Fact · Speech recognition
4times smaller memory footprint
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.
Checked against its source on October 17, 2026
In the source's own words
“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.”
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, Abstract.
Open the original- What it measures
- The paper introduces integer-only quantization for neural networks that run on mobile accelerators. The reduction is for the stored numbers in the model, not for speed.
- What it does not show
- Measured on vision networks such as MobileNets, not on speech models. The reduction is close to 4x, not exactly 4x, and accuracy depends on how the model is trained for quantization.