Research map / Architectures and efficiency / Quantization
A2Q: Accumulator-Aware Quantization with Guaranteed Overflow Avoidance
Ian Colbert, Alessandro Pappalardo, Jakoba Petri-Koenig
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with A2Q instead of heuristically reducing data bit widths to hit a target accumulator width | QNNs on CIFAR10 and BSD300 · accumulator-constrained Pareto frontier | task performance vs accumulator bit width: most models show less than a 1% performance drop with a 16-bit accumulator | Improved |
| reduce accumulator bit width under A2Q l1-norm constraint | A2Q QNNs · benchmark average | unstructured sparsity and relative accuracy: up to 92% unstructured weight sparsity while maintaining 99.2% of the floating-point model accuracy | Improved |
| use layer-wise minimized accumulator width instead of fixed 32-bit accumulator | baseline QAT on FINN · LUT utilization vs task performance Pareto | resource-accuracy trade-off | Improved |
| minimize accumulator width from final trained weight l1-norm instead of data type bound | baseline QAT on FINN · LUT utilization vs task performance Pareto | resource-accuracy trade-off | Improved |
| train with A2Q rather than baseline QAT with post-hoc accumulator minimization | FINN accelerators · LUT utilization vs task performance Pareto | resource-accuracy trade-off: up to 2.3x reduction in resource utilization over 32-bit accumulator counterparts with 99.2% of the floating-point model | Improved |
| train with A2Q instead of wraparound or saturating accumulators at very low P | 8-bit binary MNIST linear classifier · binary MNIST | top-1 accuracy | Improved |
| use saturation instead of wraparound accumulation (ablation) | 8-bit binary MNIST linear classifier · binary MNIST | top-1 accuracy | Mixed |
| apply A2Q (round-to-zero) in post-training quantization | A2Q · PTQ | accuracy | Hurt |
Closest papers by meaning
- A2Q+: Improving Accumulator-Aware Weight Quantization (ICML 2024 · similarity 0.96)
- Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 2024 · similarity 0.90)
- Data-Free Quantization Through Weight Equalization and Bias Correction (ICCV 2019 · similarity 0.87)
- SYQ: Learning Symmetric Quantization for Efficient Deep Neural Networks (CVPR 2018 · similarity 0.87)
- HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021 · similarity 0.87)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
