Why Cerebras Refuses to Cut Its Wafer Into Chips @FARZAD-FM
Why Cerebras Refuses to Cut Its Wafer Into Chips  @FARZAD-FM
Uploaded November 2026 | Updated November 2026, 4 days ago
Why does Cerebras keep a whole wafer together instead of cutting it into chips?

Normally a silicon wafer is cut into separate chips, so a defect only ruins one small chip. Cerebras keeps a huge square of the wafer as one chip with four trillion transistors, and makes it work with tiny cores and a mesh network that routes around dead ones. Each core is about one percent the size of an Nvidia H100 SM, the chip has about 970,000 physical cores with 900,000 active, and spare cores plus redundant paths let software see a uniform grid of working cores. In Cerebras's own simplified calculation, GPU chips cut from a wafer lose about 361 square millimeters to defects, while the wafer chip loses about 2.2 square millimeters. That calculation leaves out defects in shared circuitry and clustered defects.

Sources:
cerebras.ai/blog/100x-defect-tolerance-how-cerebras-solved-the-yield-problem
cerebras.ai/press-release/cerebras-announces-third-generation-wafer-scale-engine
hc2024.hotchips.org/assets/program/conference/day2/72_HC2024.Cerebras.Sean.v03.final.pdf
spectrum.ieee.org/cerebras-chip-cs3

#AI #Chips #Cerebras #shorts #Shorts
Why Cerebras Refuses to Cut Its Wafer Into ChipsWhy Tesla Wants Lasers, Not Wipers, on Its Robotaxi CamerasTeslas 5 Million Cars by 2030 Goal Needs About 30% Growth a YearWhy Amazons Robotaxi Has No Front or BackWhy Starships Booster Called Off Its Own CatchTeslas New Teaser is CRAZY!Can Teslas Flying Roadster Actually Fly? The PhysicsWhy Some Rectal Cancer Patients Can Now Skip SurgeryTesla Says Optimus Production Has Started at FremontWhy Microsofts Newest AI Data Centers Evaporate Zero WaterSpaceXs Chip Fab Timeline: Test Chip 2027, Volume 2028Tesla Says Robotaxi Hit Nearly 2 Million Unsupervised Miles
Farzad |

Why Cerebras Refuses to Cut Its Wafer Into Chips

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER