Nanowire-Based Logic-in-Memory Device Stacks Transistors and RRAM to Boost Performance, Efficiency
Compact creation some 51 times smaller in footprint and with a 10,000x reduction in latency over traditional CMOS equivalents.
Researchers from Lund University have created a nanowire-based device, which could allow computers to perform processing in-memory — solving a long-standing bottleneck and potentially boosting the performance and energy efficiency of computers considerably.
"Processors have developed a lot over many years," says doctoral student and co-author Saketh Ram Mamidala, in an understated summary of break-neck progress, which has taken us from the four-bit 750kHz Intel 4004 to today's multi-core many-gigahertz beasts in just fifty years. "On the memory side, storage capacity has steadily increased, but things have been pretty quiet on the function side."
That's where the team focused its efforts: Building 3D circuits which integrate both memory and processing capabilities, taking resistive RAM (RRAM) and adding new functions — including, unusually, what Mamidala says is the ability to work "without a power supply."
"Our version [of RRAM] is a nanowire with a transistor at the bottom, and a very small memory element located further up on the same wire," says Lars-Erik Wernesson, professor of nanoelectronics and co-author of the paper. "This makes it into a compact integrated function where the transistor controls the memory element."
"The idea has been around before, but it has proven difficult to achieve performance. Now, however, we have shown that this can be achieved and that it works surprisingly well."
While it's early days for the team's technology, it's possible it could help resolve the von Neumann bottleneck in computing: The fact that it takes time to transfer data between the processor and the memory, creating a bottleneck which saps the overall performance for all but the smallest of data.
By processing in-memory, and with a dramatic reduction in latency, the team's device could offer big gains — particularly for compute- and memory-hungry workloads like machine learning.
The team's work, which resulted in the creation of single-transistor and single-memory devices capable of Boolean logic operations, has been published in the journal Nature Electronics under closed-access terms.