A Silicon Valley startup on Monday debuted what it called the world’s largest computer chip and outlined its ambitions to help deep learning capabilities reach their full potential.
Cerebras Systems, established by the co-founders of microserver developer SeaMicro, unveiled the Cerebras Wafer Scale engine in a blog post. The Wafer Scale, at nearly 72 square inches, is more than 56 times larger than the world’s biggest graphics processing unit.
Andy Hock, the company’s product director, wrote that GPUs and other conventional machines were originally designed for different tasks than the massive and growing demands of artificial intelligence. New models and hypotheses can take weeks or months to test under current systems.
“Time and again, our industry has shown that special applications require specialized accelerators,” Hock wrote.
“In the same vein, we need a new processor — a new engine — for deep learning compute.”
The Wafer Scale, he wrote, includes more than 1.2 trillion transistors and 18 gigabytes of memory distributed amongst 400,000 cores and connected with 100 petabits per second of bandwidth. Its unprecedented size enables 3,000 times the on-chip memory and more than 10,000 times the memory bandwidth of the largest GPU.
The company argued that building the system from scratch to be optimized for deep learning allows testing times to be slashed from days to minutes.
“Every element ... has been made to enable deep learning research at unprecedented speeds and scale,” Hock wrote. “And this is just the beginning.”