Cerebras Systems, which is a Los Altos, California-based startup that has been working in stealth mode since 2016, has just released the largest computer chip ever in the world. How big? It is 8.5 inches on each side and features 1.2 trillion transistors inside. In comparison, a conventional computer ship only millions or at times billions of transistors.
As per the CEO and co-founder of Cerebras Systems, Andrew Feldman, the rate at which research and development in the field of artificial intelligence are being carried out is bound to rise exponentially. Feldman says, ‘The chip is a similar solution to AI as a minivan is a solution to having three kids. What we saw were characteristics of the AI world. And what the software wanted the hardware to do was very difficult.’
Speaking in terms of history, there have been quite a few hurdles in the way of producing huge computer chips. However, Feldman with his team of more than 200 is not only designing and manufacturing but also selling the world’s largest computer chips. Feldman while recalling that famous scientists attempted this back in the 1980s, said, ‘Some of the most famous engineering folks tried, and they failed spectacularly.’ However, he says, ‘We realized we had an opportunity to solve the problem very differently than [they] had tried and created a real opportunity for us.’
So, how is Cerebras Systems doing it? All of the chip design boils down to the photolithography. The equipment manufactures chips and prints them out using circular wafers of pure silicon. This involves using a grid of as many chips as is possible onto the wafer and then slicing the wafer up for the sake of creating separate chips. This is how transistors are etched onto the surface of the silicon wafer. Cerebras Systems, however, uses the entire wafer, for its chips instead of slicing and dicing chips out form silicon.
The equipment that is used prints chips out like a wafer or like a big piece of cookie dough. The chips are then punched out. However, similar to cookie dough where the leftover dough in between the cut-outs that can be mushed for making three or four more cookies; Feldman realized that it could be used with chips as well. He says, ‘We decided, holy cow, we can use the entire piece of dough.’ The team would work to find a way to make use of the maximum possible wafer. It took Cerebras Systems about 3.5 years for perfecting the system thanks to the $112 million in venture capital funding.
Feldman says, ‘The big limitation in artificial intelligence is that it takes weeks or months to test a single idea. What if instead, in that month, you could test 1,000 ideas?’ According to Cerebras Systems, its chip comes with 400,000 small AI optimized cores that enable a chip to house a complete neural network instead of having the data spread out over different GPUs. This helps the company to avoid a giant mess where a bunch of connecting cords is being used. In the end, the system takes less space and less power.
Feldman also said, ‘It’s the complexity of data, plus the characteristics of that complexity. Sometimes it might be hard to move a big boulder, other times it may be hard to move a huge number of little pebbles. They have different characteristics. AI requires a new type of truck to move it, and that’s the new type of computing.’
In the end, we totally agree with Feldman, who said, ‘In a new world where everything has to work together … the faster the communication, the faster the innovation.’