MUMBAI, India, March 13 -- Intellectual Property India has published a patent application (202641025548 A) filed by Afroz Fatima; and Shaik Mohammed Waseem, Hyderabad, Telangana, on March 4, for 'system and method for temporal reuse in a neural network accelerator.'
Inventor(s) include Afroz Fatima; and Shaik Mohammed Waseem.
The application for the patent was published on March 13, under issue no. 11/2026.
According to the abstract released by the Intellectual Property India: "System and Method for Temporal reuse in a Neural Network Accelerator The present invention discloses a system and method of temporal reuse in a neural network accelerator. The neural network accelerator (100) comprises resistive memory crossbar arrays (102) configured to store neural network weights and perform analog matrix-vector multiplication, an analog accumulation circuit (104) configured to retain analog partial sums across successive inference timesteps, and a control circuit (106) configured to compute a delta input vector from successive input vectors. One or more digital-to-analog converters (108) selectively apply either a full input vector or the delta input vector to the crossbar arrays, while an analog-to-digital converter (110) is selectively enabled based on threshold and retention conditions. A finite state machine controls full compute, temporal reuse, hold, and reset operations. During operation, incremental analog contributions corresponding to the delta input vector are accumulated onto stored analog partial sums, avoiding repeated full computation and unnecessary conversion events while maintaining bounded inference accuracy."
Disclaimer: Curated by HT Syndication.
