Tianshu Zhixin releases the general-purpose GPU flagship product Tiangai 300: supports a variety of calculations such as scalar, vector and tensor

Kuai Technology reported on July 20 that according to the "Tianshu Zhixin" public account, during the 2026 World Artificial Intelligence Conference, Tianshu Zhixin officially released a new generation of general-purpose GPU flagship product - Tiangai 300. This product is based on the SIMT general computing architecture, supports multiple computing types such as scalar, vector, and tensor, and is deeply optimized around key technologies such as Attention mechanism, MoE architecture, AF separation, PD separation, and large-scale system expansion. According to test data released by Tianshu Zhixin, the performance of Tiangai 300 under complex business loads has reached and exceeded the international mainstream solutions based on Hopper architecture. Specifically, the product shows obvious advantages in a number of key indicators: Attention efficiency: The overall efficiency exceeds 90%. In the 64k long context scenario, Attention performance is 10% higher than the Hopper solution, effectively improving the computing resource utilization in long sequence task training and inference; MoE inference performance: The average performance is 10% higher than the Hopper solution, and more inference tasks can be completed under the same computing power conditions; Communication and decoding performance: The first word delay is reduced by about 20% compared to the Hopper solution, and the average communication delay is reduced by about 13%. Especially for high-frequency, multi-card, and small data volume communication scenarios in the Decode stage, the protocol overhead and link data flow mechanism are optimized; the overall performance in the Decode stage is 10% higher than the Hopper solution. Tianshu Zhixin said that Tiangai 300 is now ready for large-scale application, and has been deeply adapted to domestic mainstream cloud manufacturers, server manufacturers, Internet ecosystems and super node systems, and can support full scenario coverage from stand-alone deployment to large-scale cluster solutions.