The hunt is on for anything that can surmount AI’s perennial memory wall–even quick models are bogged down by the time and energy needed to carry data between processor and memory. Resistive RAM (RRAM)could circumvent the wall by allowing computation to happen in the memory itself. Unfortunately, most types of this nonvolatile memory are too unstable and unwieldy for that purpose.
Fortunately, a potential solution may be at hand. At December’s IEEE International Electron Device Meeting (IEDM), researchers from the University of California, San Diego showed they could run a learning algorithm on an entirely new type of RRAM.
“We actually redesigned RRAM, completely rethinking the way it switches,” says Duygu Kuzum, an electrical engineer at the University of California, San Diego, who led the work.
RRAM stores data as a level of resistance to the flow of current. The key digital operation in a neural network—multiplying arrays of numbers and then summing the results—can be done in analog simply by running current through an array of RRAM cells, connecting their outputs, and measuring the resulting current.
Traditionally, RRAM stores data by creating low-resistance filaments in the higher-resistance surrounds of a dielectric material. Forming these filaments often needs voltages too high for standard CMOS, hindering its integration inside processors. Worse, forming the filaments is a noisy and random process, not ideal for storing data. (Imagine a neural network’s weights randomly drifting. Answers to the same question would change from one day to the next.)
Moreover, most filament-based RRAM cells’ noisy nature means they must be isolated from their surrounding circuits, usually with a selector transistor, which makes 3D stacking difficult.
Limitations like these mean that traditional RRAM isn’t great for computing. In particular, Kuzum says, it’s difficult to use filamentary RRAM for the sort of parallel matrix operations that are crucial for today’s neural networks.
So, the San Diego researchers decided to dispense with the filaments entirely. Instead they developed devices that switch an entire layer from high to low resistance and back again. This format, called “bulk RRAM”, can do away with both the annoying high-voltage filament-forming step and the geometry-limiting selector transistor.
3D memory for machine learning
The San Diego group wasn’t the first to build bulk RRAM devices, but it made breakthroughs both in shrinking them and forming 3D circuits with them. Kuzum and her colleagues shrank RRAM into the nanoscale; their device was just 40 nm across. They also managed to stack bulk RRAM into as many as eight layers.
... continue reading