At the World Artificial Intelligence Conference 2026 (WAIC 2026), Dr. Huang Xin, VPU R&D Director at Arm China, delivered a keynote speech titled "AI VPU: Opening a New Chapter in Video Compression" and unveiled "Wudang," the next-generation "Linglong" Video Processing Unit (VPU) IP.
Image | Dr. Huang Xin, VPU R&D Director at Arm China; Source: Photography by eefocus
From this speech, we see that the "consumers" of video are shifting from humans to AI, and the logic of video compression must also transition completely from "human-eye friendly" to "machine-friendly".
From "For Human Viewing" to "For Machine Viewing"
Dr. Huang mentioned that over the past two decades, the evolution of video codec technologies (such as H.264 and H.265) has always revolved around a core objective: serving the subjective visual experience of the human eye. The goal is to eliminate artifacts perceptible to the human eye as much as possible within limited bandwidth and storage space.
However, given the explosive growth of technologies such as humanoid robots, autonomous driving, and smart security, AI has gradually become the largest consumer of video data. Data shows that over 80% of the data acquired by AI comes from video, making video a critical bridge connecting the physical and virtual worlds.
This brings entirely new challenges: the tolerance for video quality differs completely between the human eye and machines.
First is the issue with traditional compression, where minor image degradation that is "imperceptible" to the human eye for the sake of saving bandwidth can be fatal in the eyes of a machine. This can directly lead to the collapse of AI inference and training accuracy (for example, an autonomous driving system might misidentify a pedestrian as background as a result).
Meanwhile, to reduce costs and increase efficiency, the design philosophy of the VPU (Video Processing Unit) must be disrupted, shifting from merely satisfying subjective vision to comprehensively improving the recognition accuracy of machine vision.
From "Passive Execution" to "Content Awareness"
Facing this industry shift, Arm China's self-developed "Linglong" VPU IP has provided a clear technology evolution roadmap:
Image | "Linglong" Coding Technology Evolution Roadmap; Source: Photography by eefocus
- CAE Technology with Embedded AI Operators
In past products, the VPU was passive and entirely controlled by the CPU. However, in the current-generation "Emei" and next-generation "Wudang" VPUs, lightweight AI operators are integrated internally, equipping them with Content-Aware Encoding (CAE) technology.
Dr. Huang pointed out that CAE is at the first level of the video encoding unit. While moving data, it can process two core actions in parallel:
- First layer (shallow): Pixel-level image preprocessing.
- Second layer (deep): In-depth image semantic parsing.
The fusion of these two layers of data jointly guides the next-level encoding decisions, thereby achieving extremely high bandwidth savings in different scenarios: 20% savings in aerial photography scenarios, 10% in surveillance scenarios, and over 5% in gaming scenarios.
- Three-Dimensional Leap of the "Wudang" VPU
As the latest masterpiece following "Emei," the "Wudang" VPU, which will be released early next year, builds upon the inherited CAE technology to achieve three major performance leaps:
- PPA (Performance/Power/Area) level: Single-core performance breaks through 8K30fps+@1GHz, and the chip area is significantly reduced by 40% compared to the previous generation. In typical AI scenarios such as irregular motion, encoding quality is additionally improved by 5%~10%.
- Specification expansion: Supports YUV422 encoding to meet professional shooting needs; specifically supports direct Raw Data compression, eliminating the precision loss of gamut conversion and directly feeding high-precision data for AI training.
- Embodied intelligence adaptation: Supports ultra-high bitrates of up to 1Gbps+, fully serving complex perception scenarios for Physical AI (Embodied Intelligence) such as humanoid robots.
Image | Unveiling the "Linglong" VPU "Wudang"; Source: eefocus
End-to-End Joint Optimization and "Task Awareness"
To break the bottleneck of "local optimization," Arm China is implementing "AI end-to-end video coding technology." Dr. Huang revealed that the core of this framework lies in "end-to-end joint optimization" and "AI task-aware compression."
First is backpropagation connecting the entire pipeline. During training, the front-end network, encoder, and downstream specific AI applications are linked, and joint training is conducted through backpropagation.
Second is flexible bitrate configuration, allowing the compression process to proactively "sense" the needs of downstream AI tasks, intelligently allocating precious bitrates to critical information. This significantly improves the compression rate while maximizing the task accuracy of machine AI.
Image | Practice of AI VPU in Specific Video Tasks; Source: Photography by eefocus
In specific video task practices, this framework has demonstrated astonishing data returns: it can save 18.3% of the bitrate in the Ostrack object tracking task; it can save 12.0% of the bitrate in the SlowFast action recognition task, and the encoding quality for both is improved by up to over 10%.
Therefore, although this optimization appears to have little difference from traditional algorithms to the human eye, for machines, its recognizability and friendliness have undergone a qualitative improvement.
Final Thoughts
"Enabling AI to understand the visual world," Arm China is extending the reach of its self-developed IP from traditional multimedia codecs to edge AI deployment and training by integrating AI operators down to the VPU hardware layer, thereby consolidating its influence in the edge heterogeneous computing ecosystem.
For downstream chip design engineers, how to perform system-level optimization under constrained power budgets is the key to whether this VPU can be effectively utilized.