TY - JOUR
AU1 - Thomas, Stuart
AB - Research highlights Neuromorphic computing https://doi.org/10.1038/s41928-022-00846-3 while maintaining accuracy and efficiency, is challenging. Weier Wan and colleagues now report a 48-core RRAM-based CIM chip that can run a range of model architectures at different bit-precisions. The researchers — who are based at Stan- ford University, the University of California San Diego, the University of Pittsburgh, the University of Notre Dame and Tsinghua University — developed the chip by working on all levels of its design. At the device level, they monolithically integrated 3 million ana- logue RRAM devices with complementary metal–oxide–semiconductor (CMOS)-based circuits. A compact and energy-efficient voltage-mode neuron circuit was created to allow variable bit-precision computing and analogue-to-digital conversion. At the array level, an architecture was developed to allow control of the dataflow and routing. The 48 cores are able to perform infer- The use of compute-in-memory (CIM) ence tasks in parallel, and to illustrate the hardware using resistive random-access capabilities of the chip, a range of model memory (RRAM) devices is expected to architectures performing image classifica- improve the energy efficiency of artificial tion, voice recognition and image recovery intelligence systems by eliminating the tasks were successfully run on it. transfer of data between separate memory Stuart Thomas 
TI - Artificial intelligence on a resistive RAM chip
JF - Nature Electronics
DO - 10.1038/s41928-022-00846-3
DA - 2022-09-01
UR - https://www.deepdyve.com/lp/springer-journals/artificial-intelligence-on-a-resistive-ram-chip-o0TryrnZNr
SP - 544
EP - 544
VL - 5
IS - 9
DP - DeepDyve
ER -