Wafer-Scale Integration for AI – The Holy Grail?

This paper presents motivation, challenges, and solutions to productize a wafer-scale integration (WSI) die for artificial intelligence (AI) applications. In a similar way GPUs are better suited than CPUs for AI applications, the Cerebras (CBS) WSI is even better suited to solve large problems. The solutions presented include large on-chip memory, tightly coupled vertical power delivery, a mesochronous clocking scheme, and a thorough in-system test and burn-in methodology.