Engineers at the U.S. National Institute of Standards and Technology (NIST) say that the basic theory explaining the origin of a certain type of noise produced by very small transistors is totally wrong.
Known as random telegraph noise, this aberrant signal is becoming a problem for static RAM and flash memory, and it will also become a threat to future low-power logic circuits as their dimensions continue to shrink and the voltage at which they operate decreases. Without a theory, engineers will find it difficult to reduce the noise, predicts Jason Campbell, a National Research Council postdoctoral fellow at NIST. ”If you don’t know where it’s coming from, you don’t know how to fix it,” he says.
”It’s critical to raise this issue,” says Genadi Bersuker, a fellow at Sematech in Austin, Texas. ”The whole spectrum of processes involved in noise is much richer than the simplified description everyone tends to use.”
The prevailing theory says that as an electron passes through the transistor’s channel it can temporarily get stuck in the insulation above. That happens because of a quantum mechanical phenomenon called tunneling, which makes it possible for the electrons to jump through the insulation as if it weren’t there. According to the theory, electrons randomly tunnel into defects in the insulation and then tunnel out. The random moving-stuck-moving pattern was thought to cause the noise.
The duration of the tunneling process is related to the thickness of the insulation, and decades ago, when the theories were formulated, insulation was micrometers thick. But now it is measured in nanometers. Campbell measured production-quality nanometer-scale test transistors, made by TSMC, that were much smaller than those in high-end microprocessors and whose insulation was only 1.4 nm thick. He expected to see tunneling that took nanoseconds to picoseconds. Instead, the process took the milliseconds to seconds you’d see in a much bigger transistor.
”What people thought is very likely not correct,” says Campbell, who presented the results at two IEEE-sponsored conferences in April and May.