A startup called Cerebras has designed, in collaboration with the National Energy Technology Laboratory of the US Department of Energy, what it calls "the most powerful AI computing system in the world." We are talking about the largest computer chip in the world, with 1.2 trillion transistors inside that make up about 381,000 computational cores .
So fast and powerful that it can predict future actions "faster than the laws of physics produce the same result". it's called Cerebras CS-1 , and it's a 54-square-centimeter chip, and has found its place in a computer that's about the size of a refrigerator.
But how powerful is this new chip with a trillion transistors?
To have a simple comparison on the size of this chip, just remember that Intel's first 4004 processor in 1971 had 2,300 transistors and the 80GB Nvidia A100 chip, announced a few days ago, contains 54 billion transistors.
In terms of effective computing power, when compared to the Joule supercomputer , the computer with the Cerebras CS-1 on board is 200 times faster . In a simulation of the combustion processes of an engine, its performance at the moment cannot be in any way matched by the already existing supercomputers, regardless of the number of processors and GPUs they are equipped with.
“This work opens the door to important breakthroughs in the performance of scientific computers. The CS-1 is the first system ever to have sufficient performance to simulate over a million fluid cells faster than real time. And this means that when CS-1 is used to simulate an electrical system based on data on its current operating conditions, it can tell what will happen in the future faster than the laws of physics can produce ” .
To give us an idea, Joule uses Intel Xeon chips with 20 cores each for a total of 16,000 cores and requires 6 ms to complete the calculations of the aforementioned combustion processes, while Cerebras CS-1 only took 28 us .
During the chip manufacturing process, a single huge chip is made from the silicon wafer. Every single piece of the chip, dubbed the core, is interconnected in a sophisticated way to other cores. The interconnects are designed to keep all cores running at high speed so that the transistors can work together in unison.
Finally, Cerebras CEO Andrew Feldman added:
“We can solve this and other problems in a time frame that no number of GPUs or CPUs can currently reach. This means that the CS-1 for this job is the fastest machine ever built and is faster than any combination of clusters of different processors, which is the technique currently used for the construction of supercomputers ”.