TechPrufer
Get started
Cerebras
Cerebras logoCerebras · Feature trail

Wafer-Scale Field Equation Modeling

9 posts · Mar 2025

What this trail covers

This post details advancements in wafer-scale processing for computational physics and engineering simulations, specifically molecular dynamics and materials modeling. It introduces the Wafer-Scale Engine (WSE) Field-equation API (WFA) and demonstrates its application in achieving exascale performance for molecular dynamics simulations, running 800,000-atom simulations 180 times faster than Frontier. The WSE's architecture, with its high fabric bandwidth, single-cycle core-to-core latency, fine-grained computation, and energy efficiency, is highlighted as a key enabler. The WFA provides a simple Python API for programming these simulations, making wafer-scale processing accessible to a wider range of computational scientists. The updated post highlights real-time computational fluid dynamics (CFD) simulations achieved with the WSE, enabling manipulation of nearly 200 million cells in near real-time, opening up new use cases for HPC applications.

All Cerebras trails
  1. Real-Time Computational Physics with Wafer-Scale Processing [updated]This post details the development and application of the WSE Field-equation API (WFA), a simple Python API for programming wafer-scale engines for computational science. It showcases real-time computational fluid dynamics (CFD) simulations on a grid of nearly 200 million cells, achieving unprecedented speed and enabling new real-time use cases. The post also reiterates previous achievements in solving field equations 200x faster and with 1000x better energy efficiency than conventional supercomputers, attributing this to the WSE's architecture: high memory bandwidth, high interconnect bandwidth, and single-cycle communication latency. The WFA addresses the challenge of programming the WSE for scientific computing by providing a high-level, domain-specific toolset that resembles NumPy or MATLAB, while maintaining near-native performance.
  2. Beyond AI for Wafer Scale Compute: Setting Records in Computational Fluid Dynamics - CerebrasThis post details how the Cerebras CS-1 wafer-scale computer was used to solve sparse linear equations for computational fluid dynamics modeling, achieving over 200x speedup compared to a leading US supercomputer. It highlights the CS-1's memory performance, on-wafer communication fabric bandwidth and low latency, and processor architecture optimized for high bandwidth computing as key enablers. The implementation mapped a 3D mesh of cells to cores on the wafer, utilizing local vector multiply-adds, sparse matrix-vector products requiring inter-core communication, and allreduce operations for dot products. The post contrasts this with conventional supercomputer cluster approaches that require coarser-grained data decomposition due to communication bottlenecks.
  3. TotalEnergies and Cerebras Create Massively Scalable Stencil Algorithm - CerebrasThis post details the creation of a massively scalable stencil algorithm for High-Performance Computing (HPC) applications, specifically for solving Partial Differential Equations (PDEs) like those in fluid mechanics, weather forecasting, and seismic imaging. It highlights how stencil algorithms, which are memory-bound due to high data access and low computation per access, are poorly suited to traditional hierarchical memory architectures (CPUs, GPUs). The Cerebras CS-2 System, powered by the Wafer-Scale Engine (WSE) with its 850,000 cores and 20 PB/s memory bandwidth, is presented as an ideal platform. The algorithm, implemented in Cerebras Software Language (CSL), utilizes localized broadcast patterns for efficient neighbor-to-neighbor communication. A benchmark case from TotalEnergies (Minimod) using a 25-point stencil for acoustic isotropic kernel simulation demonstrated compute-bound behavior on the WSE-2, outperforming an A100 by over 220x for the largest problem size and achieving virtually perfect weak scaling efficiency (>98%). The total throughput reached 503 TFLOPs, confirming a compute-bound implementation. Future work includes more complex HPC applications and hybridization with Machine Learning.
  4. TotalEnergies and Cerebras: Accelerating into a Multi-Energy Future (Updated) - CerebrasThis post details TotalEnergies' adoption of the Cerebras CS-2 system to accelerate multi-energy research, specifically seismic wave propagation modeling using stencil-based finite difference algorithms. The CS-2 system demonstrated over 200x performance improvement compared to a modern GPU. The post highlights the Cerebras architecture's advantage in memory-bound and communication-bound HPC workloads due to its wafer-scale integration, placing all processors on a single chip. It also mentions the Cerebras Software Development Kit (SDK) and Cerebras Software Language (CSL) as tools for developing custom kernels and leveraging wafer-scale computing.
  5. Sandia Awards Major Contract to Cerebras Systems - CerebrasThis post announces a multi-year collaboration 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 exascale computing platforms like El Capitan by leveraging Cerebras's Wafer-Scale Engine (WSE-2). The post highlights the WSE-2's specifications (2.6 trillion transistors, 850,000 AI-optimized cores, 40GB on-chip memory) and its advantages in handling sparse models and on-chip data flows. It also references previous successful partnerships and results in Molecular Dynamics (Stream-AI-MD, achieving ~50X improvement for protein folding), Computational Fluid Dynamics (CFD, outperforming supercomputers by 200X for PDE codes), and Seismic Modeling (outperforming modern GPUs by 200X). The Cerebras SDK and Cerebras Software Language (CSL) are mentioned as tools enabling these advancements.
  7. Cerebras 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 embedded atom method (EAM). It highlights the WSE's ability to overcome timescale limitations by completing simulation steps in microseconds, enabling the study of phenomena that emerge at long timescales (100+ microseconds). The post elaborates on the computational challenges of parallel MD simulations, particularly communication bandwidth, latency, granularity, and multicast communication, and explains how the WSE's architecture, with its high fabric bandwidth, single-cycle core-to-core latency, native 32-bit message granularity, and novel neighborhood multicast pattern, addresses these challenges. The implementation uses a 'neighborhood multicast' pattern where cores simultaneously multicast particle data horizontally and vertically, interleaved by the dataflow router to avoid congestion. Future plans include adding more potentials, pursuing further acceleration, leveraging multi-wafer clusters, and exploring hybrid simulation+AI models.
  8. Cerebras Wafer-Scale Engine Achieves 210x Speedup Over NVIDIA H100 - CerebrasThis post details a dataflow matrix-free finite volume solver for the Cerebras Wafer-Scale Engine (WSE) that achieves a 210x speedup over NVIDIA H100 GPUs for geological carbon capture and storage (CCS) simulations. The solver leverages the WSE's dataflow architecture, eliminating memory latency and bandwidth bottlenecks. Key findings include a 210x performance boost over H100, up to 1.217 PFlops on a single CS-2 node, and strong scalability. Algorithmic enhancements such as asynchronous communications, vectorization, and memory optimization were crucial. The research highlights the WSE's efficiency in managing intricate data communication and suggests further exploration for HPC tasks with arbitrary mesh topologies.
  9. New Tool Generates Stencil Codes Two Orders of Magnitude Faster on Cerebras WSE Than on GPUsThis post introduces StencilPy, a portable, high-performance code generator for stencil computations, and details its performance on the Cerebras CS-2 system. It highlights that StencilPy reduces the code required for stencil computations on the Cerebras CS-2 by a factor of 7 and achieves significant speedups compared to GPUs (NVIDIA H100) and CPUs (AMD MI210, AMD Genoa). The post elaborates on the Cerebras Wafer-Scale Engine (WSE) architecture, emphasizing its massive parallelism, on-chip memory, and high-bandwidth interconnect as key advantages for stencil computations. It also discusses the Cerebras Software Language (CSL) as the underlying technology enabling these optimizations.
  10. US DOE Achieves 88x Performance Speedup with Cerebras CS-2 Over H100 in Materials ModelingThis post details the implementation and performance results of the Ising model on the Cerebras CS-2 system using the WFA toolset. It highlights an 88x productivity gain over NVIDIA H100 for materials modeling, achieved through efficient checkerboard decomposition and domain folding on the WSE's wafer-scale architecture. The implementation leverages the WSE's low-latency fabric and high memory bandwidth to outperform traditional GPU architectures on this computationally intensive, yet data-light, HPC problem.