Cerebras Systems

Cerebras Systems
Name	Cerebras Systems
Type	Private
Founded	2016
Founders	Andrew Feldman; Jean-Philippe Fricker; Michael James; Sean Lie; Sean Naffziger
Headquarters	Sunnyvale, California
Industry	Semiconductor; Artificial intelligence hardware; Deep learning
Products	Wafer-Scale Engine (WSE); Cerebras CS-1; Cerebras CS-2; Cerebras Wafer-Scale Cluster; software stack

Cerebras Systems Cerebras Systems is a California-based company that develops large-scale semiconductor systems and accelerators for artificial intelligence workloads. Founded by a team of engineers and executives with backgrounds at Intel Corporation, NVIDIA Corporation, Microsoft Corporation, Qualcomm, and IBM, the company aims to deliver high-performance hardware and software for deep learning, machine learning, and high performance computing. Cerebras has drawn attention for its wafer-scale chip approach and for deployments in research centers, national laboratories, and commercial cloud providers.

History

Cerebras was founded in 2016 by Andrew Feldman, Jean-Philippe Fricker, Michael James, Sean Lie, and Sean Naffziger with early connections to Stanford University, University of California, Berkeley, and Massachusetts Institute of Technology. Early funding rounds included investors such as Benchmark, Foundation Capital, Sutter Hill Ventures, Eclipse Ventures, and Valor Equity Partners. The company publicly unveiled its Wafer-Scale Engine concept alongside collaborations with research institutions like Lawrence Livermore National Laboratory, Argonne National Laboratory, CERN, and Los Alamos National Laboratory. Cerebras expanded partnerships with cloud and enterprise providers including Google Cloud Platform, Amazon Web Services, Microsoft Azure, Oracle Corporation, and Hewlett Packard Enterprise for deployment and testing. Over subsequent years Cerebras announced product launches, architectural revisions, and funding events while engaging with academic labs at Carnegie Mellon University, University of Toronto, University of Cambridge, ETH Zurich, and Imperial College London.

Technology and Products

Cerebras developed the Wafer-Scale Engine (WSE), an unusually large integrated circuit tailored for tensor computations used in models from OpenAI and research from DeepMind Technologies and Google DeepMind. Initial commercial systems included the CS-1, followed by the CS-2 featuring larger on-chip memory and process nodes similar to offerings from TSMC. The product stack integrates with frameworks and toolchains from PyTorch, TensorFlow, Hugging Face, Ray (software), and orchestration tools such as Kubernetes and Docker. Software partnerships and tool integrations have involved NVIDIA CUDA ecosystems, Intel oneAPI, and workflow systems from Slurm Workload Manager and Apache Spark. Cerebras systems are marketed to institutions using workloads from projects by OpenAI, Meta Platforms, Inc. research groups, pharmaceutical firms like Pfizer and Roche, automotive developers such as Tesla, Inc. and Waymo, and climate modeling efforts at NOAA and NASA research centers.

Architecture and Design

The wafer-scale approach departs from traditional chiplet and multi-GPU designs used by NVIDIA Corporation and AMD (company). The WSE combines a very high count of compute cores, on-chip SRAM, and dense interconnect fabric in a single monolithic silicon wafer, addressing issues similar to those explored by Intel Corporation with large die strategies and by historical wafer-scale research at IBM Research. The design emphasizes low-latency mesh networks reminiscent of topologies used in systems from Cray Inc. and Fujitsu and implements fault tolerance strategies analogous to those in supercomputers at Lawrence Berkeley National Laboratory. Cooling and packaging integrate elements familiar to engineers from ASML Holding, Applied Materials, and Lam Research Corporation manufacturing pipelines. The architecture supports sparse and dense tensor kernels used in transformer models from Google Research and recurrent models from Facebook AI Research.

Performance and Benchmarks

Cerebras publications and third-party evaluations report performance claims comparing WSE-based systems to multi-GPU clusters built from NVIDIA A100 and NVIDIA H100 accelerators across benchmarks from MLPerf, custom transformer workloads, and large language model training from groups like EleutherAI. Performance metrics often emphasize throughput, latency, memory bandwidth, and energy efficiency, with comparisons against clusters using interconnects such as NVIDIA NVLink and InfiniBand from Mellanox Technologies. Benchmarking efforts involved collaboration with supercomputing facilities using job schedulers like Slurm and message passing libraries such as MPI (Message Passing Interface). Independent analyses from research teams at Stanford University, MIT CSAIL, and University of California, Berkeley labs examined scaling behavior, convergence rates, and cost-performance tradeoffs versus alternatives from Google TPU deployments and custom ASICs from Graphcore.

Applications and Use Cases

Cerebras systems are applied in domains including natural language processing (NLP) for models used by OpenAI collaborators, computer vision projects from Meta Platforms, Inc. research labs, genomics pipelines at institutions like Broad Institute and Sanger Institute, drug discovery programs at Novartis and GlaxoSmithKline, and climate modeling at NOAA and NASA. Additional use cases span computational chemistry efforts tied to Schrödinger (company), materials science collaborations with Sandia National Laboratories, and physics simulations relevant to CERN experiments. Enterprises in finance such as Goldman Sachs and JPMorgan Chase have explored model training for quantitative research, while automotive firms including Ford Motor Company and General Motors investigate perception model acceleration. Academic collaborations cover work at Columbia University, Yale University, Princeton University, University of Oxford, and University of Chicago.

Business and Partnerships

Cerebras has pursued strategic relationships with semiconductor manufacturers like TSMC and packaging partners associated with ASE Technology Holding and Amkor Technology. Commercial and research partnerships include cloud providers Google Cloud Platform, Amazon Web Services, Microsoft Azure, and system integrators such as Hewlett Packard Enterprise and Dell Technologies. Collaboration agreements and pilot programs have been announced with pharmaceutical corporations Pfizer and Roche, national labs such as Argonne National Laboratory and Lawrence Livermore National Laboratory, and academic consortia including PARTNER INSTITUTION NAMES for multi-institution projects. Investors and strategic backers include venture capital firms Benchmark (venture capital firm), Foundation Capital, and corporate investors with histories tied to Intel Capital and Samsung Ventures.

Category:Computer companies Category:Semiconductor companies