‘Wafer scale engine’ designed to power AI of the future


The Cerebras chip is the world’s largest, 56x bigger than any before and with 1.2 trillion transistors.

The world’s largest computer chip will enable AI models to process 4 billion parameters at once. That compares to about a million per chip typical of processors today. The new chip is made by Cerebras, a California start-up.

  • Cerebras calls their new chip the WSE, which stands for ‘wafer scale engine.’
  • The breakthrough is cutting out the interconnections. Chips often are ganged together to deliver faster outputs in their aggregate, but each connection involves time to trade results. Putting, say, 10 chips together does not offer 10x performance but some factor less.
  • FORBES calls it ‘a supercomputer-on-a-chip.’
  • ZDNET says WSE will enable new kinds of neural networks.
  • EE|Times reports the WSE offers similar performance to an assembly of 1,000 chips by its closest competitor ‘while requiring just 2-3% of its space and power.’
  • BBC NEWS says the company expects processing time for some complex operations can drop ‘from months to minutes.’
  • WIRED gives an example of $350,000 in computing time for a single piece of language software. Dropping computing time can save costs, which can be in the millions when working with large AI models. The price for the WSE and its housing has not been reported.

“When limitations of the hardware are keeping us from exploring fertile areas, that’s very much what we sought to change… Hardware should be a platform on which your thinking can take shape.”

ANDREW FELDMAN in ZDNET

SEE RELATED STORIES

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.