BNP Paribas will use Nvidia for its GPU solution but competition in the market is set to open up

Nvidia has been the most vocal graphics processing unit (GPU) vendor, promoting the use of its products in general-purpose computing environments where they can provide the benefits of massive parallelism. In the financial sector, Monte Carlo modeling, with its large number of calculations to simulate multiple paths across which an average is then computed, is a natural target for the technology.

BNP Paribas is now migrating to GPUs because it has moved to a more advanced Monte Carlo model over the last three years and found that this model was testing the limits of the clusters of CPU-based machines that it has used until now. In essence, the problem arises from the fact that its new model uses simple, so-called vanilla, options as outside input to enable it to price exotic options more accurately, in that it takes real-world pricing into account more precisely.

This means that closed-form solutions, those that are expressed exclusively in elementary functions, are no longer applicable to vanilla options. Instead, factoring the pricing of vanilla options in so as to calculate more accurate pricing of exotics requires a full-scale Monte Carlo model with more complex functions.

This shift in the way the bank works means that vanilla options now have to be calibrated just as exotics have traditionally been, tuning the parameters of the model so as to equal prices in the market. The difference is that vanilla options are far greater in number, which magnifies the computing challenge hugely. For this reason, BNP saw its CPU-based clusters reaching the limit of their capabilities and decided to migrate to GPUs. The issue was one of throughput and power, i.e. the amount of floating-point operations per second (FLOPS) that are achieved for each Watt of energy expended.

The problem with using GPUs for general-purpose computing (an activity referred to as GPU computing) is that they are constructed to interpret triangles and textures so as to produce graphics, which means that non-graphics applications need to be rewritten to run on them. More specifically, code needs to be recompiled such that algorithms can be executed on a GPU. To this end, in 2007 Nvidia launched a C compiler and a set of development tools, collectively called CUDA, which is the environment that enables GPU computing with its products.

Nvidia actually implemented an architecture for GPU computing in CUDA, while the programming environment that developers can then use to access that capability is called C with CUDA extensions. As for the rewrite issue, Nvidia points out that code designed to run on CPUs also has to be rewritten when it is going to run on a larger number of them, so there is no seamless scaling even if a company opts to remain in an all-CPU environment. Furthermore, once code has been rewritten for the GPU, Nvidia argues that it will scale seamlessly across future generations of the graphics processor, without further recoding.

BNP Paribas's experience is that it took two people working half time around nine months to rewrite its software in C with CUDA extensions, plus another six to verify its accuracy with regards to the legacy code. The bank reckons this process would be shorter if undertaken now due to the availability of double-precision silicon. This makes it easier to determine whether small differences in results, as compared to the CPU version of the software, are due to the code itself or simply a question of precision.

At the moment, anyone porting software to run on GPUs by using C with CUDA extensions is tied to Nvidia's hardware. However, there is already an emergent standard, by the name of OpenCL, which promises to enable GPU computing on any GPU silicon. It is still early days for that technology and Nvidia has actually announced support for it. AMD, the other big player in GPUs as a result of its 2006 acquisition of Nvidia's main rival, ATI, has alpha versions available.

Another vendor about to throw its hat into the GPU ring is Intel, with its Larrabee platform, which is due out next year. The company has not made a statement on its intentions with regard to GPU computing yet but is likely to support OpenCL at some point.

Meanwhile, IBM sees potential in certain capital markets applications for its Cell processor, which can almost be thought of as a hybrid CPU/GPU architecture, in that it combines a general-purpose Power Architecture core with streamlined co-processors to accelerate multimedia and vector processing applications. This is a form of parallelism akin to what GPUs can offer. Both Intel and IBM are members of the working group that came up with the OpenCL spec.

Rik Turner

http://www.datamonitor.com

Republication or redistribution, including by framing or similar means,
is expressly prohibited without prior written consent. Datamonitor shall 
not be liable for errors or delays in the content, or for any actions 
taken in reliance thereon

For full details on International Business Machines Corporation (IBM) IBM. International Business Machines Corporation (IBM) has Short Term PowerRatings at TradingMarkets. Details on International Business Machines Corporation (IBM) Short Term PowerRatings is available at This Link.

For full details on (BNPZY) BNPZY. (BNPZY) has Short Term PowerRatings at TradingMarkets. Details on (BNPZY) Short Term PowerRatings is available at This Link.