New AI supercomputing processor by Cerebras breaks 2.5 trillion transistor mark
Written by Finbarr Toesland Wed 21 Apr 2021
The largest current graphics processor unit (GPU) has 2.55 trillion fewer transistors than new Cerebras processor
California startup Cerebras Systems has announced they have created the largest AI processor ever made, the Wafer Scale Engine 2 (WSE-2).
The AI firm made waves in the industry two years ago when their original WSE chip broke records by featuring 1.2 trillion transistors and 400,000 AI optimised cores.
While the second-generation WSE is the same size as its predecessor, it offers more than twice the transistors and more than twice the cores, compared to the original.
When Cerebras founder Andrew Feldman started the company he decided to move away from traditional chipmakers, who typically take a wafer slice from a 12-inch diameter block of silicon to process and then slice the wafer again into many individual chips, ready to be placed in electronics.
Instead, Cerebras uses a wafer to create one extremely powerful interconnected chip, which resulted in the WSE-2’s 2.6 trillion transistors and 850,000 AI optimised cores.
By using an improved 7-nanometer process, where the width between circuits is seven billionths of a meter, Cerebras is able to fit far more transistors on the 12-inch wafer. The WSE-2 is set to power the Cerebras CS-2, the fastest AI computer in the industry.
To put the immense power of WSE-2 in perspective, the current largest graphics processor unit (GPU) has a mere 54 billion transistors, some 2.55 trillion fewer transistors than the WSE-2. Thanks to advancements made over the first WSE, the second-generation version has 123x more cores and 1,000x more high performance on-chip high memory when compared to GPU competitors.
As this chip has been custom-built for AI work, it represents a major step forward for AI.
“In AI compute, big chips are king, as they process information more quickly, producing answers in less time – and time is the enemy of progress in AI. The WSE-2 solves this major challenge as the industry’s fastest and largest AI processor ever made,” said Dhiraj Mallik, Vice President Hardware Engineering, Cerebras Systems, in a statement.