DateApr. 4, 2008
Speaker Martin Herbordt (Boston University)
TopicHigh Performance Computing Using FPGAs
Abstract: The current effervescence in computer architecture is leading to serious (re)examination of FPGAs as augmentations to CPU-based processors. And not just on the I/O bus, as has been done for many years, but also on the FSB and even on the processor chip itself. This interest has largely been motivated by reported per-node accelerations of 100x or more. The unique issues in "programming" FPGAs, however, have left most application writers waiting for better tools before giving them serious consideration.

In this talk we first review the state of the art of FPGA-based high performance computing. The bulk of the talk then consists of highlights from our research in developing FPGA applications for Bioinformatics and Computational Biology. This has two parts. The first is on methods of algorithm development, answering the question "How we know when we have succeeded?" This part doubles as a discussion of various effective FPGA computational modes. The second part is a case study: discrete event simulation of molecular dynamics (DMD). DMD uses simplified discretized models, enabling simulations to be advanced by event rather than time-step, with a resulting performance increase of several orders of magnitude. Our primary result here is a microarchitecture for DMD that processes events with a throughput equal to a small multiple of the FPGA's clock, resulting in a hundred-fold speed-up over serial implementations. Of particular interest is that this result appears difficult to achieve using alternative acceleration methods, independent of system cost. We end the talk by considering what must happen to make FPGA-based systems broadly effective for HPC, and the likelihood of that occurring.