Cerebras CS-3 Rack
The Cerebras CS-3 Rack packages two CS-3 systems into a single 48 RU cabinet measuring 600 mm by 1200 mm, drawing 54 kW in total: 50.4 kW carried by water and 3.6 kW by air. Networking is eight QSFP-DD transceivers at 4x100G, with a 1 GbE management port, a console port and a PoE flow controller interface. Power is 2+1 redundant at 240V/415Y, three phase, 60 A on IEC 60309 plugs, and water couples through 1.5 inch sanitary fittings with two pairs of supply and return. Between them the two engines hold 88 GB of on-chip SRAM and no DRAM. Cerebras publishes 125 petaFLOPS per engine but footnotes that figure as sparse, and the sparsity is unstructured, so no dense figure is derived.
We will point you at suppliers who have it. Free, and no signup.
System Details
GPU Configuration
System Specifications
Powered by Cerebras WSE-3
This system utilizes 2 × 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 Rack technical specifications
Cerebras CS-3 datasheet: system, rack and inference cluster specifications
Cerebras • Latest version
Typical Use Cases
The Cerebras CS-3 Rack runs 2× WSE-3 GPUs over 8x QSFP-DD transceiver, 4x100G; 214 Pb/s on-wafer fabric per engine.