MRT为企业互联网+ 赋能

According to news on August 9th, the AI application "Miaoyang Camera" quickly became popular less than a month after its launch due to its high-quality "photo taking" photos and affordable prices, attracting a large number of users. Due to the surge in user numbers, the demand for computing power has also sharply increased, leading to queuing phenomena. To address this issue, the application was expanded to Alibaba Cloud, successfully alleviating the queuing and waiting situation, significantly reducing waiting time, and improving user experience. At the same time, Miaoya Camera has also launched an app version of the Apple App Store and further enhanced its functionality.

Image source: Miao Ya Official

Since its launch on July 17th, the "Miao Duck Camera" has attracted thousands of users with a affordable price of 9.9 yuan and high-quality photos. During peak hours, up to 4000 to 5000 people queue up to use the app and need to wait for more than 10 hours to generate photos. According to the relevant person in charge, although the application has undergone emergency expansion on Alibaba Cloud to cope with the soaring demand for computing power hundreds of times. At the same time, users are also constantly increasing, but the phenomenon of waiting in line has been alleviated.

A senior IT professional stated that the photo generation of the Miao Ya camera requires a huge amount of GPU computing power for calculation. Based on peak data, it is estimated that at least a few thousand GPU servers are needed to handle such a large-scale task. Due to the impracticality of deploying servers of this scale offline, choosing to deploy on the cloud not only saves costs and time for startups, but also better meets user needs. However, currently only a few cloud computing giants have such computing power reserves.

This senior person also stated that previously, there was a lot of attention paid to the computational power required for large model training, especially for GPT model training, which often required tens of thousands of graphics cards. However, with the popularity of open-source large models such as Llama2, the computational power requirements of the inference end of large model applications are more worthy of attention. He believes that when AIGC enters the era of productivity, the demand for inference computing power may be much higher than that for training computing power.

Image source: Miao Ya Official

As various industries gradually complete the development of AIGC, the demand for reasoning and computing power required during the business development phase begins to rapidly increase. In the coming months, the total demand for inference computing power may far exceed that for training computing power.

Taking Alibaba Cloud, the largest in China, as an example, it provides a cluster that can scale up to 100000 GPU cards and can simultaneously support multiple large-scale models with trillions of parameter sizes for online training and inference. In addition, Alibaba Cloud also provides the AI acceleration kit AIACC, which can significantly improve computing efficiency. For fine-tune training of large models, the efficiency can be improved by over 60%; The inference efficiency can be improved by more than 100%.

In summary, the "Miaoya Camera" has recently received widespread attention and has successfully alleviated queuing and waiting situations by expanding its application to Alibaba Cloud, providing users with efficient and stable services. In the future, with the completion of AIGC development, the demand for inference computing power may far exceed the demand for training computing power. The computing power reserves and technical support of a few large companies such as Alibaba Cloud will undoubtedly become an important support for startups to compete in the AIGC field.