Energy-Aware Physical Synthesis of Deep Neural Networks for Edge-AI Applications in Robotics and VLSI Systems
Energy-aware physical synthesis is essential for deploying deep neural networks (DNNs) in edge-AI applications for robotics and VLSI systems, where stringent power, area, and latency constraints prevail. Beyond algorithm-and software-level optimizations such as network compression and compiler techn...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | , , , , , |
| Format: | article |
| Published: |
2025
|
| Online Access: | http://hdl.handle.net/11675/14472 https: |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Energy-aware physical synthesis is essential for deploying deep neural networks (DNNs) in edge-AI applications for robotics and VLSI systems, where stringent power, area, and latency constraints prevail. Beyond algorithm-and software-level optimizations such as network compression and compiler techniques, the physical effects during VLSI imple-mentation including clock tree synthesis (CTS), routing congestion, IR-drop, cell sizing, and placement significantly influence energy consumption and timing closure. This work presents a unified framework that co-optimizes DNN architectural features (including quantization, sparsity, and operator tiling) with physical design choices, such as multi-Vt selection, sizing, placement strategies, activity-driven buffering, and CTS. The proposed methodology formulates a multi-objective optimization targeting minimum energy at iso-throughput, meeting timing and area constraints via analytic models and sign-off cali-brated power estimates. A co-design loop iteratively reshapes the network and refines physical synthesis using sensitivity to activity factors and critical-path slack. The framework is validated with prototype RTL for representative CNN/transformer blocks, imple-mented on open PDKs and evaluated on FPGA/SoC testbeds for mobile robotics scenarios. Results demonstrate 28–41% reduction in energy per inference under constant accuracy and throughput, 12–18% lower leakage with multi-Vt and sizing, and a 1.6× improvement in worst-negative-slack closure probability across 500–800 MHz operation. Ablation stud-ies clarify the impact of quantization-aware placement and activity-weighted CTS. The framework integrates seamlessly with standard EDA flows and IEEE design rules, enabling automated hardware-software co-optimization for practical edge-AI deployments. |
|---|