• What is GPU Computing
  • GPU Applications
  • Servers and Workstations
GPU Applications

GVL Esther with GPU Acceleration

Esther is GVL's flagship product. A general purpose valuation and simulation engine for large OTC bank portfolios, GVL Esther is packaged as a software-hardware-data solution and delivered as an appliance capable of running mega-models for CVA/DVA/FVA in addition to running margin and capital simulations.


Traditionally, taking advantage of GPU accelerated computing meant firms took existing algorithms, identified the available parallelism, and ported them to the GPU using CUDA. This "retrofitting" process worked well for some firms. Other firms, however, can run into scalability limits as higher levels of portfolio and scenario parallelism may be locked into assumptions about job scheduling and library granularity. 

GVL went in the opposite direction from the beginning and re-designed the entire mathematical framework of finance around one of the things GPUs can do best: single precision matrix multiplication. Founded by a mathematician in 2006, the birth-year of CUDA, GVL’s latest release of the Esther software architecture for portfolio simulation is rapidly gaining industry recognition. It is now marketed as a combined software-hardware-data solution in partnership with Supermicro and ICAP and is licensed to TriOptima to power a centrally distributed analytics service, triQuantify.

Designed to tackle the single most massive and demanding computational tasks in financial engineering, GVL Esther is an in-memory application that executes on large heterogeneous boards with acceleration. Tailored to Supermicro equipment and accelerated by NVIDIA K10 GPUs, GVL Esther achieves a sustained performance of about 10 TF/sec on each board. A handful of servers in a compact cabinet enclosure is capable of processing workloads far in excess of what can be achieved by traditional grid computing solutions.

The GPU-centric mathematics developed at GVL departs significantly from traditional Financial Mathematics. It bypasses the use of special functions, closed form solutions, double precision arithmetic and semi-implicit PDE methods, replacing the entire mathematical toolkit with third level BLAS methods in single precision reducing down to variants of SGEMM, the matrix multiplication routine in single precision at which GPU co-processors excel.

GVL Esther not only supports nearly all traditional pricing models, it goes much further. By freeing up modelers from the straight-jacket of closed form solvability and various other bespoke shortcuts, Esther's strategy of solving models by the sheer arithmetic brute force of GPUs allows to simulate financial markets achieving groundbreaking levels of economic realism and precision.

> Consistent and realistic models provide more reliable results
> Numerical approach eliminates constraints of models with analytic solutions
> Comprehensive modeling of wrong way risk
> New hardware enables fast turn-around times and increased number of MC paths which enhance accuracy and granularity; full loss
   distributions enable more sophisticated analysis of results
> Analytics and calibrations from market data from ICAP Information Services and other leading market data suppliers
> Client specific modeling and calibration

> Light weight CVA/FVA acceleration solution
> TriOptima press release
> Supermicro whitepaper



Cumulative loss distribution with pins denoting CVA expected values XVA = CVA - DVA + FVA + CLD
Total return distributions obtained with nested simulations with a total of 100 million scenarios and 20 billion portfolio valuations.

Recommended Professional Products

Tesla is designed to deliver the highest computational performance with the most reliable numerical accuracy, and are available and supported by the world's leading professional system manufacturers.

K10 Benefits

Highest Computational Performance
> High performance algebraic engine
> Large dedicated memory
> High-speed bi-directional PCIecommunication
Best Supported
> Professional support network
> Long-term product lifecycle
> 3 year warranty
> Windows remote desktop support

Recommended Configurations

Algebraic Server
> Supermicro 2027GR-TRFHT Server
> Dual socket Intel Haswell CPUs
> 128 GB system memory
Monte Carlo Server
> Supermicro 4047R-7JRFT server
> 4-socket INTEL Haswell CPUs
> 2 TB of system memory