Don't Miss: MWC 2017 Highlights
The Securities Technology Analysis Center (STAC), presented the results recently at a conference in London to showcase its latest benchmark, STAC-A2, the first benchmark to compare the performance of traditional Intel CPUs with NVIDIA’s K20X GPU accelerators – ultra-high performance processors that dramatically accelerate the processing of massive and complex datasets. GPUs, (Graphics Processing Unit(s) are a key component in video gaming. They are extremely fast but can often be very difficult to program. Details of the benchmark and the test are on the STAC Web site.
“The independent results show that GPUs deliver considerably higher performance than Intel Xeon CPUs – up to 6x faster,” according to NVIDIA.
Peter Lankford, founder and director of STAC, said that the financial services industry has a lot of workloads where it is not so much about I/O but speed of computation.
“The big banks have supercomputers that rival some of the national labs in terms of size,” he said. “One bank has 100,000 cores to do simulations to understand risk, and that has only become more important in recent years. Another big bank plans to double their compute capacity. It is a major expense for them and a huge resource, so they are all very interested in the latest technologies. One of the drivers is there is a great need to compute risk before a deal is done,” Lankford added. “It used to be that risk was run overnight and trading desks were given guidelines and they’d go along and re-run the risk at night and see how they did. The new focus is on giving the trading desks tools to understand risk before they make a deal.”
Banks want the results, but they are often constrained by space, power and cooling in their data centers. STAC now measures speed alone, but it plans to expand its benchmarks in 2014 to include computing power per unit of power or space. GPUs offer high speeds in a smaller footprint than CPUs.
For now STAC results allow a vendor to either increase the speed of their calculations or increase their scope or accuracy.
“You may want to include more financial instruments in your calculations, or the instruments may be more complicated to price, or run more simulations.”
For a fixed workload, the fewer resources you need to run it is important, Lankford added. It might be important to reduce costs, but is usually a combination of all those things that the banks are looking for. “It is important for them to understand the tradeoffs of different computing architectures.”
STAC A2 is a particular calculation that is pretty common, but not a trivial one. It takes some real work to make it efficient, he added.
“Like all STAC benchmarks, both user firms and vendors participated in its creation, but only the users can vote. That keeps things focused on real business problems.”
STAC allows anyone to register and log into its site and see much of this information; some material is reserved for members.
“We do benchmarks with a baseline model, and then we also vary the problem size in all three dimensions so people can see how the system scales,” Lankford explained. “This was the first system that could run the baseline problem size in real-time — less than one second.”
That’s important because one of the big trends in the market is the ability to do risk management in real-time. That’s useful when a trader has someone on the phone and needs to price a deal right away.
“With this boost in computational performance,” NVIDIA said in its announcement, “financial firms can quickly run complex risk analysis scenarios and many more of them, leading to better insights and reduced risk at a lower overall IT cost.”
Lankford said that STAC is also planning to run tests with FPGAs.