130 likes | 215 Views
Big Data Reading Group Slides on MS Research’s “Accelerator”. Jason Campbell Sept 17, 2007. GPUs for non-graphics computing. Upside: high numbers of FLOPs now e.g., ATI Radeon x1900 250 GFLOPS. GPUs for non-graphics computing. Upside: high numbers of FLOPs now
E N D
Big Data Reading Group Slides on MS Research’s “Accelerator” Jason Campbell Sept 17, 2007
GPUs for non-graphics computing • Upside: high numbers of FLOPs now e.g., ATI Radeon x1900 250 GFLOPS
GPUs for non-graphics computing • Upside: high numbers of FLOPs now e.g., ATI Radeon x1900 250 GFLOPS • Downside: you can have any ops you want as long as they’re all single precision MAC and the programs are ~< 20 instructionswith no branching or (dynamic) looping MAC = multiple accumulate single precision = 8 bit exponent, 23 bit mantissa Other GPUs have less precision, down to 8 bits!
Accelerator uses texture memory & pixelshaders only
What works best • Intensely parallel operations • Limited need for precision (or integers!) • Regular memory access patterns