Cerebras Wafer-Scale Engine 2

What it is

Cerebras unveiled the Wafer Scale Engine 2 (WSE-2) and called it the largest AI processor ever made.^[1] The name describes its physical form. Conventional processors are cut from a silicon wafer into many separate chips, each packaged and connected to the others through wires or circuit boards. A wafer-scale engine instead keeps the wafer whole and treats it as one processor. The WSE-2 covers 46,225 square millimetres and stitches together 84 dies to form that single piece of silicon.^[2]

How it is built

Cerebras said the chip is manufactured by TSMC on a 7-nanometer process.^[1] Process numbers describe the generation of manufacturing technology a fab uses; smaller generations generally allow more transistors in the same area. On that process, Cerebras said, the WSE-2 holds 2.6 trillion transistors and 850,000 AI-optimized cores.^[1] Memory sits on the wafer itself rather than in separate chips off to the side: 40 GB of on-wafer memory, with 20 petabytes per second of on-chip memory bandwidth.^[2]

That bandwidth figure is the one that matters most for AI work. Training and running large neural networks involves moving enormous quantities of numbers — weights, activations, gradients — between where they are stored and where they are computed. If memory is far away, the processor spends much of its time waiting, and the arithmetic units sit idle. Putting memory on the same wafer and giving it very high aggregate bandwidth is an attempt to reduce that waiting.

How it compares

Cerebras framed the chip against two yardsticks. The first was the largest GPU at the time, which the company said had 54 billion transistors — a comparison meant to show the difference in scale between a wafer-scale design and a conventional chip.^[1] The second was its own first-generation part: Cerebras said the WSE-2 more than doubles the transistor count, core count, memory, memory bandwidth and fabric bandwidth of the original WSE.^[1] Fabric bandwidth refers to how fast the cores on the chip can pass data among themselves, which matters because a wafer-sized processor has to keep hundreds of thousands of cores working together.

Where it is used

The WSE-2 is not sold as a standalone part. It powers Cerebras's CS-2 system, which the company said more than doubles the performance of the earlier CS-1.^[1] A system built around a wafer-scale chip has to supply power and cooling across a very large piece of silicon, so the computer around the chip is as much a part of the design as the chip itself.

Why it matters

The WSE-2 is a bet on a different direction from the one the rest of the industry took. Most AI accelerators are built from many smaller chips connected by high-speed links; wafer-scale integration tries to remove those boundaries altogether by keeping a whole wafer as one device. The appeal is scale and closeness of memory to compute. The difficulty is that a wafer is a large area of silicon, and defects in manufacturing are a fact of life across such an area, so making a whole wafer behave as one reliable processor is the hard engineering problem the approach has to solve.