Cerebras WSE-3 smashes GPU stencil computations with 342x speedup

Stencil computations—critical for fluid dynamics and climate simulations—are traditionally memory-bound on GPUs due to the "Memory Wall." A new preprint from researchers Elia Belli and Daniele De Sensi shows that Cerebras' Wafer-Scale Engine (WSE-3) can overcome this limitation. They developed CStencil, a framework that maps 2D stencil kernels onto the WSE-3's massive core parallelism and high-bandwidth on-chip SRAM. For fair comparison, the team adapted ConvStencil (a state-of-the-art GPU solver) from double- to single-precision and ran it on an NVIDIA A100. The results are striking: CStencil achieves up to 342x speedup over the GPU baseline, with roofline model analysis confirming near-perfect utilization of compute and memory resources.

This work demonstrates that the WSE-3's dataflow architecture—originally optimized for AI workloads—can be repurposed for scientific computing without sacrificing performance. The key enabler is the distributed SRAM and mesh interconnect, which keep data local and eliminate off-chip memory traffic that throttles GPUs. The findings open a promising path for hybrid HPC/AI systems, where wafer-scale engines handle memory-bound kernels while GPUs tackle compute-heavy tasks. The paper is currently under review and available on arXiv.