We continue to search for something that can overcome AI. wall of eternal memory– Even fast models are bogged down by the time and energy required to carry data between processor and memory. Resistive RAM (RRAM) can get around that wall by allowing computations to be performed in memory itself. Unfortunately, most types of this non-volatile memory are too unstable and unwieldy for that purpose.
Fortunately, a potential solution may soon be found. At the IEEE International Electronic Devices Conference (IEDM) in December, researchers at the University of California, San Diego showed that learning algorithms can be run on an entirely new type of RRAM.
“We actually redesigned RRAM and completely rethought 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 critical digital operations in neural networks (multiplying arrays of numbers and summing the results) can be performed analogously by simply passing current through an array of RRAM cells, connecting their outputs, and measuring the resulting current.
Traditionally, RRAM stores data by creating a low-resistance filament around a high-resistance dielectric material. The formation of these filaments often requires voltages that are too high for standard CMOS, preventing their integration into processors. To make matters worse, filament formation is a noisy and random process, which is not ideal for storing data. (Imagine that the weights of a neural network vary randomly; the answer to the same question will change from day to day.)
Additionally, the noisy nature of most filament-based RRAM cells typically requires the use of selector transistors to isolate them from the surrounding circuitry, making 3D stacking difficult.
These limitations make traditional RRAM unsuitable for computing. In particular, it’s difficult to use filamentary RRAM for the parallel matrix operations that are important for today’s neural networks, Kuzum says.
So the San Diego researchers decided to do away with the filament altogether. Instead, they developed a device that switches an entire layer from high resistance to low resistance and back again. This format, called “bulk RRAM,” eliminates both the cumbersome high-voltage filament formation step and shape-limiting selector transistors.
Although the San Diego group was not the first to build bulk RRAM devices, it made breakthroughs in both shrinking the devices and forming 3D circuits with them. Kuzum and her colleagues scaled RRAM down to the nanoscale. Their device was only 40 nm in diameter. We also succeeded in stacking up to eight layers of bulk RRAM.
With a single pulse of the same voltage, each eight-layer stack of cells can take on one of 64 resistance values, which is very difficult to achieve with traditional filamentary RRAM. And while the resistance of most filament-based cells is limited to kiloohms, the San Diego stack is in the megaohm range, which Kuzum says makes it suitable for parallel operation. e
“You can actually tune it to any value you want, but from an integration and system-level simulation standpoint, we think megaohms is the desired range,” Kuzum says.
The dual benefits of more resistance levels and higher resistance allow this bulk RRAM stack to potentially perform more complex operations than traditional RRAM can handle.
Kuzum et al. assembled multiple 8-layer stacks into 1-kilobyte arrays that do not require selectors. We then tested the array using a continuous learning algorithm. Let the chip classify data from wearable sensors. For example, it reads data from a smartphone worn on the waist and determines whether the wearer is sitting, walking, climbing stairs, or taking other actions, while constantly adding new data. Tests showed an accuracy of 90%, which the researchers say is comparable to the performance of digitally implemented neural networks.
This test exemplifies what Kuzum thinks could particularly benefit from bulk RRAM: neural network models on edge devices that need to learn from their environment without accessing the cloud.
“We are doing a lot of characterization and material optimization to design devices specifically designed for AI applications,” Kuzum says.
The ability to integrate RRAM into arrays in this way is a major advance, said Albert Tallin, a materials scientist at Sandia National Laboratories in Livermore, Calif., and a bulk RRAM researcher who was not involved in the San Diego group’s research. “I think every step around integration is very beneficial,” he says.
But Tallinn highlights a potential hurdle: the ability to retain data for long periods of time. The San Diego group has shown that their RRAM can retain data for several years at room temperature (comparable to flash memory), but data retention at the high temperatures where computers actually operate is less reliable, Talin says. “That’s one of the big challenges with this technology,” he says, especially when it comes to edge applications.
All types of models could benefit if engineers can prove the technology. This memory barrier has only grown higher over the past decade, as traditional memory has not been able to meet the growing demands of large models. Anything that allows the model to work in memory itself can be a welcome shortcut.
From an article on your site
Related articles on the web
