Research Papers 论文研究 3h ago Updated 55m ago 更新于 55分钟前 44

NANQ: Noise-Floor-Aware Mixed-Precision Non-Uniform Quantization for Analog Compute-in-Memory NANQ:面向模拟存算一体的噪声底噪感知混合精度非均匀量化方法

NANQ is a noise-aware mixed-precision non-uniform quantization framework designed specifically for analog compute-in-memory (CIM) hardware Unlike prior methods that minimize ideal quantization error, NANQ explicitly models magnitude-dependent weight noise from real eFlash CIM array measurements The framework converts hardware noise profiles into adaptive quantization density, allocating finer resolution to low-noise weight regions while avoiding wasted precision in noise-dominated regions Layer- 提出NANQ框架,解决模拟存内计算(CIM)中器件变异和读取噪声导致低比特量化模型性能下降的问题 通过测量eFlash CIM阵列响应建模幅度相关的权重噪声,将噪声分布转换为自适应量化密度 引入统一阈值识别各层精度饱和点,实现层间混合精度分配 在2位权重幅度量化下,视觉模型准确率提升8.05个百分点,语言模型PPL平均降低54.7% 混合精度NANQ仅用3.2-3.8等效位即可捕获大部分额外量化资源带来的性能增益

55
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • NANQ is a noise-aware mixed-precision non-uniform quantization framework designed specifically for analog compute-in-memory (CIM) hardware
  • Unlike prior methods that minimize ideal quantization error, NANQ explicitly models magnitude-dependent weight noise from real eFlash CIM array measurements
  • The framework converts hardware noise profiles into adaptive quantization density, allocating finer resolution to low-noise weight regions while avoiding wasted precision in noise-dominated regions
  • Layer-wise bit-width assignment is determined by identifying each layer's precision saturation point under hardware noise using a unified threshold
  • On-chip eFlash CIM SoC experiments show NANQ improves vision model accuracy by 8.05 percentage points and reduces language model PPL by 54.7% on average compared to PowerQuant under 2-bit weight-magnitude quantization, achieving most gains with only 3.2-3.8 equivalent bits

Why It Matters

This work addresses a critical gap in the deployment of quantized neural networks on analog compute-in-memory hardware, where device variation and read noise have historically undermined low-bit quantization accuracy. By incorporating real hardware noise characteristics directly into the quantization framework, NANQ enables practitioners to extract significantly more accuracy from extremely low-bit representations, making analog CIM a more viable path toward energy-efficient edge AI inference.

Technical Details

  • NANQ models magnitude-dependent weight noise derived from measured responses of an eFlash CIM array, capturing the relationship between weight magnitude and hardware noise floor rather than assuming uniform or idealized noise distributions
  • The noise profile is converted into an adaptive quantization density function that assigns non-uniform quantization levels, providing finer resolution in low-noise weight regions and coarser resolution where noise dominates
  • A unified threshold-based method identifies each layer's precision saturation point under hardware noise, enabling automatic layer-wise bit-width assignment without per-layer manual tuning
  • On-chip validation was performed on an eFlash CIM system-on-chip (SoC), demonstrating that mixed-precision NANQ captures most of the accuracy gains from additional quantization resources while maintaining an effective bit-width of only 3.2-3.8 bits
  • Baseline comparison against PowerQuant showed substantial improvements: +8.05 percentage points in vision model accuracy and -54.7% average reduction in language model perplexity under 2-bit weight-magnitude quantization

Industry Insight

  • Hardware-aware quantization frameworks that incorporate real measured noise profiles rather than idealized assumptions will become increasingly important as analog CIM architectures move from research prototypes to production deployment
  • The finding that precision saturation points can be identified with a unified threshold across layers suggests a practical path toward automated, deployment-ready quantization pipelines that reduce the engineering overhead of mixed-precision model optimization
  • For organizations investing in analog in-memory computing hardware, NANQ demonstrates that extreme low-bit quantization (2-bit) can be made viable through noise-aware algorithm-hardware co-design, potentially unlocking significant energy efficiency gains for edge AI applications

TL;DR

  • 提出NANQ框架,解决模拟存内计算(CIM)中器件变异和读取噪声导致低比特量化模型性能下降的问题
  • 通过测量eFlash CIM阵列响应建模幅度相关的权重噪声,将噪声分布转换为自适应量化密度
  • 引入统一阈值识别各层精度饱和点,实现层间混合精度分配
  • 在2位权重幅度量化下,视觉模型准确率提升8.05个百分点,语言模型PPL平均降低54.7%
  • 混合精度NANQ仅用3.2-3.8等效位即可捕获大部分额外量化资源带来的性能增益

为什么值得看

本文针对模拟CIM硬件的实际噪声特性提出了量化方案,填补了现有方法忽略硬件噪声底噪的空白。研究成果对推动存内计算芯片在边缘AI推理中的实际应用具有重要参考价值。

技术解析

  • NANQ框架核心创新在于将硬件噪声感知融入量化过程,通过测量eFlash CIM阵列的实际响应建立幅度依赖的噪声模型,而非仅优化理想量化误差
  • 提出自适应量化密度分配策略,在低噪声区域分配更精细的量化分辨率,在噪声主导区域避免无效精度分配
  • 采用统一阈值方法识别各网络层的精度饱和点,实现层间混合位宽分配,平衡精度与硬件资源
  • 在eFlash CIM SoC上进行片上实验验证,使用视觉模型和语言模型作为基准测试对象

行业启示

  • 存内计算硬件设计需与算法量化方法协同优化,单纯追求低比特量化而忽略硬件噪声特性将导致性能瓶颈
  • 混合精度量化策略在硬件受限场景下具有显著优势,可在有限精度资源下最大化模型性能
  • 实际芯片验证对于量化方法的有效性至关重要,仿真结果与片上实验可能存在显著差异

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Quantization 量化 Research 科学研究 Chip 芯片 Inference 推理 Training 训练