Cerebras WSE-3

Wafer Scale Engine Architecture

Active

Launched March 2024

Core Specifications

VendorCerebras

ArchitectureWafer Scale Engine

Form Factor—

VRAM44 GB

Memory Bandwidth—

TDP23000 W

Precision	TFLOPs

Configuration	Precision	Performance	Source
LLaMA 70B, ultra-low latency	—	16,000 tokens_per_second	View

Configuration	Precision	Performance	Source
GPT-3 175B, single CS-3 system	—	120 hours_to_train	View
GPT-3 175B, weight streaming architecture	—	25,000 tokens_per_second	View

View other Cerebras GPUs or compare across vendors