Here's the idea. The reticule used in photolithography on leading-edge steppers these days is really big-typically 28 mm by 30 mm. The bigger it is, the fewer times the stepper has to step to print an entire wafer. That's a big deal when you are stepping your way across a 300-mm wafer. Also, of course, a big reticule means you can handle very large designs without having to break them up across multiple mask sets.

But in the era of 0.18 micron and below, not a lot of designs are over an inch square. So normally the reticule for each mask layer includes multiple instances of the mask: printing several copies of the chip at once, if you will.

The clever idea is that all the images in the reticule don't have to be of the same chip. You can combine mask images from several designs, and build different chips on the same lot of wafers, assuming the process was the same for each design.

The major foundries do exactly that. A TSMC service, for instance, divides the reticule into sixths, and pack up to six designs into one superhot-lot run. They do a run every month, in whatever technology seems to be needed at the time.

UMC has a similar program for its standard process. The company divides the reticule up into 5 x 5-mm tiles. Your design can take one tile, or several adjacent tiles. The foundry takes charge of finding designs to pack together and producing the masks.

The idea has advantages for everyone concerned. It allows the foundry to do a fast prototype run without delivering you a complete shuttle-full of 300-mm wafers. Increasingly important with escalating mask costs, it allows several designs-possibly from different companies-to share those costs for their prototype builds, so the first mistake doesn't have to cost a third of a million dollars.

But right now, the multidesign wafer is only a prototyping service. That isn't so much because of production issues as it is a result of test problems. How do you wafer-probe a wafer that has 24 different kinds of dice with 24 pad topologies and 24 test programs? The simple answer is you don't. Multidesign wafers are usually delivered as unprobed dice. After this problem is solved, the whole economics of customer-owned tooling design could change.