WACT: Lossless Compression of INT8 Weights and Activations to Reduce Data Traffic in Edge NPU Memory Hierarchies
Year: 2026
Type: article
Abstract: Data movement constrains edge convolutional neural network (CNN) accelerators. We present Weight and Activation Compression for Tensor Traffic (WACT), a lossless framework for 8-bit signed integer (INT8) weights and consumer-facing activations. WACT calibrates a mode, tile geometry, and Rice parameter per invocation-aware tensor identity, freezes the policy, and uses a complete metadata- and alignment-inclusive comparison with strict uncompressed (RAW) fallback. Across five image-classification CNNs with one cold weight load and 320 activation passes per model, WACT reduced measured downstream occupied packet bytes by 44.48% (1.801×) over 78,339 events, with zero decoded-INT8 mismatches. A minimal aligned-RAW sensitivity changed this saving by 0.0293 percentage points. On MobileNetV2, FULL_WACT exceeded the best recalibrated single mode and a matched zero-value-compression-style baseline by at least 5.402 and 8.587 percentage points, respectively. A separate labeled 50,000-image validation measured FP32 and INT8 quantize–dequantize (QDQ) Top-1 accuracies of 72.148% and 71.392%, respectively; the latter used scales frozen from the canonical 32-image calibration subset, denoted S0. Explicitly bounded aggregate cold-start SCALE-Sim modeled-access projections were 11.00–17.70%. The complete WACT policy was evaluated in software, whereas the implemented hardware was limited to the single-tile Rice-mode encoder and decoder paths. The routed 28 nm CMOS blocks had standard-cell areas of 0.010942 and 0.010571 mm2, passed post-layout SDF loopback, and consumed 4.939 mW of codec-logic power. The measured downstream-byte reduction therefore provides an energy-saving opportunity whose realization depends on codec overhead and the cost of memory movement in the target hierarchy.
Source: Electronics
Authors: Minseo Lee, Munhyeon Kim
Institution: Seoul National University of Science and Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7204155323