A drawer of Brains - Intel updates neural network system

williamcll · March 19, 2020

Still on 14nm though.

Quote

Intel today announced it has reached a new milestone for its neuromorphic chip called Loihi. It comes in the form of a system targeting research that's packed with 786 Loihi chips for a total of over 100 million neurons. The system goes by Pohoiki Springs and is available to Intel’s research community members through the cloud and programmed via an SDK. One 14nm Loihi chip comprises a 60mm-squared chip and 2 billion transistors. It can process certain specific workloads up to 1,000 times faster and 10,000 times more efficient than standard processors, so a system with 786 of them might deliver unprecedented performance in some workloads. according to Intel. The new Pohoiki Springs system comprises 786 Loihi chips, which means it packs a total of 1.5 trillion transistors on almost 50,000mm-squared of silicon. This gives it a total of 100 million transistors. Despite this enormous scale, Intel said it operates at under 500W, due to it using an asynchronous architecture: there is no global clock on which all transistors switch simultaneously. As Intel’s largest neuromorphic system to date, Pohoiki Springs is mounted in a data center rack. The chassis is the size of five standard servers (5U) amd consists of eight "rows" that each contain three 32-chip Intel Nahuku boards.

Pohoiki Springs has an amount of neurons comparable to small mammals and should be capable of supporting more advanced neuromorphic workloads, although Intel cautions that it is still in the research phase. Intel provided three categories where Loihi excels: constraint satisfaction, searching graphs and patterns, such as finding the shortest paths, and optimization problems. Intel set out use Loihi in more powerful designs by scaling the numbers of chips. Last year, it released the 64-chip Pohoiki Beach system with 8 million neurons. Using Loihi, Intel also started a neuromorphic research community (INRS) to research all aspects of the neuromorphic computing, towards commercialization. To that end, last year the first Global 500 companies joined this community.