Maximize GPGPU performance-per-watt across real scenarios with 10x efficient power solution
TimeWednesday, July 13th1:30pm - 1:45pm PDT
Location2010, Level 2
DescriptionHigh perfomance of general-purpose GPU has fundamentally changed from chip-level peak performance to system-level overall performance. This system level performance refers to maximizing system-level throughput while staying within or further reducing power. Analyzing milliseconds duration of real stimulus early in the design cycle for GPGPU with its programmability becomes increasingly important. This paper is using the streaming flow of emulator+power anlaysis tools, which is software and hardware collaboration. Worst-case scenarios can be obtained by identifying critical time windows for chip level IR-drop analysis and system level thermal analysis and PDN analysis. And power analysis of the long vectors can be within a reasonable runtime and memory capacity. This enables rapid iteration of RTL design, chip hotspot/EM optimization, and package and board optimization.