A paper from Jason Moore's group looks very exciting. They took their MDR kernal and parallelised it using CUDA across three NVidia GTX 280 GPUs.

That's a total of 840 cores!

Moore's law is the extrapolation of the observation that historically the number of transistors that can fit on an integrated circuit doubles every 18 months. This generally leads to a proportional doubling of performance for all dependent components; however heat dissipation issues have stunted the expected clock rate growth of single core processing units, which, with current technology, can't surpass 6GHz.

Figure 1 (taken from Asanovic *et al*) demonstrates this problem. The solution is to use multiple cores in parallel, and while this is slowly emerging in CPUs, the gaming industry has driven multi-core processing much harder in GPUs. For problems that are geometrically parallelisable one has the opportunity to get super-cluster levels of performance increase from just a desktop computer. For example, for an exhaustive 2D epistatic scan on 300,000 markers 4.5x10^{10} independent calculations need to be done, so parallelisation is an intuitive way to solve this computational problem.

It doesn't, however, solve the resulting multiple testing statistical issues.

Typing and design by Gib Hemani

2009-10-22 17:37:31