29 June 2010—Random-number generators make cryptography possible, thereby making safe digital communication possible, but because the generators rely on analog components, they are notoriously difficult to reduce in size. Engineers at Intel’s Circuit Research Lab, in Hillsboro, Ore., bet they could build one without the analog parts using the complementary metal-oxide-semiconductor (CMOS) processes that will soon be turning out chips with feature sizes as small as 32 nanometers (and eventually 22 nm).
At the VLSI Symposium earlier this month in Honolulu, the Intel researchers revealed they were close to winning that bet. They reported that they had made an all-digital version of a random-number generator using the 45-nm CMOS process that has been used to build Intel processors since 2007. ”Historically, RNGs have been analog,” says Greg Taylor, director of the Circuit Research Lab. ”But porting to smaller technology nodes [with analog devices] requires a lot of fine-tuning that is unnecessary with digital versions.”
Analog circuits require extra design work to manage things such as the signal-to-noise ratio, Taylor explains. “Supply voltages scale down as we move to more advanced technologies. This supply voltage reduction reduces signal power without changing device noise, consequently reducing the signal to noise ratio.”
The benefits of going digital were immediately obvious. ”The device generates billions of random bits per second and can run at very low voltage,” says Taylor. What’s more, making the generator all digital made it more random. The circuit takes advantage of a phenomenon that is a bugaboo for designers of logic that uses more than one clock. The Intel team engineered the number generator so that in each string of numbers turned out by the machine, every bit is the result of ”metastability.”
Usually, you sample a digital device’s output when it has settled on a definite value, either a one or a zero. Metastability is what occurs when the voltage is sampled during a transition and the bit is caught between a zero and a one. Eventually, the bit drops down to one state or the other, but there is no way to tell on which side it will land. The Intel researchers deliberately sample during transitions; they enhanced the randomness even further by tuning the metastability so that the bit falls to one or zero with fairly equal probability, ”making it essentially a coin flip,” says Taylor.
Taylor says his research group is trying to help make the Internet a secure place, and the all-digital random-number generator is just one piece of the puzzle.