MUMBAI, India, Aug. 31 -- Intellectual Property India has published a patent application (202621074588 A) filed by Indian Institute Of Technology Indore on June 16, 2026, for A System And Method With Bias-Controlled Sensing For Adc-Free Neural Network Inference.
Inventors include Ankit Kumar Tenwar; and Santosh Kumar Vishvakarma.
The application for the patent was published on August 28, 2026, under issue no. 35/2026.
Abstract: ABSTRACT A SYSTEM AND METHOD WITH BIAS-CONTROLLED SENSING FOR ADC-FREE NEURAL NETWORK INFERENCE The present invention relates to a compute-in-memory (CIM) system and method for neural-network inference, and more particularly to an ADC-free RRAM-based CIM architecture employing a plurality of 3T1R bitcells, each comprising three transistors and one resistive random-access memory (RRAM) element. Binary weights are stored in the RRAM elements as low- resistance and high-resistance states, while input activations are applied through bitlines to perform in-memory XNOR-based multiply-accumulate (MAC) computation. An analog MAC output generated within the memory array is applied to a bias-controlled NMOS sensing circuit, which uses a programmable bias voltage (VBN) to shift an effective sensing threshold in accordance with batch normalization parameters. The sensing circuit thereby directly generates a rail-to-rail binary output corresponding to sign activation of the analog MAC output, without requiring an analog-to-digital converter (ADC), a separate comparator block, or separate digital batch-normalization and activation circuitry. The architecture enables integration of multiply-accumulate, batch normalization, and sign activation within a single in-memory framework, thereby reducing data movement, hardware complexity, latency, and power consumption. The invention further provides a scalable array architecture suitable for binary and multi-precision neural-network operations, with improved energy efficiency, robustness under process-voltage-temperature variations, and suitability for edge artificial intelligence and real-time inference applications.
Disclaimer: Curated by HT Syndication.
