CS

Cerebras CS-3

rack
1× WSE-3
1.2 Tb/s system I/O; 214 Pb/s on-wafer fabric
2024
Power
27.0
kW Total
Memory
GB Total

The Cerebras CS-3 is a single 16 RU liquid-cooled system built around one Wafer Scale Engine 3, drawing up to 27 kW. Its memory is 44 GB of on-chip SRAM running at 21 PB/s, with no DRAM of any kind, and system I/O is 1.2 Tb/s. Cerebras rates it at 125 petaFLOPS of FP16, but a footnote on its own datasheet states that FLOPS are shown as sparse, and the sparsity is unstructured rather than 2:4 structured, so no dense equivalent can be derived by halving and none is published. This entry therefore carries no precision figures rather than a sparse number that would be compared against the dense figures used everywhere else on this site. Cooling is water at 100 L/min per system, 20 +/- 2 C, with 50 um filtering and PG25 chemistry.

We will point you at suppliers who have it. Free, and no signup.

System Details

GPU Configuration

GPU Model: Cerebras WSE-3
GPU Count: 1 GPUs
Architecture: Wafer Scale Engine 3
Interconnect: 1.2 Tb/s system I/O; 214 Pb/s on-wafer fabric

System Specifications

Form Factor: rack
Total Power: 27.0 kW
Total Memory:
Memory Bandwidth: 21000000 GB/s

Powered by Cerebras WSE-3

This system utilizes 1 × Cerebras WSE-3 GPUs, each delivering exceptional performance for AI and HPC workloads.

Per GPU TDP

Per GPU Memory

Process Node

TSMC 5nm

Architecture

Wafer Scale Engine 3

Documentation & Resources

Official Datasheet

Cerebras CS-3 technical specifications

Download PDF

Cerebras CS-3 datasheet: system, rack and inference cluster specifications

Cerebras • Latest version

View Document ↗

Typical Use Cases

AI/ML Training
High-Performance Computing
Data Analytics

The Cerebras CS-3 runs 1× WSE-3 GPUs over 1.2 Tb/s system I/O; 214 Pb/s on-wafer fabric.