Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AdAM: Adaptive Fault-Tolerant Approximate Multiplier for Edge DNN Accelerators

Mahdi Taheri, Natalia Cherezova, Samira Nazari, Ahsan Rafiq, Ali Azarpeyvand, Tara Ghasempouri, Masoud Daneshtalab, Jaan Raik, Maksim Jenihhin

TL;DR

It is demonstrated that the proposed architecture enables a multiplication with a reliability level close to the multipliers protected by TMR utilizing 63.54% less area and having 39.06% lower power-delay product compared to the exact multiplier.

Abstract

In this paper, we propose an architecture of a novel adaptive fault-tolerant approximate multiplier tailored for ASIC-based DNN accelerators.

AdAM: Adaptive Fault-Tolerant Approximate Multiplier for Edge DNN Accelerators

TL;DR

It is demonstrated that the proposed architecture enables a multiplication with a reliability level close to the multipliers protected by TMR utilizing 63.54% less area and having 39.06% lower power-delay product compared to the exact multiplier.

Abstract

In this paper, we propose an architecture of a novel adaptive fault-tolerant approximate multiplier tailored for ASIC-based DNN accelerators.
Paper Structure (11 sections, 5 figures, 2 tables)

This paper contains 11 sections, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. AdAM architecture
  4. Adaptive adder
  5. Experimental Results
  6. Experimental Setup
  7. Hardware utilization
  8. DNN accuracy
  9. Reliability analysis
  10. Conclusion
  11. Acknowledgement

Figures (5)

  • Figure 1: AdAM architecture
  • Figure 2: Adaptive adder architecture: $a$ and $b$ are inputs, $c$ is carry values and PFA stands for partial full adder
  • Figure 3: Fault-tolerance and error introduced based on different LOD cases
  • Figure 4: Hardware efficiency (area) and fault resilience (fault coverage) trade-offs in AlexNet (up) and VGG-16 (down). Unp-exact: unprotected exact multiplier, Unp-AxM: unprotected approximate multiplier, Pro-TMR: exact multiplier protected by TMR, Pro-AdAM: proposed multiplier
  • Figure 5: PDP and vulnerability tradeoffs (considering SDC-5) in different methods