At the beginning of the Year of the Dragon, there was another blockbuster news in the AI field: OpenAI released the Wensheng ** large model SORA, and generative AI ushered in a new milestone.
From a technical point of view, the rate of evolution of SORA is almost incredible. Gen-2 released in June 2023 only supports 4 seconds of **generation and frame drops are obviously like slides, in November, Meta released **generation of large model emu video can generate 512*512, 16 frames per second**, 3 months later, SORA has been able to generate arbitrary resolution and aspect ratio**, and can also perform a series of image and **editing tasks, according to text prompts to create detailed**, generated by static images**.
The rapid development of the AGI industry requires a large number of model training and inference, which drives the continuous high demand for computing power. In practical applications, not all computing resources can be fully utilized, and a large amount of computing power is "idle" in the process of computing and data processing.
Ubiquitous computing power requires a stable network to connect various computing resources, and the high bandwidth, low latency, transmission stability, and reliability characteristics of open networks provide more application scenarios and possibilities for ubiquitous computing power. Domestic enterprises want to take advantage of AI technology to promote the development of digitalization and intelligence, but they don't know which vendor to choose to provide network services.
Xingrongyuan is committed to building an open network for ubiquitous computing power, covering cloud network, high-performance computing, artificial intelligence, enterprise data center, campus access and other fields, and supporting distributed storage, network visualization and other functions, greatly reducing costs while ensuring scale, bandwidth, latency and stability and other performance.
Taking SORA as an example, because SORA is trained based on "patch" rather than the whole **, similar to text labeling in large language models (LLMs), all types of visual data are converted into unified representations for large-scale generative training, which requires efficient processing of large amounts of data
Without affecting the data transmission performance, the network architecture is simplified, which greatly reduces the cost of user network construction.
The number of network path hops is reduced to 1 hop, which greatly reduces service latency.
Simplify the network structure and reduce the difficulty of O&M and troubleshooting.
In terms of network performance, AsterFusion's AI network solution has the following advantages:
1.The bandwidth of a single network has been improved.
1) Increase the number of network cards, consider CPU and GPU sharing when the initial business volume is small, prepare 1 to 2 separate network cards for the CPU in the later stage, and prepare 4 or 8 network cards for the GPU;
2) To improve the bandwidth of a single network card, it is necessary to match the PCLE bandwidth of the host and the bandwidth of the network switch, and Xingrongyuan 200G, 400G, and 800G Ethernet switches will cooperate with the network card to ensure high bandwidth for data transmission;
2.Application RDMA Network (ROCE).
1) Reduce the number of data replications in the process of GPU communication with the help of RDMA technology, optimize the communication path, and reduce the communication delay;
2) Easy ROCE delivers complex ROCE-related configurations (PFC, ECN, etc.) in one piece to help users reduce O&M complexity.
3.Reduce network congestion.
1) Reduce the latency on the network side and improve the efficiency of GPU usage: ultra-low latency is reduced to 400ns;
2) Reduce network congestion through DCB protocol groups: Build an all-Ethernet zero-packet loss and low-latency network through PFC, PFC Watchdog, and ECN.
3) Dual-network traffic distribution: CPU traffic is completely separated from GPU traffic to reduce the occupation and interference of different network traffic.
As a pioneer in the field of open network, Xingrongyuan continues to provide customers with products and solutions with superior performance and obvious cost advantages to help enterprises achieve more efficient operation and development. Relying on advanced technology and rich experience, Xingrongyuan will open up a broader space for the development of ubiquitous computing power and bring more opportunities and possibilities to the industry.
Follow VX's official account "Xingrongyuan Asterfusion" to get more technical sharing and the latest product trends.