Wafer-Scale HPC for Scientific Simulations

2 posts · Mar 2025

What this trail covers

The Cerebras Wafer-Scale Engine (WSE) is being used to accelerate High-Performance Computing (HPC) tasks, demonstrating significant speedups in solving sparse linear equations for computational fluid dynamics and molecular dynamics simulations. The WSE's architecture, with its high fabric bandwidth, single-cycle core-to-core latency, and fine-grain message granularity, overcomes limitations of traditional HPC systems and enables real-time or faster-than-real-time performance in scientific simulations. This post details a specific stencil algorithm that achieves over 200x performance compared to an NVIDIA A100 GPU for seismic wave propagation modeling.

All Cerebras trails

Mar 2025Beyond AI for Wafer Scale Compute: Setting Records in Computational Fluid Dynamics - CerebrasThis post details the successful application of the Cerebras CS-1 wafer-scale system to solve sparse linear equations for computational fluid dynamics, achieving over 200x speedup compared to a leading US supercomputer. It explains the mapping of a 3D mesh of cells to the CS-1's cores, the utilization of 16-bit arithmetic for vector multiply-adds and 32-bit for dot products, and the efficient handling of sparse matrix-vector products through on-wafer communication and a novel global color assignment. The post emphasizes the CS-1's advantages in memory bandwidth, low-latency fabric, and its ability to sustain high communication rates, contrasting it with the communication bottlenecks in conventional supercomputers.
Mar 2025TotalEnergies and Cerebras Create Massively Scalable Stencil Algorithm - CerebrasThis post details a novel implementation of a stencil algorithm on the Cerebras CS-2 System using the Cerebras Software Language (CSL). It highlights the system's ability to handle memory-bound workloads by leveraging its massive memory bandwidth (20 PB/s) and efficient neighbor-to-neighbor communication. The implementation uses tailored localized broadcast patterns for concurrent send, receive, and compute operations at the hardware level. A weak scaling study on the Minimod benchmark (25-point stencil) showed constant solve time regardless of problem size, indicating compute-bound behavior, and outperformed an A100 by over 220x for the largest problem size. Weak scaling efficiency was over 98%. Roofline analysis confirmed compute-bound behavior with a total throughput of 503 TFLOPs on the WSE-2.
Mar 2025TotalEnergies and Cerebras: Accelerating into a Multi-Energy Future (Updated) - CerebrasThis post details a specific stencil algorithm for seismic wave propagation modeling that achieves over 200x performance improvement compared to an NVIDIA A100 GPU on the Cerebras CS-2 system. It highlights the wafer-scale architecture's advantage in overcoming memory and communication bottlenecks in HPC workloads, enabling compute-bound performance. The post also mentions the Cerebras Software Language (CSL) and SDK for developing custom kernels.
Mar 2025Sandia Awards Major Contract to Cerebras Systems - CerebrasThis post announces a multi-year R&D contract between Cerebras Systems and Sandia National Laboratories (along with LLNL and LANL) to accelerate advanced simulation and computing applications for the nation's stockpile stewardship mission. The partnership aims to achieve a 40X performance improvement over the forthcoming El Capitan exascale supercomputer by leveraging Cerebras's WSE-2. The post highlights previous successful collaborations in Molecular Dynamics (achieving ~50X time-to-solution improvement), Computational Fluid Dynamics (outperforming a US supercomputer by >200X for PDE codes), and Seismic Modelling (outperforming a modern GPU by >200X). It emphasizes the WSE-2's wafer-scale integration, massive core count, on-chip memory, and native acceleration for unstructured sparsity as key differentiators for HPC workloads.
Mar 2025Cerebras Breaks Exascale Record for Molecular Dynamics Simulations - CerebrasThis post details a new benchmark for molecular dynamics (MD) simulations using the Cerebras Wafer-Scale Engine (WSE), achieving 180x speedup over Frontier for 800,000-atom simulations using the EAM potential. It highlights the WSE's ability to complete simulation steps in microseconds, enabling millisecond-scale simulations that were previously infeasible. The post details the computational challenges of parallel computation at the atom-per-core scale, including communication bandwidth, latency, granularity, and multicast communication, and explains how the WSE's architecture (fabric bandwidth, single-cycle core-to-core latency, native 32-bit message granularity, dataflow marshalling, and neighborhood multicast pattern) addresses these challenges.
Advertisement
Mar 2025Cerebras Wafer-Scale Engine Achieves 210x Speedup Over NVIDIA H100 - CerebrasIntroduced a dataflow matrix-free finite volume solver for the Cerebras WSE, achieving a 210x speedup over NVIDIA H100 GPUs for geological carbon capture and storage (CCS) simulations. This solver eliminates the need to store large Jacobian matrices by computing Jacobian-vector products on the fly, aligning with the WSE's strengths in localized communications and single-level memory. The implementation on the Cerebras CS-2 reached up to 1.217 PFlops and demonstrated strong scalability. Algorithmic enhancements included asynchronous communications, vectorization, and memory optimization.
Mar 2025New Tool Generates Stencil Codes Two Orders of Magnitude Faster on Cerebras WSE Than on GPUsThis post introduces StencilPy, a new tool developed by researchers that generates stencil codes for scientific computing. StencilPy translates high-level stencil definitions into optimized Cerebras Software Language (CSL) code. The post details performance benchmarks showing the Cerebras CS-2 system achieving 95x speedup over CUDA on an NVIDIA H100, 292x over HIP on an AMD MI210, and 570x over OpenMP on an AMD Genoa for stencil computations. It highlights the WSE's architecture, including its massive on-chip memory, high bandwidth, and distributed core design, as key differentiators for scientific computing tasks like stencil computations.
Mar 2025US DOE Achieves 88x Performance Speedup with Cerebras CS-2 Over H100 in Materials ModelingThis post details a new benchmark demonstrating an 88x performance speedup for materials modeling (Ising model) on the Cerebras CS-2 compared to an NVIDIA H100. The implementation utilizes a checkerboard decomposition strategy on a single column of PEs, folding the domain to handle periodic boundary conditions efficiently. The results show a flip rate of 61,854 flips/ns on the WSE, significantly outperforming V100 and H100. This work validates the WSE's suitability for low data-intensity HPC algorithms.