Cut GPU Costs by 75%: AI‑Driven Car Fault Detection with UCloud Hot‑Standby

Many drivers struggle to interpret the myriad warning lights on a car’s dashboard. The WeiChe (Microcar) app offers a solution by using AI to scan a dashboard photo with a phone and instantly identify each fault indicator and its recommended remedy.

Built on data from 130 million registered vehicles, WeiChe has accumulated extensive expert knowledge. It applies deep‑learning models to classify and label every object in a fault‑light image, enabling accurate recognition of diverse warning symbols.

The service runs on UCloud’s UAI‑Inference online inference platform, which supplies scalable GPU resources and frees developers from building and maintaining underlying infrastructure.

A key challenge is the tension between the need for real‑time response and the high cost of dedicated GPU compute. Traffic patterns show sharp peaks during short periods and long idle stretches, leading to wasted GPU capacity.

Cost comparison highlights the trade‑off: CPU inference incurs ~20 seconds latency at 0.32 RMB/h, while GPU inference delivers ~0.5 seconds latency at 5.1 RMB/h. For deep‑learning workloads, GPU performance is essential despite the higher price.

UCloud’s Hot‑Standby feature resolves this dilemma. When a GPU‑exclusive service receives no requests for 30 minutes, the platform automatically migrates the workload to a lower‑cost GPU standby pool, preserving GPU capability while reducing expenses. As soon as traffic resumes, the service instantly returns to exclusive GPU mode.

Example calculation: without Hot‑Standby, a single GPU node costs 5.1 RMB/h → 3,672 RMB per month. With Hot‑Standby, the same node costs 0.99 RMB/h for 22.5 hours of idle time plus 5.1 RMB/h for 1.5 hours of peak usage, totaling 898 RMB per month—a 75 % saving.

To enable Hot‑Standby, users must first ensure the service type is GPU‑exclusive (not elastic mode). Then, in the auto‑scaling management console, activate Hot‑Standby and configure trigger rules based on QPS (e.g., switch after 30 minutes of zero QPS). Minimum node settings can control how many standby nodes are kept.

This feature is ideal for workloads that require high‑performance single‑node GPU compute but experience clear idle periods.

Hot‑Standby is currently available in the Beijing‑2 and Shanghai‑2 regions.