China's Zhonghao Xinying unveils self-developed TPU chip 'Xuyu' with 896 TFLOPS power

Chinese chip company Zhonghao Xinying launched a new fully self-developed TPU AI chip called 'Xuyu'. It delivers 896 TFLOPS of mixed-precision floating-point power per chip, three times more than the previous generation. The chip also offers 1792 TOPS for 8-bit inference, making it suitable for high-concurrency tasks. It supports long context with improved memory and interconnect speed, while consuming only 600W per card—50% less than traditional chips.

Alongside, the 'Taize 2.0' computing platform was introduced, featuring two CPUs and eight TPUs, with a total power of 7.168P. Its energy use is 20% lower than traditional GPU servers. Both the chip and platform are fully self-developed, with no reliance on foreign core technologies, meeting security and compliance needs for government and finance sectors.

The platform supports mainstream AI frameworks like PyTorch and vLLM, and is compatible with dozens of large models such as Qwen, DeepSeek, and GLM, allowing developers to quickly migrate their models.