NANQ: Noise-Floor-Aware Mixed-Precision Non-Uniform Quantization for Analog Compute-in-Memory
AI Summary
NANQ is a mixed-precision, non-uniform quantization framework designed around the hardware noise floor of analog compute-in-memory systems. It derives adaptive quantization density from magnitude-dependent noise measured on an eFlash CIM array and selects layer-wise bit widths using each layer’s precision saturation point. According to the supplied abstract, on-chip eFlash CIM SoC experiments show an 8.05 percentage-point average vision accuracy improvement and a 54.7% average language-model perplexity reduction over PowerQuant under 2-bit weight-magnitude quantization. Mixed-precision configurations reportedly retain most additional-precision gains using 3.2–3.8 equivalent bits.
Why it's worth reading
Analog CIM deployment increasingly depends on handling measured device noise, and NANQ offers a concrete precision-allocation method, although its unusually future-dated metadata and headline results require verification from the full paper.
Deep Read
1. What happened
Original fact: The paper introduces NANQ, a noise-floor-aware mixed-precision non-uniform quantization framework for analog compute-in-memory (CIM). It aims to avoid spending quantization precision where device variation and read noise already dominate the represented signal.
2. Core technology
Original fact: NANQ models magnitude-dependent weight noise using measured responses from an eFlash CIM array. It converts that profile into adaptive quantization density, placing finer resolution in lower-noise regions and reducing ineffective precision in noise-dominated regions. Layer-wise bit widths are selected from each layer’s precision saturation point under hardware noise using a unified threshold.
3. Key evidence and numbers
Original fact: The abstract reports on-chip experiments using an eFlash CIM SoC. Under 2-bit weight-magnitude quantization, NANQ reportedly improves vision-model accuracy by an average of 8.05 percentage points and reduces language-model perplexity by 54.7% on average versus PowerQuant. Mixed-precision NANQ reportedly captures most gains available from extra quantization resources with 3.2–3.8 equivalent bits. The abstract does not identify models, datasets, sample sizes, or absolute metrics.
4. Why it matters
Analysis: Conventional quantization usually optimizes rounding error under ideal digital arithmetic, while analog CIM accuracy is also bounded by device states, read noise, and magnitude-dependent errors. Connecting measured noise directly to codebook density and layer precision could produce a better accuracy-resource tradeoff than uniformly adding bits.
5. Practical impact
Analysis: NANQ could fit CIM deployment pipelines that can characterize hardware before mapping a model: measure array noise, derive a non-uniform codebook, and assign layer bit widths. If the reported 3.2–3.8 equivalent-bit result survives chip, temperature, and aging variation, it may reduce storage, conversion overhead, and wasted precision. Those system-level benefits are not quantified in the supplied abstract.
6. Limitations and uncertainty
Known gaps: The abstract omits model and dataset names, process technology, array size, energy, latency, area, calibration cost, statistical variation, and comparisons beyond PowerQuant. Unverified inference: The measured noise profile may be specific to one eFlash implementation and could require recalibration across chips, temperatures, or aging states. The supplied publication date is August 3, 2026, a future date, so the paper’s status and reported results require verification.
7. Original sources
- arXiv abstract page: arXiv:2608.02700
- This assessment uses only the title, abstract, and metadata supplied by the user; no unsupported experimental results or author conclusions have been added.