[News] China’s Zhonghao Xinying Claims TPU 1.5× Faster than NVIDIA A100 Amid Google TPU Buzz

With Google reportedly pursuing TPU deals with Meta and pressure mounting on NVIDIA’s dominance, a rising Chinese TPU startup is drawing attention. According to South China Morning Post, Zhonghao Xinying has positioned itself as a domestic alternative to NVIDIA, saying its self-developed general-purpose tensor processing unit (GPTPU) entered mass production as early as 2023.

The company is also advancing its plans to go public. According to the report, Zhonghao Xinying signed a performance-guarantee agreement with investors requiring it to complete an IPO by the end of 2026, or face a share buy-back obligation.

Fully Self-Developed TPU Architecture and Scalable Compute Infrastructure

As the report highlights, the company’s flagship chip, Chana, is claimed to deliver up to 1.5 times the compute performance of NVIDIA’s A100 tensor core GPU, while cutting energy consumption by 30% for comparable large-model workloads and reducing per-unit compute costs to 42% of NVIDIA’s.

The company’s TPU incorporates fully self-developed IP cores, a custom instruction set, and a wholly in-house compute stack that does not rely on foreign technology licenses, as the report highlights, citing founder Yanggong Yifan. EastMoney also notes that it has already reached mass production on a 12nm process.

In addition to its TPU products, South China Morning Post notes that the company has launched Taize, a large-scale computing cluster that links 1,024 Chana processors, designed to support training for trillion-parameter foundation models. The report further states that Yanggong told an industry conference in June that a “next-generation TPU” is in development, though no timeline was provided.

Leadership with Deep Google TPU Experience

Notably, the company’s leadership team brings extensive experience from major global tech firms. Zhonghao Xinying was founded in 2018 by Yanggong Yifan, who previously worked on chip architectures at Google and Oracle. According to South China Morning Post, he contributed to the development of Google’s TPU v2, v3, and v4. The CTO and co-founder, Zheng Hanxun, also has a background in chip design, having held roles at Oracle and at Samsung Electronics’ R&D center in Austin, Texas.

Read more

• [News] China’s GPU Trio Rise as NVIDIA Retreats — Decoding Moore Threads, MetaX, and Cambricon
• [News] China’s largest STAR Market IPO of 2025, Moore Threads, Goes Public on Nov 24, Raising RMB 8B

(Photo credit: Zhonghao Xinying Wechat account)