Develop with the Cerebras SDK
Build custom kernels for the Wafer-Scale Engine — everything from your first CSL program to running at scale on a full Wafer-Scale Cluster.
Start Here
Install the Cerebras SDK
Step-by-step setup of the Cerebras SDK Singularity container.
Wafer-Scale Engine Architecture
A conceptual, “mental model” view of computing with Cerebras.
Host Runtime & Tensor Streaming
The SdkRuntime host API and memcpy infrastructure for data transfer.
Tutorials
Walk through GEMV tutorials, from basic CSL syntax to multi-PE programs.
SDK Code Examples
Tutorials and benchmarks from the Cerebras SDK examples repository.
Published Results
Highlights from research built with the Cerebras SDK. Each card links to the paper.
10–20× faster LLM inference
WaferLLM runs full LLM inference 10–20× faster than A100 GPU clusters running SGLang and vLLM.
University of Edinburgh, Microsoft Research · OSDI ‘25
University of Edinburgh, Microsoft Research · OSDI ‘25
92.58 PB/s memory bandwidth
Tile low-rank matrix-vector multiplication for seismic processing, sustained across 48 CS-2 systems.
KAUST, Cerebras · SC23 Gordon Bell Prize finalist
KAUST, Cerebras · SC23 Gordon Bell Prize finalist
130× faster than an A100
A Monte Carlo particle transport kernel, against a highly optimized CUDA version on an NVIDIA A100.
Argonne National Laboratory · Computer Physics Communications
Argonne National Laboratory · Computer Physics Communications
About 14× faster than 128 A100s
An MLIR pipeline compiles unmodified stencil codes to CSL; one WSE-3 outruns 128 A100 GPUs.
University of Edinburgh (EPCC) and partners · ASPLOS ‘26
University of Edinburgh (EPCC) and partners · ASPLOS ‘26
Up to 25× faster multigrid
Geometric multigrid mapped onto WSE-3, compared with a single NVIDIA GH200 node.
LBNL, University of Utah · SC26
LBNL, University of Utah · SC26
1.2 PFLOPS on one device
A matrix-free finite-volume solver for carbon storage simulation, two orders of magnitude faster than a GPU reference.
TotalEnergies, Rice University · SC24
TotalEnergies, Rice University · SC24
Development Guides
CSL Language Reference
Complete reference covering syntax, builtins, types, DSDs, and libraries.
Running SDK on a Wafer-Scale Cluster
Use the SDK in appliance mode on a Cerebras Wafer-Scale Cluster.
Host API Reference
SdkRuntime API
Python host runtime for loading programs, launching functions, and transferring data.
SdkLayout API
Python API for programmatic layout of multi-region SDK programs.
SDK Appliance API
Compile and run SDK programs on a Wafer-Scale Cluster.