Alibaba's PPU Zhenwu 810E: How the Chinese AI Accelerator Handles LLMs

In a new article, Selectel MLOps engineer Dmitry provides a detailed review and testing of the Chinese AI accelerator Zhenwu 810E from Alibaba Group. Amid the rapid growth of the specialized computing hardware market, the author aims to evaluate the chip's real-world performance when handling large language models (LLMs). The focus is on the accelerator's applicability in standard inference scenarios, such as its use as a node for cloud Kubernetes or as part of a dedicated server. The article details the testing process, the device's technical specifications, and its potential for integration into modern cloud infrastructure. This material will be useful for professionals interested in alternative solutions for machine learning tasks and the development of the Chinese semiconductor market.