Gasgoo Munich- On Oct. 29, XHEART officially launched its large-model chip, the X7 Ultra. Built on the principle of "large models defining chips," its architecture is designed from the ground up to meet the demands of real, mass-production models — delivering 700 TOPS of compute power.
The X7 Ultra offers six core advantages. These include a massive NPU tailored for large models and unlocked hardware parallelism. It features a restructured AI toolchain and China's first mass-produced Armv9 automotive-grade CPU. It also provides deep heterogeneous computing and robust software-hardware security capabilities.
By aligning architecture with toolchain innovations, the X7 Ultra scales deployable model parameters tenfold — enabling larger models to run efficiently on the vehicle edge.
In real-world tests using mass-production autonomous-driving models, the X7 Ultra achieves an NPU utilization rate of 98.7%. Compared to current industry flagships, it reduces memory bandwidth consumption by 80% while increasing energy efficiency by 100%.
As autonomous driving shifts from scattered small models to end-to-end large models, chip requirements keep rising. Can computing power be fully utilized in real production models? Can a chip designed today keep pace with future model iterations? And can it deliver the necessary capabilities without exceeding power budgets?
The traditional approach — building the chip first, then forcing models to adapt — fails to address all three requirements simultaneously.
Therefore, XHEART has chosen a reverse-engineering approach since its founding in 2023 — starting with real-world model demands to define the chip architecture. It reorganized compute, storage, and data flow while leaving room for future model evolution. Understanding the large model first, then designing the silicon, ensures that algorithmic advances translate into product gains. This logic underpins XHEART's entire approach.
The X7 Ultra launch marks not only a milestone for XHEART's "large models define chips" roadmap. It signals a broader paradigm shift in autonomous-driving compute — moving from "general-purpose adaptation" to "native integration."
