After having re-entered the data center server market in 2017 with 1st Gen AMD EPYC™ CPUs, formerly codenamed “Naples”, AMD has since gone on to release several new generations of processors, with the latest 5th Gen AMD EPYC processors, formerly codenamed “Turin”, introduced on October 10, 2024 at the AMD Advancing AI event. With each successive iteration, AMD pushes the boundaries of performance and power efficiency, all while delivering on the industry’s expectations of time to market.
With AMD EPYC processors now following a consistent cadence of improvement, there has been increasing focus on helping ensure suitability of these CPUs as solutions for specific verticals, such as Manufacturing, Retail, Media & Entertainment, and Financial Services.
In each of these target market segments lies a narrower set of requirements, and relevant for this blog, unique benchmarks that demonstrate the performance of AMD EPYC processors on proxy applications important to each of these verticals.
Within the Financial Servies industry (FSI) is the field of quantitative finance and risk management. As these tend to be proprietary applications, with production workloads only known to the financial institutions and closely guarded, the challenge posed to AMD is how to identify appropriate benchmarks that accurately reflect what these expert users typically deploy in their data centers.
While there are several FSI benchmarks available today, including STAC-A2, FinanceBench, and COREx, the focus of this blog is on QuantLib, a free, open-source library of tools for modeling, trading, and managing risk. In the QuantLib v1.35 release, AMD upstreamed adjustments to the benchmark to better achieve the following goals of an ideal quantitative benchmark:
- Verify that calculations return correct answers.
- Use system throughput as the overall performance metric (overnight risk is a massively parallel throughput problem).
- Express a large volume of inhomogeneous work (tasks) and allow the machine to schedule tasks as it would in a production grid environment without imposing load balancing by design.
- Minimize tail effects.
Leveraging the AMD uProf performance analysis tool to profile this workload led to the observations detailed in this related whitepaper: “Introducing a new QuantLib Benchmark Version”.
AMD benchmarked QuantLib v1.35 performance and performance per Watt across a range of dual socket systems powered by various generations of AMD EPYC processors and Intel Xeon Platinum CPUs. As the below graphs show, AMD EPYC processors provide compelling QuantLib performance and performance/W compared to the competition. These graphs are presented in the same whitepaper linked earlier.
Figure 1: Normalized relative QuantLib v1.35 performance
Figure 2: Normalized relative QuantLib v1.35 performance per TDP Watt
These significant performance uplifts on overnight risk calculations critical to quants within financial institutions around the world can present IT departments with the opportunity to consolidate within the data center, reduce software licensing costs, and improve total cost of ownership (TCO) when replacing legacy Intel Xeon-based systems with modern AMD EPYC processor-based infrastructure.