en.Wedoany.com Reported - Zhonghao Xinying recently officially launched its new generation of fully self-developed high-performance TPU AI dedicated computing chip "Xuyu", and simultaneously released the software-hardware integrated intelligent computing base "Taize 2.0 AI High-Performance Intelligent Computing Platform" equipped with this chip. The new product is a comprehensive upgrade based on the first-generation "Shana" chip and the Taize intelligent computing server, achieving improvements in dimensions such as underlying architecture, peak computing power, on-chip storage, cluster interconnection, and computing energy efficiency.

According to reports, a single "Xuyu" chip has been deeply optimized for the dedicated tensor computing logic of large models, with expanded registers and large-capacity on-chip cache. Its single-chip mixed-precision floating-point computing power reaches 896 TFLOPS, three times the performance of the previous-generation chip "Shana"; its 8-bit inference computing power reaches 1792 TOPS, suitable for high-concurrency inference scenarios with massive tokens. The rated power consumption of a single chip is 600W, reducing power consumption by 50% compared to traditional computing chips with equivalent computing performance. Leveraging multi-dimensional tensor computing units and data reuse optimization design, the comprehensive computing efficiency when executing the same AI tasks can be several times that of traditional GPU architectures.

The "Xuyu" chip continues Zhonghao Xinying's fully self-developed TPU technology roadmap, achieving complete independent research and development of chip IP cores, dedicated instruction sets, underlying operator acceleration libraries, and entire machine system software. Currently, Zhonghao Xinying's core technologies cover the entire chain of chip design, circuit development, compilation tools, and model adaptation. The simultaneously launched Taize 2.0 AI High-Performance Intelligent Computing Platform features a minimum computing unit composed of two high-performance CPUs and eight "Xuyu" TPUs working in coordination, with a whole-machine mixed-precision computing power of 7.168 P. When executing the same tasks, the whole-machine energy consumption is only 80% of that of traditional GPU servers.

As one of the earliest enterprises in China to invest in the research and development of TPU architecture AI dedicated computing chips, Zhonghao Xinying successfully taped out China's first high-performance TPU chip "Shana" in 2023 and achieved mass production. The company's products have been deployed in multiple ultra-large-scale intelligent computing centers built by Shenzhen Unicom, Tianjin Mobile, Taiji Computer, and others, with application fields covering industries such as finance, media, education, and healthcare.