BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:1.0
BEGIN:VEVENT
DTSTART:20111115T193000Z
DTEND:20111115T200000Z
LOCATION:TCC 303
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: We propose a GPU projection framework that can estimate the=0Aperformance benefit of GPU acceleration=0Awithout actual GPU programming or hardware.=0AUsers only need to skeletonize pieces of CPU code that are targets=0Afor GPU acceleration. The framework will then transform=0Athe code skeleton in various ways to mimic tuned GPU code=0Awith characteristics resembling real implementations. The synthesized=0Acharacteristics are used by an existing analytical model=0Ato project GPU performance. The cost and benefit of GPU development=0Acan then be estimated according to the transformed=0Acode skeleton that yields the best projected performance. With=0Athis framework, users can leap towards GPU acceleration only=0Awhen the cost-benefit makes sense. We validate the framework=0Ausing micro-benchmarks as well as data-parallel codes in legacy=0Ahigh performance computing applications. The measured performance=0Aof manually tuned codes deviates from the projected=0Aperformance by 17% in geometric mean.
SUMMARY:GROPHECY: GPU Performance Projection from CPU Code Skeletons
PRIORITY:3
END:VEVENT
END:VCALENDAR