Research map / Architectures and efficiency / Quantization
JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes JAQ, a differentiable joint search over network architecture, ultra-low mixed-precision bitwidths (2/3/4/8-bit), and accelerator design, using channel-wise sparse quantization to cut memory and BatchTile to speed compiler mapping search.
Problem
Joint co-design of neural architecture, low-bit quantization, and hardware accelerator is memory-hungry during quantization-aware supernet training and slow in accelerator/compiler-mapping search, and prior differentiable co-search methods fail at low bit-widths.
Approach
Quantize only the top-K% most important activation channels (ranked by BatchNorm scale factors) during supernet search to cut GPU memory, and encode all tile sizes as batch entries so a hardware generation network plus a latency/energy estimator finds the optimal tiling in parallel.
Key results
- On ImageNet, JAQ reaches about 70.2% accuracy at lambda=0.001 vs 62.8% for Auto-nba at the same setting (roughly 7-point gain as claimed).
- Channel-wise sparse quantization reduces memory about 5x versus the non-optimized scenario, per the authors.
- BatchTile reduces accelerator/compiler-mapping search time per iteration to 0.15 s versus about 30-50 s for prior search.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Auto-nba with JAQ for joint architecture, bitwidth, and accelerator search | FBNet · CIFAR-10 (λ=0.002) | accuracy: 89.643 -> 92.163 | Helped |
| replace Auto-nba with JAQ for joint architecture, bitwidth, and accelerator search | FBNet · CIFAR-100 (λ=0.004) | accuracy: 60.169 -> 72.440 | Helped |
| replace Auto-nba with JAQ for joint architecture, bitwidth, and accelerator search | FBNet · ImageNet (λ=0.005) | accuracy: 62.423 -> 69.132 | Helped |
| replace Auto-nba with JAQ for joint search | ImageNet · ImageNet | search time (GPU hours): 180 -> 160 | Helped |
| select the most important activation channels for sparse quantization rather than using heterogeneous sampling (ablation) | CIFAR-100 joint network and bitwidth search · CIFAR-100 | Top-1 accuracy: 55.704 -> 65.863 | Helped |
| quantize only the first activation channel instead of selecting important channels (ablation) | CIFAR-100 joint network and bitwidth search · CIFAR-100 | Top-1 accuracy: 65.863 -> 64.355 | Hurt |
| increase selected important activation channels from K=1 to K=5 (ablation) | JAQ sparse quantization · CIFAR-100 | Top-1 accuracy: 65.377 -> 65.863 | Helped |
| select important activation channels for sparse quantization (ablation) | CIFAR-100 joint network and bitwidth search · CIFAR-100 | Misguided Search (%): 40 -> 0 | Helped |
| quantize only the first activation channel instead of dynamically selecting important channels (ablation) | CIFAR-100 joint network and bitwidth search · CIFAR-100 | Misguided Search (%): 0 -> 5 | Hurt |
| use BatchTile for compiler mapping search | JAQ accelerator search · accelerator search | search time (s): 30 -> 0.15 | Helped |
Surprising findings
- Auto-nba's unconstrained low-bit search converges to 2-bit for many operators, causing misguided search (40% in the reported comparison).
Limitations
- Evaluated only on small image classification datasets (CIFAR and ImageNet) with a single hardware template (BitFusion-style ASIC).
- Search relies on a fixed FBNet search space and a 2/3/4/8-bit candidate set.
- Hardware sensitivity and K values are tuned by hand through lambda sweeps.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet
Benchmarks: CIFAR-10 CIFAR-100 ImageNet
Metrics: Top-1 accuracy EDAP search time GPU memory
Baselines: Auto-nba NAAS OQAT BatchQuant DANCE
Builds on: Auto-nba DANCE FBNet Gumbel-Softmax BitFusion
Closest papers by meaning
- HAQ: Hardware-Aware Automated Quantization With Mixed Precision (CVPR 2019 · similarity 0.91)
- BatchQuant: Quantized-for-all Architecture Search with Robust Quantizer (NeurIPS 2021 · similarity 0.89)
- InfoQ: Mixed-Precision Quantization via Global Information Flow (AAAI 2026 · similarity 0.88)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.86)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.86)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 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.