Computational Fluid Dynamics on Cerebras CS-1

There is an emerging processor technology that eliminates many of the performance limitations that are inherent in existing supercomputers. Existing supercomputers only deliver a very low fraction (0.5-3.1%) of their peak floating-point performance on Computational Fluid Dynamics (CFD) codes. This performance limitation stems from low bandwidth and high latency of the off-die data access. Emerging wafer-scale processors have the potential to eliminate these performance limiters and could become cost-effective "single chip supercomputers". The first wafer-scale processor to arrive on the market is the Cerebras Wafer Scale Engine (WSE), and the first system powered by such a processor is Cerebras CS-1. While the WSE was developed to be an artificial intelligence engine, its hardware arrangement is also ideal for structured CFD codes. Thus, NETL, in partnership with Cerebras, has been exploring the possibility of performing MFIX (open-source multiphase CFD code) simulations on CS-1. This presentation will provide preliminary findings from the exploratory study. Cerebras and NETL have developed the first (bare-bones) CFD solver to run on CS-1 and have found that the CS-1 could conduct fast, cost-effective CFD computations. We will discuss the CFD-code implementation, current performance, and future development plans.