Summary

  • Need for a reliable infrastructure to support large-scale personalized diet management services
  • Deployed Superb AI’s object recognition model on an NPU-based cloud service
  • Used the Rebellions SDK to run inference on pre‑trained models
  • Verified service quality and reduced TCO: performance per watt (FPS/W) improved by 2x, with accuracy on par with GPGPU

Challenge

Superb AI, an AI startup, partnered with a domestic healthcare platform that serves hundreds of thousands of monthly active users to implement AI features for intelligent, personalized diet management. They developed a “food recognition AI” that uses computer vision to identify foods in smartphone photos and analyze nutritional components. To commercialize this AI, they needed both high computer vision accuracy and cost‑efficient operations.

Food recognition requires high accuracy. The same dish can look different depending on how it is cooked or plated, and many foods look similar. Korean cuisine adds to the challenge because multiple side dishes are often served at once, so the detection scope must be both wide and precise.

Because a diet management service is a real-time service with a large user base, it requires cloud infrastructure that can deliver stable processing and scalability under heavy concurrent access. While the cloud enables real-time data processing and efficient operations, costs rise as throughput grows. It is important to design the setup carefully to ensure both accuracy and cost‑effectiveness.

Solution

Superb AI set medium‑ to long‑term goals to build a flexible model that makes it easy to add new food categories, improve performance through continuous learning, and maintain stable inference even during large spikes in concurrent traffic. To execute this roadmap with both performance and cost in mind, the company chose a Rebellions NPU‑based cloud service.

The NPU‑based cloud infrastructure delivered performance comparable to GPUs while consuming less power, resulting in high energy efficiency. Its cloud‑native architecture also enabled model optimization and seamless service scaling.

Rebellions’ software stack, the RBLN SDK, supports deep learning models trained with TensorFlow, PyTorch, and Hugging Face, allowing Superb AI to develop AI models and serve them on the infrastructure with minimal effort. Using the RBLN SDK, Superb AI could apply and deploy its large‑scale food object recognition model trained on extensive data as follows.

Result

By leveraging the benefits of NPU cloud infrastructure through this collaboration, Superb AI and the healthcare platform achieved the following outcomes:

  • Maintained the target accuracy (F1‑Score 0.87, baseline 0.7) without performance degradation
  • Demonstrated stable operation during continuous stress testing
  • Achieved performance per watt (FPS/W) up to more than twice as high as GPU‑based cloud services
  • Improved latency, which directly enhanced perceived speed for users

This adoption of NPU‑based cloud infrastructure confirmed that operating costs for the service can be effectively reduced. For real‑time services like AI‑based diet management, stable performance and reasonable operating costs open the door to expanding features and accommodating user growth in an economical way.

Appendix: NPU User Guide

Get Started Now: Prepare a pre‑trained model → Compile the model with the RBLN Compiler → Load the compiled model and run inference