# Bit-Tactical: Exploiting Ineffectual Computations in Convolutional Neural Networks: Which, Why, and How

## TL;DR

Bit-Tactical introduces a hardware accelerator that exploits various sparsity and precision properties in CNNs during inference, significantly improving performance and energy efficiency without modifying the CNN models.

## Contribution

The paper presents Bit-Tactical, a novel hardware accelerator that leverages multiple sources of ineffectual computations in CNNs to enhance efficiency and performance.

## Key findings

- TCL achieves over 2x to 8x more exploitable ineffectual work.
- TCLp improves performance by 5.05x and energy efficiency by 2.98x over dense CNN accelerators.
- TCL supports both sparse and dense CNNs, including convolutional and fully-connected layers.

## Abstract

We show that, during inference with Convolutional Neural Networks (CNNs), more than 2x to $8x ineffectual work can be exposed if instead of targeting those weights and activations that are zero, we target different combinations of value stream properties. We demonstrate a practical application with Bit-Tactical (TCL), a hardware accelerator which exploits weight sparsity, per layer precision variability and dynamic fine-grain precision reduction for activations, and optionally the naturally occurring sparse effectual bit content of activations to improve performance and energy efficiency. TCL benefits both sparse and dense CNNs, natively supports both convolutional and fully-connected layers, and exploits properties of all activations to reduce storage, communication, and computation demands. While TCL does not require changes to the CNN to deliver benefits, it does reward any technique that would amplify any of the aforementioned weight and activation value properties. Compared to an equivalent data-parallel accelerator for dense CNNs, TCLp, a variant of TCL improves performance by 5.05x and is 2.98x more energy efficient while requiring 22% more area.