we do implementations within the CUDA framework for many months now. In Nov-09 we installed a nvidia Tesla personal supercomputer in a server-rack with 2 Quadcore processors, 24 GB memory, triple channel architecture, with 2 Tesla C1060 and 1 GTX260 GPUs. A machine that let you "feel" the heat. A-wild-thing-that-makes-my-heart-sing.
We ran comprehensive tasks the UnRisk way: valuing various options under various models (from BS to Heston - represented by threads with Montecarlo, Quasi-Montecarlo, Brownian Bridges techniques) with very tight constraints related to accuracy. Including correct model calibration.
Impressive results compiled: For valuations where single precision is enough the GPU speed-up compared to 1 CPU is fantastic (over 200 for asian options).
When double precision is essential (calibration!) the recent GPUs are not so superior. We expect a great boost with the recently released FERMI architecture .
We hope to enter into concrete tests with FERMI soon.
As always we are not the quick-releasers. We take our time to do it right at once.
But taming those supercomputers with low cost of ownership is exciting. And we look forward to making them valuable for quants with highest-performance-at-accuracy requirements.
No comments:
Post a Comment