Gasgoo Munich-XHEART officially unveiled its X7 Ultra AI chip in Suzhou on October 29, introducing a product designed around the computing requirements of large AI models. Featuring 700 TOPS of computing power and six key technical features, the chip reflects the company's approach of designing hardware around evolving AI models. The company also presented its mass-production progress and outlined development targets for its next-generation X9 chip.
Image source: Momenta
From model adaptation to model-driven chip design
As intelligent driving shifts from systems built around multiple specialized models toward end-to-end large models, chip architectures must evolve accordingly. The key challenges include making better use of computing resources in production models, keeping pace with model updates, and supporting more advanced capabilities within vehicle-level power consumption and thermal constraints.
XHEART says the X7 Ultra's 700 TOPS computing capability is evaluated using production intelligent-driving models rather than theoretical benchmarks alone. The company also applies a model-output similarity requirement of more than 99.99% between quantized deployment results and the trained FP32 floating-point model, aiming to align chip optimization with mass-production requirements.
Six technical features of the X7 Ultra
XHEART says the X7 Ultra incorporates six key features addressing the computing power, precision, bandwidth and system coordination requirements of large AI models.
The first feature is a large-core NPU architecture designed for large models. By consolidating data exchanges that would traditionally take place between smaller cores into high-speed interconnects within a larger core, the design moves from bandwidth at the hundreds-of-GB/s level to the TB/s level. Dedicated high-precision acceleration engines are also used for key computing operators.
The second feature focuses on hardware parallelism. Through hardware-software co-optimization, more computing units can operate simultaneously, while computation and weight loading take place in parallel to reduce idle time. XHEART reports an NPU utilization rate of 98.7% in tests using production intelligent-driving models.
The third feature is a redesigned AI toolchain. Whole-network integration enables unified scheduling and optimization of model computing tasks, reducing repeated transfers of intermediate data to external memory. Combined with the new NPU architecture and parallel computing capabilities, this approach delivers the chip's stated 700 TOPS performance. XHEART also claims an 80% reduction in memory bandwidth consumption compared with industry flagship products and a 100% improvement in energy efficiency.
The fourth feature is an automotive-grade Armv9 CPU, which XHEART describes as China's first mass-produced implementation of its kind. The CPU handles complex control and system scheduling while improving multicore coordination across the chip.
The fifth feature is heterogeneous computing, with dedicated processing units handling specific tasks locally. The company says lidar processing capacity reaches up to 20 million points per second, while sensor preprocessing stability improves by 100%.
The sixth feature concerns functional safety. The X7 Ultra meets ISO 26262 ASIL-B requirements, while its integrated MCU has obtained ASIL-D certification. XHEART says the chip's computing capabilities and support for model updates are also intended to help meet the safety requirements of intelligent-driving systems.
Together, these features reflect XHEART's model-first design philosophy. The company says the combination of chip architecture and toolchain optimization enables a 10-fold increase in deployable model parameters, bringing larger models closer to mass-production vehicle applications.
X7 series advances to mass production
XHEART's approach is now being tested through commercial deployment. The company said it took two years from its founding to bring its chips into mass production.
According to XHEART, its first X7 chip was powered on within 24 hours of returning from fabrication, with its main data pathways successfully established. By the 17th day, the chip, underlying software and intelligent-driving model had been integrated into a complete vehicle system and tested on real roads in daytime, nighttime and heavy rain. The system was subsequently made available for demonstrations to customers and partners.
By the time of the launch event, the X7 series had been in mass production for eight months, with cumulative shipments exceeding 200,000 units. The chips have been installed in 10 production vehicle models and secured orders for more than 40 models. A computing platform based on two X7 chips has also been selected for multiple vehicle programs targeting L3 and L4 applications.
Industry collaboration and next-generation models
At a panel discussion, Turing Award laureate Professor Jack Dongarra said close collaboration among algorithm, software and chip teams before a design is finalized is an effective approach to chip development in the large-model era. He also highlighted improvements in data-transfer efficiency within and between chips as an important direction for hardware-software co-optimization.
For its next-generation X9 chip, XHEART has set development targets of 3,000 TOPS of large-model computing power and 700 GB/s of single-chip memory bandwidth. The chip is scheduled for release in 2027, with its computing and memory capabilities to be developed in tandem to support next-generation world models.
