Memristors have been — so far — mostly a solution looking for a problem. However, researchers at the University of Michigan are claiming the first memristor-based programmable computer that has the potential to make AI applications more efficient and faster.
Because memristors have a memory, they can accumulate data in a way that is common for — among other things — neural networks. The chip has both an array of nearly 6,000 memristors, a crossbar array, along with analog to digital and digital to analog converters. In fact, there are 486 DACs and 162 ADCs along with an OpenRISC processor.
According to the paper, the chip turned in 188 billion operations per second per watt while consuming about 300 mW of power. This includes power consumption attributable to the 180 nM CMOS technology, which is not even close to cutting edge. Newer technology would drive the chip’s performance even higher.
By analogy, you can consider a memristor accumulator as a potentiometer that gets twisted a little more with each input. The final position of the potentiometer indicates the sum of the inputs. This kind of technique has been of great interest as CPU power is becoming harder to increase. If you can do processing in memory instead of the CPU, you can achieve great performance, in theory, since you don’t have to transfer memory to a processing unit, do the computation, and then transfer back to memory.
This reminded us of how we saw memristors solving equations although that wasn’t as integrated as this chip. While some people claim that memristors are a new type of fundamental component, there are those that disagree.