HiPER Risk Engine™ powers high performance computing
Curt Burmeister, chief technology officer and co-head at SS&C Algorithmics, explains how the technology provider’s expertise in high performance computing can enable insurers to significantly speed up financial modelling and large-scale simulations, as well as deliver cost savings
Can you explain high performance computing?
High performance computing (HPC) dates to the early 1960s with the advent of the “supercomputer” when Control Data Corporation released the CDC 6600. In the 1980s, companies like Cray Research and Thinking Machines led the market with super computers based on techniques, such as vector processing or massively parallel computing. HPC consists of multiple techniques to improve computation performance to solve very large problems in a reasonable amount of time.
What expertise does SS&C Algorithmics have in high performance computing to support insurers?
We have been applying HPC techniques to calculate market risk, credit risk, and capital calculations for over 20 years.
Our clients often run tens of thousands (and higher) of simulations/scenarios for these calculations, but are constrained by the amount of hardware available and the allotted time to perform the calculation. For many years, we have focused on two main techniques:
- Parallel computing – dividing the calculations across multiple servers
- New mathematical algorithms – developing mathematical algorithms that are faster than standard algorithmics without sacrificing accuracy.
More recently, we have developed a new HPC engine, SS&C Algorithmics HiPER Risk Engine™ (HiPER™), that uses multiple techniques to improve performance, including vector processing, parallel computing, low-level code optimisations and new mathematical algorithms.
Although many of the simulations insurance companies run can be programmed to execute on Graphics Processing Units (GPUs), there are several challenges. One challenge is cost. Applications written for GPUs require specialised hardware, and as a result, most companies must purchase new servers to run these applications. A second challenge is that GPUs were designed to render graphics and thus, programming them for more general applications can be challenging.
Instead of GPUs, we decided to use vector processing for HiPER™ because vector instructions are available on all modern Central Processing Units (CPUs – designed for a wide variety of processing tasks), but most applications don’t utilise them. This means that HiPER™ can run on the same servers that an insurance company currently owns, and no new hardware investments are required.
What are the mechanics of your HPC solution?
HiPER™ uses several techniques:
1. Vector processing
- Calculations and data are organised to allow for vector processing on arrays of data
2. Parallel processing
- Execution across multiple nodes and cores
- Dynamic scheduling and planning
- Ability to add/remove compute nodes during a run
3. Low-level code optimisations
- Cache-aware data structure
- JIT compilation (just in time)
- Late-stage code optimisations
- CPU instruction set targeting
4. New mathematical algorithms
- Valuation models written to apply adjoint differentiation
Can you provide some examples of how HPC compares with traditional / legacy approaches?
The HiPER™ engine is orders of magnitude faster (factors of 10) than the Algorithmics’ legacy engine. Our clients are seeing performance improvements ranging from 30x to over 1000x. This translates to significant improvements in the time required to run a simulation. Jobs that previously took over 20 hours and now are completed in minutes.
These large improvements provide clients with multiple benefits. Now, they can:
- Run more simulations on the same hardware
- Achieve cost savings by reducing the hardware needed
- Change the operating model and achieve multiple runs per day, running on-demand based on market movements, re-run immediately after fixing data errors, and more.
How do you expect HPC to evolve over the next 3–5 years?
The current trend is supercomputers and exascale systems (10¹⁸ FLOPS), which are massively parallel machines with a hybrid mix of CPUs and GPUs.
Another technology that has promise for some types of financial modelling is quantum computing, but most estimates are that it will be another 5-20 years before it is being used for real-world applications.
Neither of these are likely to be applied to insurance modelling in the near term. But looking at a 50-year time horizon, they will likely be important as these technologies improve and their costs come down.
If we compare one of the early supercomputers, the Cray-1, released in 1976, to a MacBook Pro released in 2025, we see that the MacBook has approximately 238,000 times the raw compute power compared with the Cray-1 (160 MFLOPS vs 38 TFLOPS), but the Cray-1 is ~8000 times more expensive (in today’s dollars) than the MacBook.
In the meantime, insurance companies should be looking at engines like HiPER™ that apply today’s HPC techniques and use them to speed up financial models and large-scale simulations.
