Sparse 3x3 convolution has far lower arithmetic intensity than dense
Original title: Theoretical and Empirical Analysis of Sparse Convolution
Preliminary runs report up to 1.65x versus dense on an A100 and 1.52x on a Jetson Nano.
The paper analyzes standard convolution with sparse weights for inference on resource-constrained devices. Roofline analysis puts a 3x3 sparse convolution at about 10 FLOP/byte versus about 96 FLOP/byte for the dense kernel. The authors propose techniques for high-sparsity convolution and report preliminary speedups on an NVIDIA A100 and an NVIDIA Jetson Nano (Tegra X1).
Unconfirmed: The reported speedups are described as preliminary results.
- Arithmetic intensity is about 10 FLOP/byte sparse versus about 96 dense.
- Speedup is up to 1.65x on NVIDIA A100 versus dense convolution.
- Speedup is up to 1.52x on NVIDIA Jetson Nano (Tegra X1).
- Vendors
- NVIDIA
- Products
- A100Jetson Nano
From the original article
1 min read at arxiv.orgDeep neural networks for image processing are increasingly deployed on resource-constrained edge devices such as drones, medical systems, and smartphones, where limited computational resources and memory bandwidth pose significant challenges for efficient inference.
By Keshvi Tuteja · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.