【China Science Daily】On March 3, China Science Daily reported from the recently concluded FPGA 2025 — the premier international conference in the field of reconfigurable computing, that this year’s Best Paper Award was granted to FlightVGM, a video generation LLM inference IP jointly proposed by Infinigence, Shanghai Jiao Tong University, and Tsinghua University. According to the report, this marks the first time the FPGA conference has awarded its Best Paper distinction to a research work led entirely by a team from mainland China. It is also the first time an Asian research team has received this honor.
According to the team, this work marks the first efficient implementation of video generation models (VGMs) on field-programmable gate arrays (FPGAs). It also represents the latest milestone in a research series that began with their FPGA’24 work, FlightLLM, which accelerated large language models on FPGA platforms. Compared with an NVIDIA RTX 3090 GPU, FlightVGM achieves a 1.30× performance improvement and a 4.49× gain in energy efficiency on an AMD V80 FPGA, despite a peak compute gap of more than 21×.
China Science Daily noted that researchers from the Department of Electronic Engineering at Tsinghua University have previously had papers accepted at the FPGA International Conference in 2016 and 2017. The 2017 paper was selected as the sole Best Paper that year, although it was co-authored with an overseas team. In contrast, all authors of this year’s FPGA’25 Best Paper are based in mainland China, marking the first time the award has been presented to a single-country team from Asia.
In their paper, the authors highlight that amid ongoing industry debate surrounding the escalating cost of large-scale LLM deployment, improving hardware efficiency through flexible programmable architectures such as field-programmable gate arrays (FPGAs) and application-specific integrated circuits (ASICs) may represent a key pathway toward reducing inference costs and enabling scalable real-world adoption. In 2024, Infinigence’s custom inference IP for large language models, FlightLLM, was accepted at the FPGA conference. Building on this foundation, the team extended its work in 2025 with FlightVGM, a specialized inference IP for video generation models (VGMs), which was awarded the Best Paper Award. Both works explore how co-designing model inference pipelines with reconfigurable hardware architectures can substantially improve execution efficiency. These research outcomes have been incorporated into Infinigence’s proprietary large-model inference IP stack, the Large-model Processing Unit (LPU), and are currently being validated through collaborative deployments with industry partners.
The first author of the paper is Jun Liu, a PhD candidate at Shanghai Jiao Tong University. Co-first author Shulin Zeng is currently a postdoctoral researcher at Tsinghua University. The corresponding authors are Professor Yu Wang, Chair of the Department of Electronic Engineering at Tsinghua University and founder of Infinigence, and Associate Professor Dai Guohao of Shanghai Jiao Tong University, co-founder and Chief Scientist of Infinigence.
Paper link: https://dl.acm.org/doi/10.1145/3706628.3708864
