%0 Generic %1 delmas2018bittactical %A Delmas, Alberto %A Judd, Patrick %A Stuart, Dylan Malone %A Poulos, Zissis %A Mahmoud, Mostafa %A Sharify, Sayeh %A Nikolic, Milos %A Moshovos, Andreas %D 2018 %K prune quantization %T Bit-Tactical: Exploiting Ineffectual Computations in Convolutional Neural Networks: Which, Why, and How %U http://arxiv.org/abs/1803.03688 %X 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.

@misc{delmas2018bittactical, 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.}, added-at = {2018-03-13T15:00:43.000+0100}, author = {Delmas, Alberto and Judd, Patrick and Stuart, Dylan Malone and Poulos, Zissis and Mahmoud, Mostafa and Sharify, Sayeh and Nikolic, Milos and Moshovos, Andreas}, biburl = {https://www.bibsonomy.org/bibtex/276629178cd30bf0cb186474fa4106878/jk_itwm}, description = {Bit-Tactical: Exploiting Ineffectual Computations in Convolutional Neural Networks: Which, Why, and How}, interhash = {d238d02cfdf0c925f1740c69d19cc52d}, intrahash = {76629178cd30bf0cb186474fa4106878}, keywords = {prune quantization}, note = {cite arxiv:1803.03688Comment: An earlier version of this work titled "JaZ: Enabling Innovation Towards Chaff-Free Deep Learning Computing" was submitted for blind review}, timestamp = {2018-03-13T15:00:43.000+0100}, title = {Bit-Tactical: Exploiting Ineffectual Computations in Convolutional Neural Networks: Which, Why, and How}, url = {http://arxiv.org/abs/1803.03688}, year = 2018 }