Skip to main content

GPU acceleration for simulating financial systems

Project Member(s): Xu, Y., Liang, J.

Funding or Partner Organisation: Magellan Financial Group
Magellan Financial Group

Start year: 2019

Summary: There are two primary uses of the infrastructure: strategy optimization with a neural network and grid search. The former enables the user to find an investment strategy that maximizes multi-client utility over a long investment horizon. The latter enables the user to compare specific pre-defined strategies. Both optimization and grid search use the same underlying computation. This computation is carried out on Amazon Web Services on Sagemaker, usually on 72 CPU Core machines. In general, a single run takes between 10 seconds and 5 minutes, depending on the complexity of the simulation, the orizon, the frequency, and the vectorization degree. While this is acceptable, it means that training the networks can often take several days. The PyTorch library comes with GPU acceleration of the numerical calculations. The method for deploying to GPU is usually called “to_cuda”. In principal, this can speed up the computation between 5 and 50 times. The existing code base has not implemented the GPU functionality. The goal of this project is to incorporate the GPU acceleration feature of PyTorch and run on AWS GPU machines to substantially speed up our simulations for both grid search and optimization.

FOR Codes: Pattern Recognition and Data Mining, Application Software Packages (excl. Computer Games), Modelling and simulation, Application software packages