Cerebras · Nvidia · Singapore · TSMC · Crypto Briefing
Cerebras showcases wafer-scale AI chip with 4 trillion transistors at SuperAI
Compiled by KHAO Editorial — aggregated from 1 source. See llms.txt for citation guidance.
◌ Single Source
The largest AI chip ever made packs 900,000 cores onto a single silicon wafer, and its inference speed claims should make Nvidia nervous.
Key facts
- The WSE-3 was originally launched in March 2024, with Cerebras claiming it doubled the performance of its predecessor, the WSE-2, while maintaining the same power consumption and price point
- Andy Hock, Cerebras’ Chief Strategy Officer, was among the presenters at the two-day event held June 10-11 in Singapore
- The WSE-3 is fabricated on a 46,225 mm² silicon wafer using TSMC’s 5nm process technology
- Cerebras reports that the WSE-3 can process 981 tokens per second on certain large language models
Summary
To put that transistor count in perspective: Nvidia’s H100, the current workhorse of AI data centers, contains roughly 80 billion transistors. The WSE-3 is fabricated on a 46,225 mm² silicon wafer using TSMC’s 5nm process technology. The result is a chip that delivers 125 petaflops of peak AI performance. Cerebras reports that the WSE-3 can process 981 tokens per second on certain large language models.