Wafer-scale engine
Cerebras's design that turns a whole silicon wafer into one giant chip instead of cutting it into many small ones.
Also: WSE, wafer-scale chip
Chipmakers normally cut a 300 mm silicon wafer into dozens of separate chips. Cerebras Systems instead keeps the wafer whole and wires it into one processor. Its third generation, the WSE-3 of 2024, has about 4 trillion transistors and 900,000 cores, with large amounts of fast on-chip memory. Keeping data on one piece of silicon avoids the slower links between chips that GPU clusters depend on, which Cerebras says makes it very fast at generating tokens. The company sells systems and runs its own inference cloud.