Small VLM Quantization Explored for Edge Deployment on NVIDIA Jetson

This paper investigates the quantization of small vision-language models (VLMs) for efficient deployment on edge devices, specifically the NVIDIA Jetson Orin NX and AGX. The research systematically evaluates five hypotheses across six quantization configurations, revealing that model architecture (MoE vs. dense) significantly impacts quantization sensitivity, with MoE models better handling INT4 noise. The study also found that SigLIP encoders introduce latency on Jetson Ampere hardware due to specific kernel-hardware interactions, and while INT4 quantization reduces VRAM, it can slow token generation. Composite quantization errors are generally additive, except for modality-alignment paths, and the intelligence-per-joule profile is highly platform-dependent due to memory bandwidth. AI

IMPACT Provides insights into optimizing VLM performance and efficiency on resource-constrained edge devices.

RANK_REASON Academic paper detailing a systematic evaluation of VLM quantization techniques for edge deployment. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Hugging Face
• Innu-aimun
• Int4
• Int8
• Jetson Ampere
• NVIDIA Jetson AGX Orin 64GB
• NVIDIA Jetson Orin NX 16GB
• SigLIP

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →