RAMP: Robust Adaptive Mixed-Precision Quantization for Edge CPU Vision Models
David Población-Criado, Dario Garcia-Gasulla, Eduardo Quinones
cs.CV · cs.LG · cs.PF
Quick verdict
Claims JSD+K-Means gives near-lossless INT8 mixed precision on edge CPUs (1.81x mean speedup); a careful, honest study, but narrow: 4 models, one dataset.
Summary
The paper presents RAMP, a solver-free mixed-precision quantization (MPQ) pipeline for deploying vision models on edge CPUs, built on two empirical findings. First, a systematic study of 13 layer-sensitivity metrics for INT8 post-training quantization across four architectures (ResNet-18, EfficientNet-B0, ConvNeXt-Tiny, TinyViT) on a synthetic classification dataset shows that the gradient-based proxy HAWQ-V2 fails catastrophically on 4 of 8 model-hardware configurations and weight-based statistics on 2, while output-space Jensen-Shannon Divergence (JSD) on final logits has zero failures, with mean Recall@Top-N% of 0.766 versus 0.509 for HAWQ-V2. Second, fixed percentile thresholds are fragile under skewed sensitivity distributions; the authors replace them with 1D K-Means clustering (K=5), which improves silhouette score on TinyViT from 0.18 to 0.93. A policy is selected from K candidates via Pareto knee-point; validated on Apple M1 and Raspberry Pi 5 (Cortex-A76) with ONNX Runtime, RAMP achieves near-lossless accuracy (mean drop 0.59%) and a mean 1.81x speed-up over FP32, this figure being the mean of the eight per-configuration ratios. An ablation shows that excluding layers with negligible per-layer speed-up is counterproductive: INT8 subgraphs fragment, operator fusion is disabled, and end-to-end latency degrades in 4 of 8 configurations (e.g., ResNet-18 on Raspberry Pi 5 from 20.8 ms to 49.7 ms). Limitations: a single synthetic dataset, a binary INT8/FP32 precision set, OAT profiling that ignores inter-layer interactions, and x86-64 acceleration left to future work.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Gradient-based sensitivity estimation (HAWQ-V2) causes accuracy collapse in 4 of 8 model-hardware configurations and weight-based STD in 2, while output-space JSD causes zero catastrophic failures across the same 8.
- JSD retrieves the most quantization-sensitive layers with mean Recall@Top-N% of 0.766, versus 0.509 for HAWQ-V2 and 0.496 for STD (Table 1).
- 1D K-Means clustering of JSD scores (K=5) adapts to skewed sensitivity distributions where fixed percentile thresholds fail, with silhouette 0.93 vs 0.18 on TinyViT (Figure 2).
- RAMP attains near-lossless accuracy (mean drop 0.59%) with a mean 1.81x end-to-end speed-up over FP32 on Apple M1 and Raspberry Pi 5 (Table 2).
- Excluding layers with negligible per-layer speed-up is counterproductive because it fragments INT8 subgraphs, disables operator fusion, and degraded end-to-end latency in 4 of 8 configurations (Section 6.2).
Strengths
Genuine negative results reported and instrumented
Section 6.2 treats the paper's own hardware-aware hypothesis as falsifiable and refutes it: the JSD+Latency filter degrades end-to-end latency in 4 of 8 configurations (ResNet-18 on Raspberry Pi 5 goes 20.8 ms to 49.7 ms at identical accuracy; Figure 4 shows a fused subgraph rising from 7.7 ms to 16.5 ms), directly undermining the additive-cost assumption of ILP-based MPQ [6,7].
Consistent, pre-defended metric-evaluation protocol
Recall@Top-N% (Eq. 3) plus a collapse criterion (>15 pp below FP32) that the authors show is non-load-bearing (no configuration falls in the 8-15 pp range) give Table 1's 13-metric comparison a reproducible basis; the authors also demote Spearman correlation to a sanity check because output-space proxies are favored by construction (Section 4.2).
Headline numbers are arithmetically verifiable from Table 2
Deployment realism over analytic cost models
Thresholding ablation with an internal validity metric
Weaknesses
All accuracy evidence comes from one synthetic dataset
Binary precision set shrinks the stated problem
Sparse statistical reporting
KL is identical to JSD in all 48 cells of Table 1
No sensitivity analysis for K or knee-point selection
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
RAMP (BMVC 2026; Barcelona Supercomputing Center / UPC) targets layer-wise INT8/FP32 mixed-precision quantization for edge CPU inference. It makes three claims: (1) output-space Jensen-Shannon Divergence (JSD) is a more reliable sensitivity proxy than gradient-based (HAWQ-V2) or weight-based criteria - catastrophic collapse in 4/8 and 2/8 model-hardware configurations respectively versus 0/8 for JSD (Sections 4.2, 6.1, Table 2); (2) fixed percentile thresholding is fragile under skewed sensitivity distributions, and 1D K-Means (K=5) fixes this (Section 4.3, Figure 2); (3) per-layer latency filtering is counterproductive because it fragments computational graphs and disables operator fusion (Section 6.2, Figure 4). Headline result: near-lossless accuracy (mean drop 0.59%) at a mean 1.81x end-to-end speed-up over FP32 on Apple M1 and Raspberry Pi 5.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What speedup does RAMP achieve for quantized models on edge CPUs?
RAMP reports a mean 1.81x end-to-end speed-up over FP32 with near-lossless accuracy (mean Top-1 drop of 0.59%) across four vision models on Apple M1 and Raspberry Pi 5. For comparison, uniform INT8 quantization is faster (up to about 5x) but collapses accuracy (drops of 28-70 points) on 3 of the 4 tested architectures.
Why do gradient-based quantization sensitivity metrics like HAWQ-V2 fail?
The paper shows HAWQ-V2's Hessian-trace approximation misranks the layers most vulnerable to INT8 quantization on modern architectures with skewed sensitivity distributions, causing accuracy collapse (over 15 points below FP32) in 4 of 8 model-hardware configurations, whereas output-space Jensen-Shannon Divergence on final logits caused zero collapses.
What is graph fragmentation in mixed-precision quantization?
It is the effect where keeping individual slow layers in FP32 breaks contiguous INT8 subgraphs and inserts Quantize/Dequantize conversion nodes that disable operator fusion in inference engines like ONNX Runtime. The paper measured a fused subgraph's latency rising from 7.7 ms to 16.5 ms as a result, so excluding individually slow layers can worsen total latency.
Is code available for the RAMP quantization method?
Yes. The authors state that code is available at github.com/davidpob99/ramp-mpq. The pipeline uses standard tooling (ONNX Runtime CPU execution provider, post-training quantization), so accuracy and policy-generation results should be reproducible on commodity hardware, though exact latency figures require the specific ARM64 boards.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 are present only as captions and axis labels in the extraction; per-layer speedup distributions, clustering boundaries, and silhouette plots could not be independently verified.
- Table 1's dense layout was extracted as a single block; the printed k values per threshold (k=2 to k=20) were taken as printed and cross-checked against L per model.
- Some aggregate claims (mean drops 0.59% and 0.45%) were verified by recomputation against Table 2 and are consistent within rounding.
- No supplementary material or appendix was included in the extracted text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.