Improving Medical Research with GPUs
May 15, 2012, GPU Technology Conference, San Jose, CA—Scott Rupport from Lenovo described their efforts to support medical researchers. The advent of GPU computing and visualization is changing the nature of the scientific investigations.
They help researchers by supplying the compute platforms with Xeon CPUs and Quadra GPU boards. These platforms facilitate the testing and validation of new code for the simulation of various hypotheses. These new capabilities are changing the nature of the investigations, from direct measurements of many data sources to direct simulation of physical processes.
Tanmay Dharmadhikari from Beckman-Coulter demonstrated how the compute hardware can improve the science. His area of interest is flow cytometry, a technique for identifying and sorting cells and their components (as DNA) by staining with a fluorescent dye and detecting the fluorescence usually by laser beam illumination. This requires locating and tracking suspended particles in a fluid and spans many levels.
They have to detect parameters like size, complexity, absence or presence of proteins, and manage all of this while maintaining operational control of optics and fluidics. The researcher is looking for distinguishing features in size, scatter, fluorescence across many variables and the number of variables continues to increase as we learn more about biology.
The challenge is to classify upwards of 10M cells within 10-18 measurements each. The data set can exceed 400MB and the interpretations of the data have to cross-correlate the many variables. They generate gates or windows of interest and create plots for each variable. The area of interest in any plot has to be related to all of the other variables and their windows. This visualization process requires immense compute resources.
The value of a high-performance CPU-GPU platform is that the acceleration enables increased sample sizes and, therefore, a grater likelihood of finding causality within the spectrum of measured cells. In addition, the higher speeds possible for viewing allow for interactivity and discovery, enhancing the chance for detecting subtle patterns.
The software flow is a follows: compensation for coordinate transforms and sizes, compute the gates or windows of interest, sort by parameters, and generate plots. The plots require identification of in-and out-of gate limits and scaling so the plot has resolution at the areas of interest. The programs also generate statistics on all of the measured parameters. All told, this set of operations requires over 200 GFLOPs and is not possible on any general purpose computer. The processing requires the maximum possible acceleration to be of value to the researcher.
