Efficient processing of numeric algorithms is critical to the finance industry, from trading and risk management to collateral management and portfolio optimization. Much of this analysis employs numerical methods that are very computationally intensive. Large firms have datacenters packed with thousands of compute nodes dedicated to this task. The workload has become more taxing as market volatility increases, portfolios get more complex, and more trading desks incorporate risk information into their decision making, requiring shorter turnaround times for analysis.
A firm's goals with respect to numeric analysis can depend on the extent to which it is focused on cost reduction, revenue expansion, or regulatory compliance. It may wish to reduce the power and space required for calculations, to analyze more information, to analyze more scenarios for more users, to increase the accuracy of the calculations, or all of the above.
No matter what the requirement, innovation in technology infrastructure has proven to be a crucial enabler of these business goals. New processors, memory, and interconnects, as well as innovative software libraries, development tools, and grid software can create favorable shifts in the tradeoffs (e.g., accuracy vs speed, capacity vs power consumption, etc.).
As a quick scan of recent STAC Summits attest, these sorts of "big compute" issues are a hot topic in the STAC Benchmark Council. A few years ago, the Council began to develop standard benchmarks based on computationally intense workloads. These standards enable vendors to publish apples-to-apples comparisons and end-user firms to baseline their existing systems using the same tests.
The first such standard to emerge is STAC-A2, a vendor-independent benchmark suite based on real-world market risk analysis workloads. STAC-A2 was specified by banks and made actionable by leading HPC vendors. STAC-A2 satisfies all of the requirements important to end-user firms: relevance, neutrality, scalability, and completeness. Numerous STAC-A2 results have been published in the past two years, on systems using CPUs, GPUs, and other co-processors. Work on additional platforms, programming frameworks, and server form factors continues.