Latch-based RISC-V core with popcount instruction for CNN acceleration
Myllynen, Ohto (2021-05-31)
Latch-based RISC-V core with popcount instruction for CNN acceleration
Myllynen, Ohto
(31.05.2021)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2021060233187
https://urn.fi/URN:NBN:fi-fe2021060233187
Tiivistelmä
Energy-efficiency is essential for vast majority of mobile and embedded battery-powered systems. Internet-of-Things paradigm combines requirements for high computational capabilities, extreme energy-efficiency and low-cost. Increasing manufacturing process variations pose formidable challenges for deep-submicron integrated circuit designs. The effects of variation are further exacerbated by lowered voltages in energy-efficient designs. Compared to traditional flip-flop-based design, latch-based design offers area, energy-efficiency and variation tolerance benefits at the cost of increased timing behavior complexity. A method for converting flip-flop-based processor core to latch-based core at register-transfer-level is presented in this work.
Convolutional neural networks have enabled image recognition in the field of computer vision at unprecedented accuracy. Performance and memory requirements of canonical convolutional neural networks have been out of reach for low-cost IoT devices. In collaboration with Tampere University, a custom popcount instruction was added to the cores for accelerating IoT optimized vehicle classification convolutional neural network.
This work compares simulation results from synthesized flip-flop-based and latch-based versions of a SCR1 RISC-V processor core and the effects of custom instruction for CNN acceleration. The latch core achieved roughly 50\% smaller energy per operation than the flip-flop core and 2.1x speedup was observed in the execution of the CNN when using the custom instruction.
Convolutional neural networks have enabled image recognition in the field of computer vision at unprecedented accuracy. Performance and memory requirements of canonical convolutional neural networks have been out of reach for low-cost IoT devices. In collaboration with Tampere University, a custom popcount instruction was added to the cores for accelerating IoT optimized vehicle classification convolutional neural network.
This work compares simulation results from synthesized flip-flop-based and latch-based versions of a SCR1 RISC-V processor core and the effects of custom instruction for CNN acceleration. The latch core achieved roughly 50\% smaller energy per operation than the flip-flop core and 2.1x speedup was observed in the execution of the CNN when using the custom instruction.