Efficient Mixed-Signal Neuromorphic Computing Via Successive Integration And Division
Tech ID: 31649 / UC Case 2020-053-1
From artificial intelligence to neural networks, vector-by-matrix multiplication is a ubiquitous and key operation of network accelerators. As the adoption of groundbreaking network technology grows, the devices that execute the operation must become more efficient. Advancements in time-domain circuits based on non-volatile memory (eNVM) afford new opportunities for efficiency. The main obstacle for fully realizing the increased efficiency are large-footprint, energy-intensive load capacitors. However, they are necessary for handling inputs to time-domain based vector-by-matrix-multipliers (VMMs).
Researchers from UC Santa Barbara have developed a novel successive integration and division (SID) approach for implementing a highly efficient mixed-signal time-domain VMM for low-to-medium precision computing. This approach, which introduces analog computing techniques to digital applications, has demonstrated more than double the efficiency of conventional time-domain VMMs. This technology favorably addresses the paramount concerns of implementing efficient VMMs: power, size, and speed.
• Increased energy efficiency
• Increased speed efficiency
• Compact device foot-print
• Computer Hardware
• IoT Devices
• Mobile Devices
• Artificial Intelligence
• Deep/Recurrent Neural Networks