Peer-reviewed research
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference
Benoit Jacob et al., IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2018, 2018
Read the originalRead on October 17, 2026
- What it is
- A Google paper that shows how to store the weights and activations of a neural network as 8-bit integers instead of 32-bit floating-point numbers, so the model runs on integer-only hardware. Its experiments use image classification and detection models, not speech models.
- How much weight it carries
- Peer-reviewed at CVPR 2018 and widely cited as the standard reference for 8-bit quantization. The results are for vision networks such as MobileNets, so the size reduction is a general principle that speech models follow, not a measured speech result.
Facts taken from this source
- Jacob and colleagues report that storing both activations and weights as 8-bit integers gives a close to 4x memory footprint reduction compared with 32-bit floating-point numbers.
“By quantizing both activations and weights as 8-bit integers, we obtain a close to 4x memory footprint reduction compared to 32-bit floating-point representations.” (Abstract)