REQAP: Resilient Weight Packing and Quantization for Edge DNN Acceleration

Mubassher Ansari presents REQAP, a reliability-aware framework for weight packing and quantization in edge DNN accelerators. Quantized neural networks reduce memory and computation for edge deployment, but different bit-widths across layers can leave hardware registers underused and make quantized weights more vulnerable to bit-flip faults. REQAP explores how this unused register space can also be used for fault protection. The framework assigns bit-widths based on layer sensitivity and packs mixed-precision weight and activation pairs into fixed-width registers using the Safe-FFD algorithm. The remaining register space is then used for selective MSB replication with Triple Modular Redundancy, providing additional protection with almost no extra memory overhead. The approach is evaluated on AlexNet, VGG-11, and ResNet-18. The results show up to 56% fewer MAC operations and up to 62% memory savings, while also improving resilience under fault injection. The presentation also includes a bare-metal implementation validated on real FPGA hardware using a RISC-V core. This work was carried out by Mubassher Ansari as part of his Master’s thesis at the RSS Lab, under the supervision of Dr. Mahdi Taheri. The paper will be presented at the IEEE 20th Baltic Electronics Conference (BEC 2026) on 7 October 2026 in Tallinn, Estonia. 💻 GitHub Repository: https://github.com/Reliable-and-Secur... 📄 Paper: https://arxiv.org/abs/2609.17555 🔗 LinkedIn: https://lnkd.in/p/dtXzfgUj 🌐 RSS Lab: https://www.rss-lab.de