Aaron J. Hand
Today's typical microprocessor is a complex creature. The continuing trend toward faster, smaller and denser CMOS circuits does not help engineers measure a chip's signals for possible defects. Traditional inspection methods, which involve probing the device, face difficulties as processor technologies advance.
Researchers at IBM's T.J. Watson Research Center have a solution that would allow inspection by examining infrared light through the back of the processor. The fact that it has become common to mount microprocessors face down makes probing even more problematic and this new method that much more viable.
James C. Tsang and Jeffrey A. Kash call their technique picosecond imaging circuit analysis. Every time a field effect transistor switches, it emits a pulse of light lasting 10 to 100 ps, most of which is in the infrared. Because silicon is transparent to infrared light, the light shows through the back of the microprocessor.
To view the pulses of infrared light, the researchers experimented with a few types of detectors: a liquid nitrogen-cooled charge-coupled device, a microchannel-plate photomultiplier and an avalanche photodiode. The photomultiplier -- supplied by Quantar Technology Inc. of Santa Cruz, Calif. -- gave the best spatial and temporal results, allowing the photons to be detected one at a time.
Everything at once
Because each pulse of light is so weak, the engineers must record the circuit operations for several minutes to really see the big picture. "You have to run the chip for millions and millions of cycles to get enough light," Tsang said. "But if you add up enough switching events, you can get a good picture of what's going on." Optics allow the researchers to see the whole sample at once, and by viewing light emission in the image, they can see exactly which gates are malfunctioning.
Other inspection techniques use lasers to probe the microprocessor. But the laser may perturb the sample so much that it actually causes failures. This is an area where Tsang and Kash's technique has the advantage, since it is entirely passive. The imaging circuit analysis still has heat issues to deal with, since the heat sink must be removed and the back planed down to help the light come through. "It's something you have to worry about," Tsang said, "but it's not a show stopper."
Laser probing, like physical probing, has difficulty dealing with ever-smaller devices. Tsang and Kash's method, on the other hand, actually improves as circuit size decreases, since the voltages used to operate transistors do not scale down as quickly as their dimensions. The researchers have successfully tested linewidths as small as 0.125 µm, half the size of commercial linewidths.