# HGQ: High Granularity Quantization for Real-time Neural Networks and LUT-Based Inference

- Institution: CERN
- Speaker: Chang Sun
- Year: 2026
- Language: English
- Source: [CERN](https://videos.cern.ch/record/3025769)
- Licence: [CERN](https://copyright.web.cern)
- Who can watch: Teens profiles and older (sources give no ratings; placed by Invisible Network from the source's own words)

Neural networks with sub-microsecond inference latency are required by many critical applications. Targeting such applications deployed on FPGAs, we present High Granularity Quantization (HGQ), a quantization-aware training framework that optimizes parameter bit-widths through gradient descent. Unlike conventional methods, HGQ determines the optimal bit-width for each parameter independently, making it suitable for hardware supporting heterogeneous, arbitrary precision arithmetic. Simultaneously, we introduce HGQ-LUT, a new class of LUT-based layers implemented within HGQ with regular tensor operations during training , enabling the efficient optimization of LUT-based or hybrid neural networks with more than 2 orders of magnitude faster training speed compared to previous methods. We show that the HGQ framework achieves superior performance compared to previous arts, achieving significant reduction in resource consumption and latency while maintaining the accuracy.

00:00:00 Slide 1

00:00:51 Slide 2

00:01:17 Slide 3

00:02:01 Slide 4

00:02:46 Slide 5

00:04:04 Slide 6

00:04:39 Slide 7

00:06:59 Slide 8

00:07:18 Slide 9

00:07:55 Slide 10

00:08:15 Slide 11

00:09:27 Slide 12

00:10:52 Slide 13

00:11:29 Slide 14

00:12:33 Slide 15

00:13:41 Slide 16

00:14:37 Slide 17

00:15:04 Slide 18

00:15:21 Slide 19

00:15:37 Slide 20

00:15:52 Slide 21

00:16:09 Slide 22

00:16:32 Slide 23

00:16:50 Slide 24

00:17:12 Slide 25

00:17:37 Slide 26

00:18:16 Slide 27

00:18:54 Slide 28

## Lectures

1. [HGQ: High Granularity Quantization for Real-time Neural Networks and LUT-Based Inference](https://invisiblenetwork.ai/learn/cern/3025769/c864fbab5d4a), 21 min (CERN's player)