Numerical computations widely used in quantitative finance can be harassing for standard servers and CPUs. Hence it is natural to consider to use dedicated hardware. The names that come often that for are "FPGA" and "GPU". Here are some first simple facts about these two kind of hardware:
- GPU is very good for parallel computations (already underlined in remarks)
- but bad for memory sharing between the master software and the GPU-hosted library
- FPGA is good for parallel computations too (but harder to use compared to GPU)
- FPGA works well for working on data flows
- if you need to solve numerically a PDE with few parameters (coefficients of the derivative terms) or to run intense Monte Carlo simulations: GPU is nice;
- if you need to compute on the flight analytics for high frequency trading: FPGA is nice.
For DPE solving using GPU, one must carefully think about how to parallelize such computations; this is not a straightforward question.