Tianshu Zhixin Launches Tianai 300 GPU, Challenges Hopper Performance

At the 2026 World AI Conference, Tianshu Zhixin unveiled its new flagship GPU, the Tianai 300. It uses a SIMT architecture and supports scalar, vector, and tensor computing, with optimizations for Attention and MoE. The chip achieves over 90% Attention efficiency, and is 10% better than Hopper in 64k context scenarios. MoE performance is also 10% higher, while first-token latency drops by 20% and average communication latency by 13%. Decode efficiency is 10% better than Hopper. Tianai 300 is ready for large-scale use and has been adapted for cloud and server deployments.