Stories, not Words: Accelerating the Processing of Similarly Structured Data

ECE Seminar: Stories, not Words: Accelerating the Processing of Similarly Structured Data

Starts at: September 21, 2017 4:30 PM

Ends at: 6:00 PM

Location: SH 125

Speaker: Dr. Lisa Wu

Affiliation: PostDoc Research Fellow

Link to Abstract

Link to Video (1)



The failure of Dennard scaling and the rapid growth of data produced and consumed daily have made mitigating the dark silicon phenomena and providing fast computation for processing large volumes and expansive variety of data while consuming minimal energy the utmost important challenges for modern computer architecture. This talk will introduce the concept that grouping data structures that are previously defined in software and processing them through hardware specialization can significantly improve the application performance and energy efficiency. We target application domains where the common data structures are widely used and show that specializing both the compute and memory subsystem provides orders of magnitude improvements in performance and energy efficiencies. As case studies for regular and irregular data-access application domains, we present a database accelerator that processes tables and columns, and a graph accelerator that processes vertices and edges as similarly structure data. Creating specialized encapsulated data accesses and datapaths allows us to mitigate unnecessary data movement, take advantage of data and pipeline parallelism, and consequently provide substantial energy savings while obtaining significant performance gains.