Skip to main content

Cerebras WSE-3 (CS-3)

Overview

Cerebras WSE-3 was launched in March 2024 as the third-generation wafer-scale AI chip. It features 4 trillion transistors (~1.5× WSE-2), 900,000 cores, 44GB on-chip SRAM, and up to 125 PFLOPS FP16 sparse compute. CS-3 addresses the core bottlenecks of large-model training: the memory wall and communication overhead.

Core Specifications

ItemSpecification
ArchitectureWSE-3 (Wafer-Scale Engine 3)
Process NodeTSMC 5nm
Transistor Count4 trillion
Core Count900,000
On-chip SRAM44 GB
FP16 Sparse Compute125 PFLOPS
TDP (system)15 kW
Form FactorCS-3 complete system

WSE-2 vs WSE-3

MetricWSE-2WSE-3Improvement
Process Node7nm5nm
Transistors2.6 trillion4 trillion1.5×
Core Count850,000900,0001.06×
SRAM40 GB44 GB1.1×
FP16 Sparse~62 PFLOPS125 PFLOPS

Vendor Information

ItemDetails
ManufacturerCerebras Systems
Official Websitehttps://www.cerebras.net
Product Pagehttps://www.cerebras.net/product-cs-3/
ReleaseMarch 2024
Cloud Rentalhttps://cerebras.net/ai-cloud/

Key Features

  • CS-3 System: 15kW liquid-cooled complete system
  • Cluster Solution: Cerebras Cluster, multiple CS-3 interconnected
  • Weight Streaming: No memory wall
  • Supports GPT-class model training

Use Cases

  • Giant LLM training (no model parallelism)
  • Research institutions (Genentech, Argonne National Laboratory)
  • Pharmaceutical and materials science