notesum.ai

Published at November 3

BF-IMNA: A Bit Fluid In-Memory Neural Architecture for Neural Network Acceleration

cs.AR

cs.AI

Released Date: November 3, 2024

Authors: Mariam Rakka, Rachid Karami, Ahmed M. Eltawil, Mohammed E. Fouda, Fadi Kurdahi

Arxiv: http://arxiv.org/abs/2411.01417v1

Latency

Constraint

Per Layer Bitwidth

(weight and activation)

Average

Bitwidth

Normalized

Energy

Normalized

Latency

EDP (J.sec)

Size (MB)

Top 1%

Accuracy

19x{4}

3.29

1.004

0.58

5.6

68.45

High

{8,8,8,8,8,8,8,8,4,8,8,8,8,4,8,4,8,8,4}

7.16

1.13

1.001

1.69

8.7

70.4

Medium

{8,8,8,8,8,4,8,8,4,8,8,8,4,4,8,4,8,4,4}

6.53

1.22

1.002

1.56

7.2

70.34

Low

{8,8,8,4,8,4,8,4,4,4,4,4,4,4,4,4,4,4,4}

5.05

1.90

1.004

1.00

6.1

68.56

19x{8}

1.91

11.2

71.56