The electronic system of the future will monitor itself, change the functions it performs, and repair its damaged circuits—all without external intervention. Such is the dream of autonomic computing. Although the dream is not yet reality, engineers and scientists at IBM Corp. have taken a big step with the development of an on-chip fuse that is electrically blown—or programmed, as the company prefers to call the process—by using a physical effect heretofore considered a serious reliability problem in semiconductor circuits.

According to IBM, combining the new eFuse technology with already available on-chip built-in self-test and -repair circuitry will yield a chip capable of diagnosing its failures and then fixing them by blowing fuses to reroute its circuits. The built-in self-test circuitry determines which parts of the chip do not work and sends the information to the self-repair circuitry, which figures out what fuses to open to replace the failing circuits with spare, redundant ones.

For several generations of semiconductor process technology, fuses and redundant circuits have worked hand in hand to repair chips and increase the number of usable chips on a wafer. They are particularly important in chips with embedded DRAMs, in which cutting fuses allows extra rows or columns of memory cells to replace damaged ones.

But those fuses are cut from outside the chip, with a laser slicing through metal lines placed on the chip's topmost layer. "That process is a lengthy and costly one," says Subramanian Iyer, distinguished engineer and manager of 90-nanometer bulk technology development at IBM's Semiconductor Research and Development Center, in East Fishkill, N.Y., "because the wafer has to shuttle back and forth from the testing station to the laser station and back to the tester—and must be carefully aligned after each move. What's more, laser fuses can be programmed only while the chips are still on the wafer."

Another problem with the laser fuses, Iyer says, is that their dimensions have not been shrinking even as microchip wiring and components have gotten smaller. That's because the fuses' dimensions are tied to the wavelength of the laser and the resolution limits of the optics used to cut them, which are several times as large as the features that make up the transistors on new chips.

What the IBM team came up with for the eFuse is a tiny strip of polycrystalline silicon—or polysilicon—roughly 1.2 micrometers long and 0.12 µm wide, covered with a thin layer of cobalt silicide, the same materials that make up a transistor gate. The fuse is opened through a process called electromigration, in which current pushes the atoms in small wires out of place. Electromigration has been a reliability headache for years because it creates voids that can break the interconnect wires on an integrated circuit, resulting in chip failure. And it is becoming a bigger problem as new generations of chips rely on narrower wires and throw off more heat, both factors that increase electromigration.

But the IBM scheme puts electromigration to good use.

At room temperature the polysilicon in the fuse is a poor conductor, because it has few of the impurities that help transport charge in diodes, transistors, and similar devices. Cobalt silicide, on the other hand, is a good conductor, so most of the current applied to the polysilicon-cobalt-silicide strip goes through the cobalt silicide. At sufficiently high current, electromigration occurs, and atoms in the silicide begin to drift along with the electrons in the current, from the negative to the positive side of the circuit, eventually making a gap in the material. At the same time, the high density of current through the fuse causes it to heat up. Once it is hot, electromigration increases in the silicide, and the conductivity of the underlying polysilicon goes up as well, allowing current to pass through it. So electromigration continues even after a break forms in the silicide. "It's like two conductors in parallel, but the top conductor is not continuous," Iyer explains.

After a time the current is removed, the fuse cools down, the polysilicon becomes a poor conductor again, and the fuse stays permanently open.

Iyer says that eFuses have many advantages. The wafers are tested and the fuses are programmed in a single step, without moving the wafer from one station to another. Fuses can also be programmed after the wafer is diced and the chips are packaged. They can even be programmed after the products are in the field. Another advantage of the technology is that its manufacture requires no additional materials, processes, or photolithography masks. Moreover, the programming uses only those voltages that typically operate chips.

In addition to replacing failing circuits with redundant ones, the programming of eFuses can also vary the operating voltages of individual circuits on a chip by adjusting the inputs to the on-chip voltage regulator. The built-in self-test circuitry determines the voltage range over which a circuit works properly. And if it fails at the nominal operating voltage, the circuitry calculates which fuses to program in order to bring the supply voltage within the circuit's range of operation. Also, Iyer says, for chips that are running too hot, bringing down the voltage will allow them to consume less power and thus run cooler.

IBM's approach is not the only one for programming fuses electrically, but it's the only one that uses electromigration. Actel Corp., in Mountain View, Calif., a manufacturer of field-programmable gate arrays, has long used what it calls an antifuse to program its FPGAs. The antifuse is a layer of insulator sandwiched between two conductors. Applying a voltage greater than the breakdown voltage of the insulator blows a hole in it, establishing an electrical connection between the two conductors.

Jim Turley, an independent semiconductor-industry analyst, sees the IBM development as only an incremental improvement. "It is a manufacturing bonus in that you don't have to focus a laser. After testing, you can make the chips do their own lobotomies," Turley says. "But it doesn't strike me as a huge fundamental upheaval of technology." IBM is already using the new fuses to improve yield in its most advanced manufacturing process, which produces chips with line widths of 90 nm [see photo, "Self-Healing Circuits"].

Richard Doherty, director of the Envisioneering Group, an industry analysis firm in Seaford, N.Y., sees things differently. Once customers understand that the technology will allow them to design circuits in new ways, we could see some interesting new chip applications, Doherty says. "The fact that it fits into existing systems and can be driven by software is going to be exciting to system designers," he predicts, "especially for the design of space probes, biomedical devices, security and defense systems, and other mission-critical applications where you don't want to pay a house call to reconfigure a chip."