Semiconductor Startup Cerebras Systems Shows Off the World’s Biggest Processor for AI
Startup Cerebras Systems on Monday unveiled a processor measuring roughly 8 inches by 8 inches - at least 50 times larger than similar chips available today.
This very big chip is designed to meet the needs of the AI community which is repurposing graphics processors to meet new compute needs. You see, big chips process information more quickly and produce answers in less time. This is incredibly important for researchers crunching enormous amounts of data, it enables them to train models faster.
The Cerebras team have built the Cerebras Wafer Scale Engine (WSE), aka, the largest chip ever built. With a 1,000x performance improvement over what’s currently available, the Cerebras WSE is comprised of more than 1.2 trillion transistors and is 46,225 square millimeters. It also contains 3,000 times more high speed, on-chip memory, and has 10,000 times more memory bandwidth. For comparison, the first chips created in the 1960s featured a few hundred transistors.
On a chip of this size, the startup was able to deliver 400,000 AI-optimized cores. The company's specialized memory architecture ensures each of these cores operates at maximum efficiency. It provides 18 Gigabytes of fast, on-chip memory distributed among the cores in a single-level memory hierarchy, one clock cycle away from each core. All of these AI-optimized cores are connected entirely on silicon by the Swarm fabric in a 2D mesh with 100 Petabits per second of bandwidth. Swarm delivers breakthrough bandwidth and low latency at a fraction of the power draw of traditional techniques used to cluster graphics processing units. Software configures all the cores on the WSE to support the precise communication required for training user-specified models.
Altogether, the WSE takes the fundamental properties of cores, memory, and interconnect to their logical extremes. A vast array of programmable cores provides cluster-scale compute on a single chip. High-speed memory close to each core ensures that cores are always occupied doing calculations. And by connecting everything on-die, communication is many thousands of times faster than what is possible with off-chip technologies like InfiniBand.
Andrew Feldman, the company's founder, co-founded server maker SeaMicro Inc. and sold it to chipmaker Advanced Micro Devices Inc. for more than $300 million in 2012. Cerebras has raised over $100 million from Silicon Valley investors including Benchmark, Andy Bechtolsheim and Sam Altman. Feldman has a team of 174 engineers and Taiwan Semiconductor Manufacturing Co.is manufacturing the massive Cerebras processor.
Cerebras won’t sell the chips because it’s so difficult to connect and cool such a huge piece of silicon. Instead, the product will offered as part of a new server that will be installed in data centers. The company said it has test systems working at several large potential customers and will start shipping the machines commercially in October.