Bhumika Mittal

Modern language models like BERT have revolutionized natural language processing tasks but are computationally intensive, limiting their deployment on edge devices. This paper presents an energy-efficient accelerator design tailored for encoder-based language models, enabling their integration into mobile and edge computing environments. A dataflow-aware hardware accelerator design for language models inspired by SIMBA, use of Approximate Fixed-Point POSIT (AFPOS) based multipliers, and use of high bandwidth memory achieves significant improvements in computational efficiency, power consumption, area, and latency compared to the hardware-realized scalable accelerator Simba. Compared to SIMBA, AxLaM achieves 9x energy reduction, 58% area reduction and 1.2x improved latency, making it suitable for deployment in edge devices. The energy efficiency of AxLaN is 1.8 TOPS/W, 65% higher than FACT which needs pre-processing of language model before implementing it on the hardware.

A balanced generalized de Bruijn sequence with parameters (n,l,k) is a cyclic sequence of n bits such that (a) the number of 0’s equals the number of 1’s, and (b) each substring of length l occurs at most k times. We determine necessary and sufficient conditions on n,l, and k for the existence of such a sequence.