Cerebras Systems, an AI start-up that focuses on deep learning, has developed an AI chip with 1.2 trillion transistors. By doing so, the company claims to have built the largest processor in the world to date.
The Wafer Scale Engine Cerebras unveiled today, in short WSE, measures 21.6 by 21.6 centimetres and has no less than 1.2 trillion transistors. The WSE’s circuits are organised into 400,000 processing cores specially optimised for AI. The chip has 18 gigabytes of high speed memory on board.
Cerebras comes with the chip as part of a data centre device with its own water cooling system. This is intended to counteract the great heat that this chip develops. Cerebras director Andrew Feldman told Fortune that the chip is 150 times more powerful than a server with multiple Nvidia graphics cards. In fact, the startup chip consumes only 2 to 3 percent of the space and energy that a similar traditional server farm based on Nvidia would use.
The efficiency is due to the size of the WSE. As a result, all compute is executed on the same chip, instead of on multiple chips. A bottleneck caused by a slow connection is therefore a thing of the past.
According to SiliconAngle, another important challenge lies in creating a sample chip like this during the production process. It is very difficult to get all 1.2 million transistors out of production without defects. However, faulty transistors are bypassed by Cerebras. This is done by building an internal network in the chip that can avoid faulty parts.
The design “takes the fundamental properties of the cores, memory and connections and brings them to their extremes”, wrote Andy Hock, director of product management at Cerebras, in a blog post. “By connecting everything on the chip, communication is many thousands of times faster than is possible with off-chip methods such as InfiniBand.This news article was automatically translated from Dutch to give Techzine.eu a head start. All news articles after September 1, 2019 are written in native English and NOT translated. All our background stories are written in native English as well. For more information read our launch article.