The American company Cerebras Systems introduced the world's largest processor, called Wafer Scale Engine. At 21.5 x 21.5 cm, this chip includes 1.2 trillion transistors and a whopping 400,000 cores. In size, it is comparable to the 11-inch iPad Pro.
The assurances Cerebras Systems, the new chip is 56 times greater than the largest GPU. Its total area is 46,225 square millimeters. On this huge canvas housed 1.2 trillion transistors and 400 thousand cores, optimized for the tasks of artificial intelligence. The chip is made by 16-nm TSMC process technology.
The Wafer Scale Engine includes 18 GB of ultra-fast SRAM. Its capacity is 9 PB/s.
Wafer Scale Engine is designed for use in data centers. The chip is part of a device with its own water cooling system for the normal operation of thousands of cores. According to the Executive Director Cerebras Systems Andrew Feldman, the processor is able to deliver 150 times the compute capacity compared to a server that has multiple graphics cards NVIDIA. What's more, WSE consumes 2% to 3% of the space and power required for a server based on NVIDIA chips with the same processing power.
In traditional servers with several separate graphics processors focused on performing artificial intelligence tasks, the processed data is constantly transmitted between different chips through relatively slow network channels, which significantly reduces the efficiency of such solutions. In the case of development Cerebras Systems, this is not a problem since all computations are performed on the same Board.
Prior to this, the creation of such processors was impossible, since if at least some of the 1.2 trillion transistors worked incorrectly, it would render the entire chip unusable. Cerebras Systems managed to get around this problem, by creating a network in Wafer Scale Engine in such a way that the data can "bypass" the faulty part.
The company intends to start deliveries of the data center on the basis of WSE in September this year. Some organizations have already received early samples.