer. It combines POWER9 CPUs and NVIDIA GPUs using high bandwidth interconnects and has been at work solving some of the world's most challenging science problems. As technology scaling becomes harder, more complex and costly, value at the system level for today's workloads is achieved more by a diversity of parts, improved connectivity and bandwidth rather than sheer compute power in a monolithically integrated SoC. This will drive advanced packaging techniques such as heterogeneous integration to enable high inter-connectivity among these components as well as advanced cooling techniques and will be crucial for continued progress in AI. Compute at reduced precision will be another lever for acceleration and offer ample opportunities for innovation across the stack. Analog accelerators based on non-volatile elements such as phase change memory, resistive RAM and even electrochemical RAM that are currently in a research phase present another knob for accelerating deep learning. With all these advances in the pipeline, it become possible to envision a future where a variety of heterogeneously enabled form factors driving different workloads for inference and training are pervasive on the edge and in the cloud.