ARK Invest
Andrew Feldman, co-founder and CEO of Cerebras, joins us this week to discuss the Wafer Scale Engine, or WSE, an AI chip that is 50 times larger than the largest chips produced by Nvidia and Intel. This radical design has raised a lot of eyebrows and it is already being heralded as the biggest breakthrough in semi-conductor technology in decades.
Andrew helps us unpack why AI work needs chips this large, how Cerebras was able to leapfrog industry incumbents, and what why the Wafer Scale Engine is the ideal AI training accelerator. Join us on this podcast as we talk with Andrew about some of the biggest hurdles that his team encountered on this market shifting journey.
Key Points From This Episode
-A high-level description of the chip Cerebras has created and what makes it different.
-Why this size and level of the chip has eluded companies far larger than Cerebras.
-The path to the Wafer Scale Engine, starting with the founding of Cerebras.
-Understanding the system architecture of the Wafer Scale Engine.
-The relationship between Cerebras and TSMC, their fabrication partner.
-Choices that were ultimately made around size and particularly memory.
-Getting past the challenge of bottlenecks and bringing data onto a chip this size.
-Reasons startups like Cerebras are able to do things more cheaply than larger competitors.
-Examples of the power that this Wafer Scale Engine offers through its incredible flexibility
-Cerebras’ go-to-market strategy and Andrew’s thoughts on the size of the training market.
Tweetables
“Our guys are not afraid of invention and I think sometimes that’s not the incentive structure at large companies.” — @CerebrasSystems [0:07:01]
“We set out to be extraordinarily ambitious and that was in the system design, that was in the chip design and architecture. That was at every stage of our thinking.” — @CerebrasSystems [0:14:13]
More FYI Podcasts: https://ark-invest.com/research/podcast
Learn more about ARK: https://ark-invest.com/
Disclosure: http://bit.ly/1C5DBVL
Source
Very insightful!
Amazing episode. Andrew's explanations are clear and concise and uplifting. A must hear for anyone in AI. Thanks James for steering the conversation into every avenue of interest
Great episode, James keep your feelings about Nvidia to yourself
Wouldn’t quantum computers be faster than this chip?
The problem is quality. Even if only one transistor in a billion is faulty, there would still be 12,000 faulty transistors on a single chip.
This is why the 1% is rich and the rest are watching cat videos and fortnite.
Truly revolutionary and these chips are already being used at argonne for ai cancer research, Feldman is a genius and should be praised for the medical and various different breakthrough product this will yield.
1 person afraid of the big bad AI.
How much does the system cost?
Made in TSMC
The matter here if cerebras has a commercial ready product to market and overall price!!! But again this system is an experiment or a prototype.
Would be interesting to see what kind of yield he's getting.0