The SciGPU website is about to undergo a migration. In the meantime, the latest information on upcoming events will be available via a public Google calendar. Search Google calendars for "SciGPU" to subscribe.

Location: Maxwell Dworkin, 33 Oxford Street, Room 323

Speaker: Won-Ki Jeong, Research Scientist, Harvard IIC/SEAS

Abstract

In order to model the performance of any HPC system, certain characteristics must be known. Those relevant to a GPU are

1. Memory copy latency between device and host
2. Bandwidth between device and host
3. Kernel launch latency

This article analyses some of these for the Orgoglio cluster at Harvard. Orgoglio pairs S1070s with Harpertown Xeon processors and ECC RAM.

The Murchison Widefield Array is using a real-time GPU correlator to enable engineering and early science for a 5% prototype. Read more about how this system works! See online coverage of the MWA showcasing GPU computing efficiency, as described at the NVIDIA GPU Technology Conference, San Jose 2009. Take a look at the related talk, Diesel-Power GPU Computing.

In our recently submitted paper (R. Olivares-Amaya et al, JCTC), the Alan Aspuru-Guzik group has presented a new implementation of the quantum chemistry method RI-MP2 (resolution-of-the-identity second-order Møller-Plesset perturbation theory) accelerated using GPUs and the MGEMM library published on this website. For the 168-atom valinomycin molecule in a cc-pVDZ basis set, we observed speedups relative to CPU double-precision results of 13.8x, 7.8x and 10.1x using single-, double- and our new mixed-precision general matrix multiply (SGEMM, DGEMM and MGEMM), respectively.

A popular account of the SciGPU project has been posted online by the Harvard News Office.

Writer Alvin Powell describes the "trio of projects at Harvard whose massive computing needs have prompted investigators to join forces to pioneer new computing techniques that will benefit not just radio astronomy, but quantum chemistry and neuroscience as well."

With the initial port of the MWA pipeline complete, I thought it would be a good time to reflect on some of the lessons learned. Particularly relevant have been the experiences in trying to maintain a CPU and GPU based version of the code, and managing the data when everything is located in two places.

Matrix-matrix multiplications are common in quantum chemistry calculations, and can benefit enormously from GPU acceleration. Although NVIDIA provides an implementation of the BLAS *GEMM routines with its CUDA distribution, two key problems exist when trying to use these from existing code

The following news was released April 2, 2009 by NVIDIA Corporation. The company's release can be found here.

The SciGPU collaborators welcomed four students who came to Harvard for NSF-funded Research Experiences for Undergraduates during summer 2009: Dominik Gothe (University of South Carolina; astronomy), Matthias Lee (Wentworth Institute; time series analysis), Beatrice Perez (University of Puerto Rico; quantum chemistry), and Bo Wang (University of Pittsburgh; neuroscience).