Content

Intended customers: all users. New features/specifications: EFC client supports standalone and distributed read caches on the computing side (disk and memory) of CPFS, NAS, and OSS. Scenarios: 1. Inference scenarios: During model loading, model files need to be quickly loaded to each GPU server. Customers expect to control the model loading time within 1 minute. The greater the number of GPUs, the higher the performance requirements of the storage system. Based on the GPU local disk and memory, the distributed cache capability allows you to decouple the performance of the storage system and the model loading speed to achieve fast model loading. 2. Training scenarios: In training scenarios, the training data is read repeatedly. The cache on the computing side enhances the reading performance during the training process, which can greatly improve the GPU utilization and improve the overall training speed.