Actuarial modeling vendors have not yet been able to translate their existing domain-specific languages (DSLs) to execute efficiently on graphics processing unit (GPU) cards. GPUs are specialized microprocessors designed to render high-resolution images and video. This specialism can be leveraged to perform floating-point arithmetic at higher degrees of parallelism than a central processing unit (CPU).
The expression of business rules does not easily translate to the kinds of operations suited to GPUs, so there is little penetration of these options. DSLs for GPUs leak too much of the implementation detail, which creates a chasm between mass-market opportunity and the promise of the increased performance. Crossing the chasm is expensive and requires specialized talent, so the application of this hardware is limited to narrow use cases, such as economic scenario generation and narrow modeling use cases (e.g., hedge calculations with variable annuity embedded guarantees).
The pace of innovation in the cloud has provided an alternative, using commodity CPU instructions. The density of CPUs in a server has increased, as has available memory and memory bandwidth. The restrictions of input/output (I/O) communication between the CPU, its cache, its memory and hard disk, and across networks, are less impactful thanks to solid-state drives (hard disks without a mechanical head that provide much better random-access behavior and more I/O operations per second) and remote direct memory access (RDMA) networks, which allow direct memory access across servers at high throughput and low latency.
Combined, this delivers significant performance at a lower cost than GPUs due to more optimal thermal efficiency and lower power consumption. Even though transistor density is no longer doubling each year (as predicted by Moore), cloud innovations are finding ways to stay ahead of demands without a paradigm shift to alternative hardware. As the cloud evolves without any capital investments, vendors quickly can bring new options to customers without complex changes to the models.
Copyright © 2020 by the Society of Actuaries, Schaumburg, Illinois.