Apr 26, 2021 09:00:00

Cerebras develops large chip 'Wafer Scale Engine 2' equipped with 2.6 trillion transistors

On April 20, 2021, Cerebras Systems, a company that manufactures chips dedicated to machine learning and deep learning, has developed a computer chip 'Wafer Scale Engine 2 (WSE2) ' with 2.6 trillion transistors mounted on a wafer. Announced.

We're serving up a record-breaking chip. It contains 2.6 trillion transistors and covers more than 46,225 square millimeters of silicon, about the size of a dinner plate. Pic.twitter.com/vg272j5E6A

— Cerebras Systems (@CerebrasSystems) April 22, 2021

Cerebras' New Monster AI Chip Adds 1.4 Trillion Transistors --IEEE Spectrum
https://spectrum.ieee.org/tech-talk/semiconductors/processors/cerebras-giant-ai-chip-now-has-a-trillions-more-transistors

WSE2 has 2.6 trillion transistors, 850,000 cores, and 40GB of on-chip memory in an area of 20cm x 22cm. Cerebras Systems was

developing the Wafer Scale Engine (WSE) chip with 1.2 trillion transistors in 2019, but the WSE2 has more than doubled in most performance. The spec comparison table of WSE and WSE2 is as follows.

	WSE 2	WSE
size	46,255 mm 2	46,255 mm 2
Transistor	2.6 trillion	1.2 trillion
core	850,000	400,000
On-chip memory	40GB	18GB
Memory bandwidth	20PB / s	9PB / s
Fabric bandwidth	220Pb / s	100Pb / s
Manufacturing process	7nm	16nm

The biggest change from WSE is the manufacturing process. WSE2 uses TSMC's 7nm process and more than doubles the transistor density, resulting in a 40% increase in speed and a 60% reduction in power consumption.

Cerebras Systems explains that the reason the size of the chip itself has not changed from WSE is that this size is the limit size for mass production. Dhiraj Malik, Vice President of Cerebras Systems, said, 'Large chips are important for the information processing speed of AI computing,' and Cerebras intends to offer the new WSE to various customers.