according toChina UnicomOfficially, recently, China Unicom Research Institute and Zhejiang Unicom, Unicom clothing manufacturing legion collaborated on research and put forward an innovative business model for the local storage of AI-sensitive data off-site training requirements, and successfully implemented between Hangzhou and JinhuaThe industry's first 30TB sample data across 200 kilometers of storage and computation separation far training, after the actual calculation, the training efficiency is as high as 97% or more..
1AI learned from the official introduction that the test fully verified the security, feasibility and efficiency of the storage and computation separation technology, providing new ideas and directions for the development of future AI technology.
Separation of storage and computation technology refers to the separation of the warehouse that stores the data and the processing plant that computes the data.The data is pulled directly from the remote storage device for computation during training, instead of being stored to the local disk before processing, which can effectively ensure the security and consistency of user data.
According to China Unicom, there are two major challenges in the process of processing massive sample data:firstlyData is mostly stored on the enterprise side, and some data with high security requirements are not convenient to move out;secondlyAs the volume of sample data surges, AI computing centers need to be equipped with additional storage resources while possessing powerful computing power, which significantly increases the construction cost. In this context, the industry has an urgent need to realize the "separation of storage and computation, and sample training and pulling".
The key features of this storage and calculation pull-away test validation include:
Firstly, we have innovatively reconstructed the smart computing training model with cross-location AI large model training capability.The traditional centralized training mode of Smart Computing requires users to upload samples to the Smart Computing Center for drop disk training, but some users have security concerns about the drop disk of private samples. Zhejiang Unicom has realized the "data without dropping disk" remote training of Hangzhou storage and Jinhua training through IP wide-area non-destructive program, and explored a new way for enterprise users to train their privacy samples with the ability of computing network synergy.
Second, the total amount of sample data reaches 30TB, the transmission distance is over 200 kilometers, and the computational pull-out efficiency is greater than 97%.Through the AI training storage-calculation separation test of the "Clothes Pupil Industry Model" of Unicom's clothing manufacturing army, the technical feasibility of storage-calculation remote training for AI training business has been fully verified. It fully verified the technical feasibility of the storage-calculation pull-away for AI training business, and in the future, users with relevant data-sensitive business needs can use the operator's arithmetic services to complete the pull-away training of private samples without leaving the campus, realizing the best balance between cost and security.