Sora Goes Global Viral

Who doesn't know about this by now?

On February 16th, OpenAI released a video generated by the video giant model Sora, and its lifelike visual effects caused it to go viral overnight.

This sudden surge in interest around Sora has set off waves of excitement among tech professionals, investors, and the film industry, even sparking anxiety about "AI replacing humans." This unexpected development prompted companies like Adobe and domestic tech giants (Baidu, Alibaba, Tencent) to swiftly establish their AI video research teams, reigniting discussions everywhere about the potential for AI to supersede human capabilities.

Currently, Sora is still in the GPT-3 stage, with some successful attempts and some failures evident from the released videos. However, the immersive and highly realistic visuals have left a lasting impression, sometimes blurring the lines between reality and AI-generated content.

▲ Hyper-realistic effects

Unlike other AI-generated videos, Sora's groundbreaking significance lies in three aspects: Firstly, it can produce "60-second ultra-long videos," whereas previous AI-generated videos were limited to a maximum of 4 seconds. Secondly, the state of objects depicted in the videos faithfully replicates real-world light and shadow changes, with seamless camera transitions. Thirdly, the synthesized content by Sora adheres to the physical laws of the real world, with moving objects following fluid dynamics principles.

So, what's the technology behind it?

In simple terms, Sora is a diffusion model trained on videos and images of varying durations, resolutions, and aspect ratios. It employs an advanced diffusion model + Transformer architecture to meaningful combinations of contextually relevant units within the synthesized content. This is akin to using various roots and affixes to randomly arrange and combine words to form a sentence. Similarly, Sora uses batches of visual blocks to meaningfully combine into an image in spatial layout dimensions, which are then seamlessly pieced together into a video.

Sora's popularity has firmly shifted the AI spotlight from text to video, and it's anticipated that its model parameters will expand further, accompanied by a significant increase in computational power demands. Automated video generation—whether it's about nuanced visual textures, smooth frame rates, or precise effects processing—poses higher requirements for AI servers, GPU cards, and other computational equipment. Gooxi, as a leading server solutions provider in China, has introduced AI inference servers and large-scale model training services to meet the demands of visual model training, providing powerful storage and computing capabilities for rapidly generating high-quality video content.

In terms of computational power, the Gooxi Eagle Stream platform's 4U10-GPU AI server is built on the Intel fourth and fifth-generation Xeon® Scalable processor platforms, employing the latest DDR5 and PCIe 5.0 technologies to significantly accelerate data transfer rates. It supports 10 dual-width GPU cards from domestic and international brands, with a maximum of 19 PCIe expansion slots, offering abundant PCIe expansion capabilities. Additionally, the server system features redundant design for critical components and modular design for functional parts, effectively enhancing product reliability and maintainability. Its efficient and stable data processing and computing capabilities can meet the high-performance and high-scalability requirements of various video generation businesses.

▲ Gooxi Eagle Stream Platform 4U10-GPU AI Server

On the platform side, Gooxi's Bai Ze Intelligent Computing Fusion Platform is a self-developed computational power scheduling platform tailored for artificial intelligence (AI) scenarios. It adopts advanced technical architecture, supports mainstream X86 servers, configures NVIDIA GPUs for high-performance accelerated computing, and utilizes mainstream distributed storage devices, supporting TCP/IP, InfiniBand high-speed network interconnection. It enables efficient management, scheduling, and monitoring of heterogeneous resources; moreover, the platform comes pre-installed with images such as TensorFlow, PyTorch, MxNet, CUDA, Gromacs, NAMD, LAMPPS, OpenVINO, oneAPI, and can connect to NGC, Docker HUB, and other image sources, allowing users to connect and configure image environments via Docker Exec, supporting various methods such as requirements.txt, pip, conda for installing and configuring python packages to meet the requirements of different backgrounds and levels of users for image creation.

▲ Gooxi Bai Ze Intelligent Computing Solution

Furthermore, a major advantage of the Bai Ze platform is its ability to be locally deployed, improving the efficiency of large model development, aiding enterprises in building full-stack development capabilities, and effectively promoting the industrialization of AI. Besides offering complete functionalities from AI model development, training to deployment, Bai Ze is also suitable for high-performance computing (HPC) scientific computing fields such as life sciences, meteorological forecasting, quantum mechanics, and oil and gas exploration.

The release of the video giant model Sora signifies a new milestone in video generation technology. Gooxi will continue to advance hand in hand with its ecosystem partners, collectively driving the AI big model industry forward.

Image sources: Visual China Group, Sora Official Video Screenshots