A wafer-scale chip is a single chip made from a whole silicon wafer that is never cut into pieces. It is about 21.5 cm on each side, roughly the size of a dinner plate. One AI chip start-up, based in the United States with engineers in Bengaluru, builds it to train and run AI models.
Normally, chips are made many at a time. A wafer is a thin, round disc of pure silicon, 300 mm across. A semiconductor foundry prints the same chip design on the wafer again and again, side by side, in a grid. Each copy carries billions of transistors, the microscopic electrical switches with which a chip computes. The wafer is then cut apart, and each chip-sized piece is called a die. A large AI processor's die measures about 800 square millimetres, so one wafer gives about 60 of them. Each die is then sealed in a package, the case with metal contacts that joins a chip to a circuit board.
Why keep the wafer whole? In a cluster of ordinary GPUs, data must leave one chip and travel through its package, a circuit board and often a cable to reach the next chip. That trip is slower, and it uses far more energy, than moving data within one piece of silicon. On a wafer-scale chip, all the cores and their memory sit on the same silicon. So data passes from core to core along wires a fraction of a millimetre long.
The problem is defects. Every wafer has a few tiny flaws, for example where a speck of dust landed during printing. On an ordinary wafer, a flaw ruins only the die it falls on. That die is thrown away and the rest are sold. The share of good dies is called the yield. On a wafer-scale chip, a single flaw could ruin the whole chip. So the design includes spare cores. After manufacture, tests find the faulty cores. They are switched off, and the wiring is set to use the spares instead.
Consider how the current, third generation of this chip is made at a foundry in Taiwan. The foundry prints 84 identical blocks of cores in a grid, just as it would print 84 separate chips. Then it adds something unusual: extra wires across the narrow gaps between the blocks, where the wafer would normally be cut. The wafer is not cut. Only its round edge is trimmed, leaving the largest square that fits. The finished chip has about 4 trillion transistors, about 9 lakh cores and 44 GB of fast memory on the silicon itself. A GPU, by contrast, keeps most of its data in separate HBM memory chips beside it.
The trade-offs are size and cost. A wafer-scale chip needs its own power supply, its own water cooling and a special cabinet. Each system costs millions of dollars. It is sold mainly to research labs and AI companies in the United States and the Middle East, and as a cloud service. It is one of the bets made by AI accelerator start-ups.