AI Paper Atlas

Research map / Architectures and efficiency / Quantization

PsumQuant: In-line Post-training Partial Sum Quantizer for Energy Efficient NPU Inference

Sangwoo Hwang, Yeeun Hong, Jaeha Kung

ICML 2026 · Architectures and efficiency › Quantization · approach: Ultra-low-bit LLM quantization

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 changedWhereMeasuredEffect
use PsumQuant instead of static psum quantizationW8A8 SmoothQuant Llama 3.1 (8B) · WikiText-2, 128x128 arrayPerplexity: 6.46 -> 6.14Improved
use PsumQuant instead of static psum quantizationW8A8 Phi-3 (3.8B) · WikiText-2, 128x128 arrayPerplexity: 6.54 -> 6.27Improved
use PsumQuant instead of static psum quantizationW8A8 Qwen-2.5 (1.5B) · WikiText-2, 128x128 arrayPerplexity: 9.17 -> 8.87Improved
use PsumQuant instead of static psum quantizationW8A8 Llama 3.1 (8B) · Zero-shot average of 5 tasksAverage accuracy (%): 66.1 -> 66.4Improved
use PsumQuant instead of static psum quantizationResNet-18, B_PQ=6 · ImageNet Top-1Top-1 accuracy (%): 67.66 -> 70.04Improved
use PsumQuant instead of static psum quantizationResNet-34, B_PQ=6 · ImageNet Top-1Top-1 accuracy (%): 73.97 -> 74.57Improved
use PsumQuant instead of static psum quantizationSwin-T, B_PQ=6 · ImageNet Top-1Top-1 accuracy (%): 78.78 -> 81.02Improved
add crest-factor (kappa_nz) gradient modulation to scale predictor calibration (ablation)PsumQuant, Llama 3.1 · WikiText-2Perplexity / max absolute errorImproved
use PsumQuant instead of static scaling for scale-psum correlationResNet-18 · Scale vs max absolute psumPearson correlation r: 0.32 -> 0.92Improved
reduce psum bit-width from 24-bit to 8-bit with PsumQuant128x128 weight-stationary array · Total average energyenergy reduction: 45% reductionImproved
use PTQ-based PsumQuant calibration instead of QATPsumQuant · Calibration latencylatency: 3.5x lowerImproved
reduce NPU array size from 256x256 to 64x64PsumQuant calibration, Llama 3.1 8B · Calibration timeTime (min): 18.92 -> 55.97Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy