New 3D NAND Memory Based Neuromorphic Computing Platform
American computing and engineering experts have developed an ultra-compact 3D NAND memory-based neuromorphic computing platform. His proposal combines several innovations that increase the entry time window, reducing the delay in data transfer generally generated by this memory concerning other technologies applied to neurocomputing.
Neuromorphic computing is one of the most exciting development fields in computing today. Many researchers are working on technologies that enable this ability to build networks of devices that try to mimic the functioning of the human brain. These are based on devices with their computing capacity, interconnected in a sophisticated network in which each party collaborates in the calculation tasks.
To achieve this, different approaches to in-memory computing are being evaluated. One of them intends to use the widespread 3D NAND flash memory technology to perform the computational operations of neuromorphic computing. Compared to other emerging memories, the problem with this technology is that its performance is inferior. Still, it also offers the advantage of being a proven and accessible technology at a much lower cost than other proposals.
Recently, a team of computer experts from the University of California has presented a new technology that improves the performance of 3D NAND memory in neuromorphic computing applications. They define its development as an ultra-compact and energy-efficient time-domain vector-per-matrix multiplier (VMM) based on a commercial 3D NAND memory structure.
As explained in their work, recently published in the specialized journal Neuromorphic Computing and Engineering, 3D-VMM uses a new successive rescaled resistive integration scheme (RSIR) to dispense with a charging capacitor as bulky as possible the one commonly used in other similar technologies. This reduces the overall device size and power consumption of conventional VMMs.
According to their analyzes, performed on a 55-nanometer technology node, RSIR-3D-VMM’s area efficiency and energy efficiency outperform existing 3D-VMM approaches by a factor of 1.3 to 5 times. And, unlike these, it is possible to adapt this technology to work in a smaller output range. Another innovation that these researchers have developed is another innovation called 3D-cortex, a multipurpose neuromorphic inference processor capable of taking advantage of the 3D-VMM block as a central processing unit.
Their tests have shown that 3D-cortex performance modeling targeting various next-generation neural network benchmarks achieves a record storage efficiency of 30.7 MB mm 2 and maximum energy efficiency of 113.3 Top / J and 10.66 Top / s in neuromorphic computing tasks. And they claim that the RSIR area efficiency improvement provides a lower delay in data transfer, managing to compensate for the VMM performance reduction thanks to a larger input time window.