When Smartglasses Meet AI: A First Performance Look

Abstract

Smartglasses are emerging as hands-free and low-friction platforms for wearable voice AI assistants, but their end-to-end performance and system characteristics remain poorly understood. This paper presents the first empirical measurement study of five consumer-grade AI-enabled smartglasses—Ray-Ban Meta (RBM), Ray-Ban Meta Display (RBM DP), Demabon, EvenG1, and Halliday—and compares them with three phone-based counterparts: Meta AI, OpenAI ChatGPT, and Microsoft Copilot. We find that smart-glasses consistently lag behind phones in responsiveness: their mean time-to-first-action (TTFA) spans 2.85–9.13 s, versus 1.79–2.79 s for phones. However, this gap is not a fixed wearable penalty, as smartglasses themselves exhibit large internal variation, from near phone-class TTFA at 2.85 s to slow-tier TTFA above 8 s. TTFA breakdown further shows that different devices are limited by different pipeline designs. For example, Demabon is server-bound, with 8.43 s server-side latency, whereas EvenG1 has a fast backend but is client-bound, with 4.28 s client-side latency. Under uplink packet loss, robustness depends not only on the transport protocol but also on query-submission granularity: QUIC-based RBM/RBM DP remain resilient, TCP-based EvenG1/Demabon degrade gracefully, while Halliday's single long-lived POST leads to cliff-like failure. Overall, our results call for smartglasses AI pipelines that co-design cloud services, cross-stage coordination, and robust transport.