Pianissimo: A Sub-mW Class DNN Accelerator with Progressive Bit-by-Bit Datapath Architecture for Adaptive Inference at Edge

Junnosuke Suzuki, Jaehoon Yu, Mari Yasunaga, Ángel López García-Arias, Yasuyuki Okoshi, Shungo Kumazawa, Kota Ando, Kazushi Kawamura, Thiem Van Chu, Masato Motomura

Abstract: Pianissimo is a sub-mW class inference accelerator that adaptively responds to the changing edge environmental conditions with a progressive bit-by-bit datapath architecture. SWHW cooperative control with the custom RISC and the HW counters allows Pianissimo adaptive/mixed precision and block skip, providing a better accuracy-computation tradeoff for low-power edge AI. The 40 nm chip, with 1104 KB memory, dissipates 793-1032$\mu$W at 0.7 V on MobileNetVl, achieving 0. 49-1.25TOPS/W at this ultra-low power range.