Cerebras Unveils Trillion-Transistor Wafer-Scale Chip After Solving Massive Design Hurdles
Cerebras emerges from stealth to launch the Wafer Scale Engine, a massive chip with 1.2 trillion transistors designed to accelerate deep learning. The company overcomes significant manufacturing and design challenges by using an entire silicon wafer as a single processor.
Cerebras emerges from stealth mode to unveil the Wafer Scale Engine, a massive computer chip that boasts 1.2 trillion transistors, 400,000 processing cores, and 18 gigabytes of on-chip memory. The company introduces this record-breaking hardware at the Hot Chips conference at Stanford University, marking the culmination of three years of quiet development backed by $112 million in venture capital.
Instead of etching dozens of small individual chips onto a standard silicon wafer, Cerebras uses the entire wafer as one gigantic, continuous processor. This unconventional approach allows the processing cores to communicate directly with one another, which vastly speeds up the critical feedback loops required for deep learning algorithms.
Abandoning traditional chip geometry introduces massive manufacturing and design challenges, particularly regarding lithography errors and yield management. Because typical chipmakers rely on small chip sizes to minimize defects and maximize profits, Cerebras faces unique hurdles in ensuring a fully functional, error-free wafer-scale design that completely rewrites standard silicon manufacturing rules.