National Formosa University: The Compute Engine Powering Taiwan’s National Drone Team

Aerospace heritage × AI cloud computing platform — accelerating drone R&D, teaching transformation, and industry–academia co-creation

I.  Project Background

As the war in Ukraine reshaped the nature of modern warfare, drones have become a strategic industry that nations are competing to master. Taiwan’s Executive Yuan has announced an investment of US$1.49 billion over six years, beginning in 2025, to build a “national drone team.” Drawing on its deep aerospace heritage, National Formosa University (NFU) has been executing a Ministry of Education program since 2022, establishing its Advanced UAV Manufacturing and Inspection Laboratory. With more than two decades of work in the drone field, the university also leads the operation of the Asia UAV AI Innovation Application R&D Center, which has attracted flagship companies such as Thunder Tiger into joint development — including the Made-in-Taiwan drone built on morphing-wing structural technology.

Forward-looking projects such as intelligent swarm flight and real-time autonomous navigation, however, involve large-scale virtual environment generation and enormous volumes of sensor data — 4K imagery, LiDAR point clouds — all of which demand exceptional AI computing power. Under the traditional physical development cycle of design, manufacture, and flight test, a single failure sends the team back to the drawing board, making schedules long and costs high. NFU therefore set out to build an integrated AI compute cloud platform, using digital twin technology to substantially shorten development and validation cycles. The goal was to complete the final piece of its R&D capability while opening AI capacity to faculty and students across the entire campus — transforming teaching and driving intelligent upgrading throughout the Yunlin–Chiayi–Tainan regional industrial base.

II.  Challenges Before Adoption

• Siloed resources and uneven allocation.  Individual departments purchased their own GPUs, resulting in a persistent mismatch: busy laboratories ran short of compute resources while equipment in less-active units sat idle. With no mechanism for sharing across colleges and departments, investment was wasted.
• Time-consuming setup and a high maintenance barrier.  Different AI models require different frameworks and CUDA versions. Faculty and students routinely spent days or even weeks configuring environments, and the maintenance burden fell on professors and graduate students — with no assurance of stability or information security.
• Limited flexibility for teaching and student projects.  It was difficult to provide dozens of students with identically specified computing environments at the same time. High-end resources were reserved for core projects, leaving general faculty and students facing either no available GPU or excessive queuing.
• Compute specifications unable to support high-complexity models.  Conventional workstations could not handle training on massive datasets such as 4K imagery and LiDAR point clouds, nor meet the low-latency digital twin computing requirements of platforms such as NVIDIA Omniverse.

III.  The Solution: Deploying the AI-Stack Infrastructure Management Platform

NFU’s Computer Center introduced AI-Stack to virtualize and centrally manage the underlying GPU, CPU, and memory resources, establishing a complete intelligent compute governance architecture spanning resource scheduling, container environments, user permissions, and system operations.

3.1  Resource Scheduling

• GPU partitioning allows multiple users and multiple tasks to share a single card. Well suited to development, testing, and lightweight inference, it markedly improves hardware utilization and lowers the barrier to entry for equipment.
• Multi-GPU and cross-node aggregation supports large language model fine-tuning and high-performance training. Combined with dynamic quota management, compute can be dispatched flexibly on demand and idle capacity reclaimed automatically.

3.2  Container Environment Management

• Support for open-source ecosystems including PyTorch, JupyterHub, Hugging Face, and Ollama. Users select from a web interface and deploy an environment with the matching CUDA version within seconds — an open-source AI container in under a minute, accelerating RAG system development.
• Support for custom image uploads, allowing drone R&D teams to replicate specialized environments containing NVIDIA Omniverse, LiDAR point cloud processing, and similar tooling.

3.3  User and Permission Management

• Integration with the campus LDAP/AD account system means faculty and students sign in with their existing campus credentials rather than memorizing multiple passwords, achieving single sign-on (SSO).
• Group quotas and approval workflows are established by department and project, with compute resource isolation ensuring that industry–academia collaboration data remains protected.

3.4  System Operations

• Course containers are created in batches and automatically reclaimed when a course ends or after an idle timeout, preventing resources from being held indefinitely.
• A graphical dashboard monitors server and GPU utilization, temperature, and health status in real time.
• Complete usage histories are recorded for every account and project, providing the basis for cost allocation and performance reporting across inter-institutional collaboration, government programs, and industry–academia partnerships.

IV.  Results

• Improved management efficiency.  The Computer Center shifted from reactive repair to proactive management. A single back-end console now governs campus-wide compute, and maintenance costs have fallen by more than 50%.
• Maximized resource utilization.  GPU partitioning redistributes previously idle fragments of compute to multiple students for lightweight testing, significantly improving return on hardware investment.
• Substantially improved user experience.  Final-project environments that once took days to prepare are now ready within minutes, greatly shortening project development cycles.
• Upgraded teaching and talent development.  AI has been embedded deeply into teaching across aerospace, electrical engineering, smart manufacturing, and other departments, cultivating cross-disciplinary talent with genuine hands-on capability — graduates who are job-ready on day one and are highly regarded by industry.

V.  Looking Ahead: Broader AI Applications and Regional Industry–Academia Empowerment

• Cross-disciplinary talent development.  A Cross-Disciplinary Drone Program covering AI system modeling, intelligent autonomous flight, and smart manufacturing applications, enabling faculty and students without a programming background to launch a standardized development environment with a single click.
• Campus AI assistants in production.  Representative applications include the OriStork Calculus Teaching Assistant, which uses Socratic questioning to guide students’ reasoning rather than supplying answers outright; and the Smart Campus AI Administrative Assistant, which digitizes regulations, course selection rules, and travel policies and applies an LLM to compress what was once 15–20 minutes of manual cross-checking into precise retrieval and summarization in 5–10 seconds.
• On-premises deployment and information security autonomy.  All model inference runs on the university’s core servers. Sensitive data never leaves campus, balancing information security with compute sovereignty while avoiding the substantial commercial API fees otherwise paid to overseas technology providers.
• Driving regional industry.  The AI compute cloud platform will continue to expand capacity in partnership with Dell Technologies, with the goal of becoming a regional AI computing center serving the Yunlin–Chiayi–Tainan industrial cluster and driving intelligent upgrading across central Taiwan’s precision machinery and aerospace industries.

VI.  Customer Testimonial