Wednesday, April 10, 2024 11am to 12:15pm
About this Event
Engineering 2 1156 High Street, Santa Cruz, California 95064
#CSEcolloquiumPresenter: Raph Levien
Abstract:
GPUs have about 10x throughput per dollar (or watt) than CPUs, fueling their massive adoption in 3D graphics and AI workloads. However, for general purpose computing, CPUs still reign. Why is that? A big part of the reason is that GPUs have an impoverished execution model tailored either to the specifics of 3D rasterization, or to grids of work where the dimensions are known ahead of time. An intriguing possibility is a parallel computer with the same throughput advantages as GPUs, but more agile with respect to workloads that don't fit those templates. This talk will discuss historical and industry trends, exploring why we're in the current situation, and two possible paths to a future good parallel computer, namely an evolution of GPUs with more flexible programmable dispatch, and the new generation of AI accelerators based on a large grid of CPUs (often RISC-V) with vector units. It ends with proposing a number of directions for academic research which might bring us closer to that future.
Bio:
Raph Levien is a research software engineer on the Google Fonts team, primarily working on GPU font rendering. He’s been actively involved in the Rust community for over seven years, and has worked on pulldown-cmark, the Xilem UI toolkit, and other popular crates. He has a PhD from UC Berkeley on the topic of interactive curve editing, and has been involved in font creation and 2D graphics tools most of his career.
Hosted by: Professor Lindsey Kuper
Zoom: https://ucsc.zoom.us/j/94784153196?pwd=cmZFeTJGZHNPaE9XV3NxRHRhOTR1dz09
User Activity
No recent activity