Using NVIDIA GPUs for Computational Finance

HPC (High Performance Computing) has become very important in the finance sector. Based on the list of the world's TOP 500 supercomputers it is arguable that the finance industry is now the leading HPC user in the UK.

The big news in HPC these days is the emergence of manycore GPUs as a new parallel architecture. NVIDIA's current "Fermi" GPUs have up to 480 cores on a single chip, and this is complemented by the CUDA software development environment which is based on C/C++ with a few extensions.

Some of the largest HPC systems in the world are now based on GPU accelerators:

• Press release discussing new Chinese system called Tianhe-1A which has 7168 NVIDIA Fermi GPUs and is currently ranked #2 in the world
• Tsubame 2 at Tokyo Institute of Technology has 4200 Fermi GPUs in dense HP GPU servers
• Oak Ridge National Laboratory plans a 10-20 Petaflop supercomputer called Titan

Within the finance sector:

• Bloomberg has installed almost 200 NVIDIA Tesla GPUs instead of getting about 2000 CPUs
• BNP Paribas is also using NVIDIA Tesla servers with factor 10 energy savings over the CPU alternative.
• JP Morgan has recently started using NVIDIA Tesla GPUs for risk calculations

In addition, several ISV's have developed computational finance applications running on NVIDIA GPUs:

• SciComp
• Quant Catalyst
• UnRisk
• Hanweck Associates
• Level 3 Finance

An ISV with a different approach is Xcelerit which has developed an SDK with a very clean API for streaming Monte Carlo applications on a variety of platforms. The user develops their own application in C++, defining certain classes and methods, and then through some C++ magic it is incorporated to produce executable code for NVIDIA GPUs as well as distributed multicore CPUs. A White Paper is available here.

For more information on the use of NVIDIA GPUs in finance see:

• Wall Street & Technology article
• NVIDIA's computational finance webpage
• NAG's Numerical Routines for GPUs