Cerebras makes the worlds largest AI chip with 2.6 trillion transistors and nearly 1 million cores

Cerebras Systems has unveiled its largest AI chip based on the 7nm process, called Wafer Scale Engine 2. Following the first WSE, WSE2 is a single monolithic chip with a number of new features including the highest number of transistors and cores ever.

The WSE2 has two parts: the chip itself and the CS-2 system. The WSE2 is the largest AI chip ever produced. Based on Taiwan Semiconductor Manufacturing Company (TSMC)'s 7nm process, the WSE2 will power the company's AI computer, the Cerebras CS-2, which doubles the performance of Cerebras' first-generation CS1.

The WSE2 measures 46,225 mm2, which is 56 times larger than the largest GPU ever made, the NVIDIA Ampere A100 . The WSE2 chip has 850,000 cores, 123 times more than the Ampere A100. The WSE2 has 40GB of memory, compared to 40MB for the Ampere A100. The memory bandwidth is 20 PB/s.

As for the cores on WSE2, they are AI-optimized SLAC sparse matrices and are tuned for AI processing. Compared to WSE1, WSE2 offers 2.16x increased transistor density and has more than double the core count.

“TSMC has long collaborated with industry leaders to create cutting-edge processors with the best performance. We are pleased with the results of our collaboration with Cerebras Systems in manufacturing the Cerebras WSE2 based on our groundbreaking 7nm process. This marks a significant milestone in wafer scale after introducing the Cerebras 16nm WSE less than two years ago,” said Sajiv Dalal, SVP of Business Management, TSMC North America.

The CS2 is 66.04 cm tall and features a 15U with 12 100GbE lanes, a custom water cooling solution via a closed-loop setup layer, and 12 standard PSUs in a 9+3 redundancy configuration delivering up to 23,000W of power to the CS2 system.

There is no price information for CS2 AND WSE2 yet but WSE1 is priced around $2 million.