XHEART New Silicon Journey Releases Large-Model Chip X7 Ultra, 700 TOPS Compute Accelerates Mass Production

On October 29, XHEART officially launched the X7 Ultra large-model chip. With 700 TOPS of large-model compute and six major product advantages, it translates the philosophy of "large models defining chips" into product capabilities, and together with industry partners it showcased mass-production results and announced the development goals for the next-generation chip, X9. Currently, intelligent driving is moving from the division of labor among multiple small models toward end-to-end large models. Changes in model architecture place new demands on chips: whether compute can be fully utilized in real mass-production models, whether the architecture can keep pace with model iteration, and whether stronger intelligence can be supported under limited energy consumption and heat dissipation conditions. When the company was founded in 2023, XHEART established the technical route of "large models defining chips": starting from model requirements in real applications, working backward to the chip architecture, and reorganizing computation, storage, and data flow. This approach also runs through performance evaluation. The X7 Ultra's 700 TOPS of large-model compute is based on evaluation with real mass-production intelligent-driving large models, and is constrained by requiring the similarity between the model output after quantized deployment and the trained FP32 floating-point model to reach 99.99% or above, so that chip optimization directly targets mass-production requirements and the driving experience. Centered on the large model's requirements for computation, precision, bandwidth, and system coordination, the X7 Ultra offers six major product advantages: an ultra-large-core NPU designed specifically for large models, unlocked hardware parallelism, a restructured AI toolchain, China's first mass-produced Armv9 automotive-grade CPU, deep heterogeneous computing, and deep heterogeneous computing. Based on actual testing with mass-production intelligent-driving large models, the X7 Ultra NPU achieves a compute utilization rate of 98.7%. The X7 Ultra delivers 700 TOPS of large-model compute, reduces memory bandwidth consumption by 80% compared with industry flagship products, and doubles energy efficiency. Multi-core collaboration enables the entire chip to work together efficiently and supports continuously evolving intelligent-driving systems. At the same time, efficient compute and room for model iteration support driving safety, fastening the seatbelt for intelligent-driving systems. XHEART CEO Li Zonglin said that these six capabilities come from the same starting point: first understand large models, then design chips. Through collaborative innovation in architecture and toolchains, the X7 Ultra supports a 10-fold increase in the number of deployable model parameters, bringing stronger model capabilities into mass-produced vehicles. The technical route of "large models defining chips" is now being tested by large-scale mass production. From the company's founding to chip mass production, XHEART took two years. After the first X7 chip came back from the fab, it was powered up in less than 24 hours, with all main data paths connected; on the 17th day, the chip, low-level software, and intelligent-driving model achieved a full-vehicle closed loop and were tested on real roads in daytime, nighttime, and heavy rain conditions, after which experiences were opened to customers and partners. Rapid vehicle integration was followed by completing mass-production delivery together with customers. As of the launch event, the X7 series had been in mass production for 8 months, with cumulative shipments exceeding 200,000 units, installed in 10 mass-production vehicle models, and orders secured for more than 40 vehicle models. The L3/L4 computing platform based on dual X7 chips has also won designated projects for multiple vehicle models, expanding from assisted driving to higher levels of autonomous driving. Behind "large models defining chips" is the shared judgment of XHEART and industry partners on the direction of technology. At the launch event, Momenta CEO Cao Xudong publicly shared the company's third insight for the first time, namely "large models defining chips." This judgment was formed at the end of 2022 and was put into action jointly by Momenta and XHEART. Cao Xudong believes that chips defined by large models can achieve a 10-fold performance improvement compared with general-purpose chips. Achieving scaled L4 requires models with 10 billion to 100 billion parameters. How to carry such models under the constraints of vehicle-side cost, power consumption, and reliability requires both parties to start from model requirements and design chips collaboratively. The soon-to-be-released R7 Ultra demonstrates the latest result of this collaboration. With the support of the X7 Ultra, the R7 Ultra's model parameter scale is expanded by 5-10 times, and safety performance is improved by more than 5-10 times, driving world models from "understanding the world" to "simulating the world." Cao Xudong said that large-model chips are accelerating Moore's Law for intelligent driving. From 2022 to 2026, intelligent-driving capability will improve tenfold every two years, driving urban NOA from early exploration to an application tipping point and accelerated adoption. Looking ahead, to drive this pattern toward a tenfold improvement every year, the key is "large models defining chips." He judges that L3 and L4 are expected to reach an important tipping point in 2028 and move toward large-scale deployment in 2030.