high performance computing on graphics processing units: hgpu.org

The multi-directional (MD) technique is a general purpose tool for optimisation, that is, finding the global maxima or minima of some objective function in a given domain. Any function that produces a relatively continuous surface may therefore be suitable. Using a graphics processing unit (GPU) for MD optimisation demonstrates an increase in speed of up to 400-fold compared to using a central processing unit (CPU). More than a 100-fold speed up was seen across a range of problems. This was achieved despite non-trivial amounts of branching in the algorithm. The main problem investigated was a particular form of portfolio optimisation. As a further test, the same algorithm was used to find the minima of the Schwefel function, which has several local minima and is a particularly difficult candidate for numerical optimisation. The great exibility of the MD technique and the performance obtained strongly suggests that the GPU implementation has great potential in finance.