**Breakthrough for Video-Generation AI Model Using Off-the-Shelf Chip Signifies a Milestone for China and Changes the Global AI Hardware Landscape** **Reading Time: 2 minutes** **Why You Can Trust SCMP** *Zhang Tongin, Beijing* **Published: 7:00 PM, March 12, 2025** In an ambitious effort to challenge the leadership of US giant Nvidia in AI hardware, Chinese researchers have successfully trained an advanced video-generation model using an off-the-shelf industrial chip, demonstrating superior speed and efficiency compared to high-end GPUs. The system, named FlightVGM, achieved a remarkable 30 percent performance increase and displayed energy efficiency that was 4. 5 times greater than Nvidia’s premier RTX 3090 GPU, all while operating on the widely accessible V80 FPGA chip from Advanced Micro Devices (AMD), another major US semiconductor company. This groundbreaking innovation garnered the prestigious Best Paper Award at the FPGA 2025 conference, which concluded on March 1. This marked a historic first for a mainland Chinese team, indicating a significant shift in the global quest to optimize AI hardware. FlightVGM was developed by a collaboration of scientists from Shanghai Jiao Tong University, Tsinghua University, and the Beijing-based startup Infinigence-AI. The model has the potential to transform how industries implement cost-effective and energy-efficient AI systems across various applications, including robotics and autonomous vehicles. Field-programmable gate arrays (FPGAs) are programmable semiconductor devices that enable modifications to their circuitry and functions after manufacturing. In contrast, traditional chips like CPUs (central processing units), GPUs (graphics processing units), and ASICs (application-specific integrated circuits) have fixed functionalities once produced. In the realms of video generation and general computing, both FPGAs and Nvidia GPUs offer unique advantages.

FPGAs provide customizable architectures tailored to specific uses, leading to greater energy efficiency and reduced latency. Conversely, Nvidia GPUs are renowned for their substantial parallel processing power, making them adept at managing large data sets and executing complex computational tasks. Building on earlier research, the Chinese team developed FlightVGM as the first video-generation AI model trained on FPGAs. Through innovations in data architecture and scheduling techniques, FlightVGM achieved computational performance that outstrips that of GPUs.