Skip to content

WSE-3

A 5 nm wafer-scale engine for training large AI models

Visit official site

What WSE-3 is

The chip is described as a third-generation 5 nm wafer-scale engine from Cerebras Systems for training very large AI models.

What you can do with it

  • Deliver 125 petaflops of peak AI performance
  • Train models with up to 24 trillion parameters
  • Scale a system out to as many as 2048 nodes

Application scenarios

  • Research institutes train medical diagnostic models
  • Firms use a CS-3 system for drug discovery
  • Hyperscale data centres build AI supercomputers

Specifications

  • 4 trillion transistors are stated
  • 900,000 AI-optimised cores are stated
  • Performance is twice WSE-2 at the same power draw

Pros & cons

✓ What we like

  • Supports AI models up to 24 trillion parameters
  • Scales to as many as 2048 nodes

! What to watch out for

  • One logical memory space is described for a 24-trillion-parameter model
  • The chip is aimed at training very large AI models

FAQ

What process is used?

The material states a 5 nm process with 4 trillion transistors.

What peak performance is listed?

125 petaflops of peak AI performance is stated.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review