Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 15·5 min read
NAS-Driven Hardware Accelerator Exploration for Edge AI and Quantization Effects on the Pareto Space
TL;DR
A novel three-stage pipeline for hardware-aware Neural Architecture Search (NAS) is proposed, addressing the impact of Post-Training Quantization (PTQ) on architecture performance. The study reveals that an FP32 zero-shot surrogate outperforms an INT4-trained surrogate in exploring the Pareto space of architectures.
✦ Why It Matters
Consider adopting the proposed pipeline to improve your edge AI architecture design and quantization strategies.
Key Takeaways
- Implement the three-stage pipeline for efficient architecture exploration.
- Utilize FP32 zero-shot surrogates for better Pareto space coverage.
- Analyze the effects of quantization on architecture performance.
- Explore evolutionary Domain Space Exploration for optimal hardware mapping.
How It Works
The proposed pipeline consists of three stages: a hardware-agnostic Pareto rank surrogate that evaluates architectures, a quantization bridge that filters and controls feedback based on Pareto efficiency, and an evolutionary Domain Space Exploration (DSE) backend that maps architectures to hardware efficiently. This structured approach allows for a comprehensive analysis of how quantization impacts architecture performance.