One of the highlights of Hot Chips 2019 was the presentation of the Cerebras Wafer Scale Engine – an AI processor chip that was as huge as a wafer, containing 1.2 trillion transistors and set at over 46225 sq. millimetres of silicon. This was enabled by means of breakthrough methods in cross-reticle patterning, however with the extent of redundancy constructed into the design, ensured a yield of 100%, each time. The first WSE system, the CS-1, was put out on show at Supercomputing 2019, the place we acquired an opportunity to chew into the design with Andrew Feldman, the founder and CEO of Cerebras.
Unfortunately I by no means acquired round to writing up my discussions with Andrew, nevertheless what we did study on the time is that the CS1 is a totally built-in 15U chassis that requires 20 kW of energy to push to the chip by means of 12x four kW energy provides (some redundancy built-in). The chip is mounted vertically for the sake of ease of entry, which is sort of weird within the fashionable world of computing. Most of the chassis was customized constructed for the CS-1, together with the tooling and a good quantity of economic 3D printing. Andrew additionally stated on the time that whereas there was no minimal order amount for the CS-1, nevertheless each would price ‘a few million’.
Today’s announcement from the Pittsburgh Supercomputing Center (PSC) helps spherical that quantity right down to maybe ~$2 million. Through a $5 million grant from the National Science Foundation (NSF) to the PSC, a brand new AI supercomputer shall be constructed, referred to as Neocortex. At the center of Neocortex shall be {hardware} inbuilt partnership with Cerebras and Hewlett Packard Enterprise.
Specifically, there shall be two CS-1 machines on the coronary heart of Neocortex. The CS-1 helps asynchronous fashions by means of TensorFlow and pyTorch, with the software program platform capable of optimize the dimensions of the workloads for the out there space on the CS-1 Wafer Scale Engine.
Each entrance panel half is machined from a single piece of aluminium
The pair of CS-1 machines shall be coupled with an ‘extreme’ shared-memory HPE Superdome Flex server, which accommodates 32 Xeon CPUs, 24 TB of DDR4, 205 TB of storage, and 1.2 Tbps of community interfacing. Neocortex is anticipated for use to allow AI researchers to coach their fashions, protecting areas reminiscent of healthcare, illness, energy era, transportation, in addition to urgent problems with the day.
The machine shall be put in in late 2020. PSC has acknowledged that entry to Neocortex shall be out there to researchers within the US without charge.
When we spoke to Cerebras final yr, the corporate acknowledged that they already had orders within the ‘strong double digits’. When pressed, I managed to get that from ’12 to a number of dozen’. Quite a few machines had been ordered for the Argonne National Laboratories on the time, and I believe others are actually investing.
Interestingly sufficient, at Hot Chips 2020 this yr, the corporate is about to reveal its second era Wafer Scale Engine. At a guess, I might counsel that that is barely additional away to commercialization than WSE1 was when it was introduced, however the firm appears to have had substantial curiosity of their expertise.
