SD-VPU: Software Defined Video Processing Unit for Video Cloud

Operator video cloud applications (such as server room surveillance and video conferencing) consume vast video storage (thousands of PB), making efficient video compression an urgent necessity. However, existing compression solutions rely on GPUs, FPGAs, VPUs, but GPUs have high power and cost, and FPGAs and traditional ASIC VPUs are not userfriendly, failing to meet operators’ low-power, low-cost, easycustomization needs. This paper introduces a software-defined VPU (SD-VPU) architecture. Unlike the ASIC VPUs that only have video and AI hardware units, the SD-VPU additionally integrates high-performance general computing cores to support operating systems and realize easy-customization. We present the hardware and software design of the SD-VPU, the large-scale deployment scheme, and conduct a comprehensive evaluation in the existing applications of China Telecom. The results show our solution reduces power consumption by $70 \%$, cuts costs by $\mathbf{4 3. 8 \%}$, and improves the video compression performance by over $15 \%$ compared with the baseline. To the best of our knowledge, this is the first work to discuss video compression acceleration at scale in operator environments.