Alibaba Tuntang's official website has recently launched its high-end AI chip "Zhenwu 810E". This chip not only achieves full self-research in both software and hardware, but also marks the official formation of Alibaba's AI "Golden Triangle" strategy of "Tongyunge".
From the performance parameters, "Zhenwu 810E" shows strong performance. Its overall performance not only exceeds NVIDIA A800 and mainstream domestic GPUs, but also matches NVIDIA H20 in key indicators. The chip uses Tuntang's self-developed parallel computing architecture and ICN inter-chip connection technology, with an inter-chip bandwidth as high as 700GB/s. Currently, this chip has been deployed in a large-scale cluster on Alibaba Cloud and has successfully passed large-scale business verification.
Alibaba emphasized that "Zhenwu 810E" focuses on high usability and integrated chip-cloud. It is fully compatible with mainstream AI ecosystems, allowing developers to enjoy source code-level compilation capabilities and achieve seamless migration of algorithms. Combined with Alibaba Cloud's AI framework and model applications, this chip will provide enterprises with a full-stack integrated solution from computing power to underlying software.
Key Points:
🚀 Full Self-Research Breakthrough: Achieves full self-research in software and hardware, with a single card using HBM2e memory with a capacity of up to 96GB.
🔗 Strong Interconnection Architecture: Supports 7 independent ICN links, with an inter-chip connection bandwidth of 700GB/s, supporting linear acceleration for ultra-large scale clusters.
🏆 Competing with International Leaders: Comprehensive performance is comparable to NVIDIA H20, and has completed practical scale verification in Alibaba Cloud's thousand-card cluster.
