Get the latest tech news
100x defect tolerance: How we solved the yield problem
Conventional wisdom in semiconductor manufacturing has long held that bigger chips mean worse yields. Yet at Cerebras, we’ve successfully built and commercialized a chip 50x larger than the largest computer chips – and achieved comparable yields.
By examining the interplay between defect rates, core size, and fault tolerance, we’ll show how we achieve wafer scale integration with equal or better yields vs. reticle limited GPUs. Even though larger chips generally run faster, early microprocessors were built to a modest size to maintain acceptable manufacturing yields and profit margins. Unlike previous approaches that required massive redundancy overhead, our architecture achieves high yield with minimal spare capacity through intelligent routing.
Or read this on Hacker News
