wundr

Sand to Chip

one of the most complex things we make starts as sand.

a story in eight chapters on how a chip is built from sand: purify it to nine nines, pull one crystal out of the melt, grind it to size and slice it into wafers, spin on a film of resist, print the mask with light field by field, etch the print in, stack the copper layers, and cut it into chips

How a chip is built, from the beach to the package, one piece of silicon the whole way. Sand is mostly silicon dioxide (chips start from a very pure quartz); in an arc furnace carbon takes the oxygen away (it leaves as gas), leaving silicon about 98-99% pure, which is then refined (as a gas, distilled and deposited again) to the purity chips need: 99.9999999% or better, nine nines, about one stray atom in a billion (SemiconductorX lists 9N feedstock for Czochralski crystals). That silicon is melted (it melts at 1414 C) and a seed crystal is dipped in and slowly pulled up while it turns: the melt freezes onto it in the same atomic order, one single crystal. The pull speed sets its width: pull faster and it grows thinner (as the industry's diameter-control patents put it). The crystal is grown slightly wider than the wafer it is for and then ground down to the exact diameter with diamond wheels (SVMI's silicon ingot page; eesemi.com); a crystal pulled too thin for 300 mm can only make a smaller standard wafer (200, 150 or 100 mm; Wikipedia, Wafer), and the card grinds it to that and says so. A wire saw slices the whole crystal at once into wafers 300 mm across and under a millimetre thick (775 um), polished to a mirror. A drop of photoresist is spun out into a thin film; its thickness goes roughly as one over the square root of the spin speed, so four times the speed gives half the thickness (Ossila's spin-coating guide, after Meyerhofer's model; real films are calibrated by measuring). A scanner shines UV light through the mask and a 4x reduction lens (ASML's TWINSCAN XT:860N has one), printing one field, then the stage steps to the next and repeats across the wafer (a modern scanner sweeps the light across each field through a slit; the card flashes it whole, as an older stepper does). In a positive resist the exposed parts dissolve in the developer (EPFL's cleanroom notes), plasma etches through the windows they leave, and an oxygen plasma strips the rest. Then layer on layer, the same loop (print, etch, fill with metal, polish flat): the transistors at the bottom, then wires of copper, the fine ones low and the thick ones on top, more than ten metal layers in a modern chip (Wikipedia, back end of line). Last, the wafer is cut into dies and each die is packaged. How many dies fit depends on their size: here the count is whole squares inside the wafer less a 3 mm edge, with no saw lanes between them, and a die is at most 26 mm, the short side of a scanner's 26 by 33 mm field (ASML), which agrees within about a tenth with the usual gross-die formula (AnySilicon) on a 300 mm wafer. Apple's A17 Pro holds 19 billion transistors (Apple's figure) on about a square centimetre (third-party measures give 104 to 114 mm2). Drawn to show, not to scale: a real wafer has more fields and dies than drawn, and UV is invisible (violet here is a false colour). Made from a team member's brief, a 60 s YouTube short on chip manufacturing; its order is followed, its claims of 90 billion transistors in a fingernail, a $300,000 mask and 80 layers of circuits are not repeated.

Concepts: chemistry, crystals, scaling.

An idea you find out by playing, on Wundr, made by @neelu.

On the shelf matter, and also on earth & life.

More on matter