AsHES 2018 Keynote
Michael J. Wolfe · 2018
Accelerated systems have been around in HPC for 50 years or more. IBM produced the 2938 Array Processor for mainframes in 1969, and Floating Point Systems developed the AP-120B array processor for minicomputers in 1976. Ten years ago, the IBM PowerXCell coprocessor was used in Roadrunner, the fastest machine in the TOP500 list. Yet, repeatedly, the improved performance from each accelerator was eventually obviated by faster and more capable CPUs. Now we see GPUs becoming common as compute accelerators, FPGAs achieving some success, and other accelerator ASICs being designed for specific applications. How do today's accelerators differ from the previous generations? What will it take to make the current move to accelerators successful in the long term? Is there something unique about the jump from petascale to exascale computing that begs for accelerators? How will compilers and languages have to evolve for these systems? We explore these questions and more.