On October 29, XHEART (Xinxin Hangtu) officially released its new-generation smart-driving large-model chip, the X7Ultra. In response to the industry trend of intelligent driving evolving from multiple traditional small models to "end-to-end large models," XHEART upholds the core philosophy of "large models define chips." From the very beginning of its founding, it decided to start from the underlying requirements of real production models, work backward, and redesign the chip's computing architecture, memory solution, and data flow, while reserving ample room for the continuous evolution of future models.
In terms of core computing power and architecture design, the X7Ultra achieves large-model computing power of 700 TOPS, with measured alignment for production models as high as 99.99%. The chip is purpose-built for large models and adopts a super-large-core NPU architecture. Intra-core data interaction achieves a TB/s-level speed improvement compared with traditional inter-core interaction. At the same time, its multi-level cache architecture combines large capacity with high concurrency, effectively reducing latency and power consumption during inference.
