I attended a good set of presentations from one of our implementation partners, D-Fine, yesterday on high performance computing (HPC). They had one of their hedge fund clients present talking about some work they had done in valuing derivatives on a cluster, however the most interesting part of the event was talking about Graphics Processing Units (GPUs) being used within HPC grids and clusters.

Seems that GPUs such as those made by Nvidia offer performance levels that easily exceed that possible on traditional CPU-based HPC solutions. Key to it seems to be the GPU being more specifically designed for vector and floating point operations with several hundred processing cores being available on one chip, whereas CPUs are understandably designed for more general processing requirements. So if the problem is well suited to GPUs (not always the case), one of the technologists from QuantCatalyst said that with optimisation, the performance could be improved by up to 10,000 times relative to CPU based cluster solutions.

As with Chris’s article on Solid State Drives then GPU usage is not exactly a free lunch, with specific tools and compilers needed to be used for the moment until a more generic, industry accepted abstraction comes along. At 10,000 times quicker it might be worth the effort though – maybe Hank Paulson should take a look given the CDOs he will need to be valuing any day soon…