15 June 2009—You can tell a car’s age and how long you expect it to last just by looking at the mileage on its odometer. University of Minnesota electrical engineering professor Chris H. Kim wants to do essentially the same for silicon chips. This week at the VLSI Symposia, in Kyoto, Japan, he reported the development of a silicon circuit that tracks the three signs of aging that can slow down a chip’s performance. The hope is that integrating such a circuit into a microprocessor could help the microprocessor compensate for its own aging.
”The notion that a transistor ages is a new concept for circuit designers,” says Kim. Transistor aging has traditionally been the bailiwick of engineers who design the processes that make transistors; they also formulate recipes that guarantee the transistors will operate within a certain frequency and other parameters typically for 10 years or so.
But as transistors are scaled down further and operated with thinner voltage margins, it’s becoming harder to make those guarantees. With time, even the smallest of variations can drag on a transistor’s operating frequency, and an aging device may not perform adequately at a very low voltage. Because of such dilemmas, transistor aging is emerging as a circuit designer’s problem.
”The process people are working very hard to try to reduce [age-related] degradation,” says Vijay Reddy, a senior member of the technical staff at Texas Instruments working on complementary metal-oxide-semiconductor reliability. ”But what people are realizing is that we need help from the design side.” Reddy thinks that knowledge gleaned from circuits like Kim’s ”silicon odometer” could help circuit designers incorporate reliability issues early on in the design process.
Or the odometer could have a more direct use. Say, for example, you want to hold down the power consumption in a laptop processor. You could use a technique called dynamic voltage scaling to dial down the voltage to conserve power. But a chip’s transistors may be too old to operate correctly at a very low voltage. A detector embedded in the circuitry could tell the voltage regulator circuit the state of the chip’s transistors and allow it to dial the voltage up a notch to compensate.
”I think these sorts of wear-out monitors will show up on ICs,” says Kenneth Shepard, a professor of electrical engineering at Columbia University in New York City. Shepard expects that as chips include more and more processor cores, some will be kept in reserve to compensate for others that age. On-chip monitors will watch active cores and shut them down when age degrades their performance. Unused cores will then be started up to replace them. ”That’s where things are going,” says Shepard.
Since 2005, Kim’s group has been working on a silicon odometer in collaboration with IBM’s T.J. Watson Research Center, in Yorktown Heights, N.Y. The earliest version could measure aging, but it could not distinguish among the three types of aging: hot carrier injection (HCI), bias temperature instability, and time-dependent dielectric breakdown. The new version can distinguish among them, potentially making it easier for circuit designers to compensate.
”Simply put, HCI is an aging that happens when a transistor is switching,” explains Kim. Charges get stuck in the transistor’s gate insulation and alter the voltage at which the device turns on.
Bias temperature instability, an effect that occurs when a transistor is held in the ”on” state, has a similar impact: Charge traps form at the interface between the transistor channel and the gate insulation, affecting the turn-on voltage. This kind of instability is a problem for more and more transistors, notes Kim, as advanced chipmakers adopt high-k dielectrics and metal gates.
Time-dependent dielectric breakdown ”is more like a catastrophic failure,” says Kim. As the transistor ages, defects pile up in the gate insulation, and a short circuit forms.
In order to measure the signs of all three, Kim and his colleagues created a circuit that measured the beat frequency—the interference between two signals of slightly different frequency—of two oscillators. One oscillator was stressed to simulate aging, and the other wasn’t. The circuit was able to discern changes in the range of 0.01 percent—enough detail to tease out all three signs of aging. It was even sensitive enough to see aging under normal operation, says Kim.
The next step will be to extend the age-detection circuitry to find signs of aging in a chip’s interconnects, not just in the transistors.