The announcement: a chip designed for language models
Presenting the new chip at a dedicated press conference, XHEART CEO Li Zonglin outlined the company’s strategy: to build a vehicle processor designed around artificial intelligence models, rather than adapted after the fact. The news, reported by IT之家, centers on the new X7 Ultra unit, which the company describes as a chip dedicated to large models for automotive use.
The architecture revolves around an NPU with a “super core” structure, designed to enable data communication within a single core rather than between different cores: according to the company, this approach raises interconnection speed to the order of terabytes per second. This is complemented by a multi-level cache hierarchy intended to support large volumes of data and simultaneous accesses, reducing inference times and power consumption.
The novel element for the automotive sector is the integration of an acceleration engine dedicated to Transformer model operators, such as Layernorm, Softmax and Transpose, with native support for FP16/FP32 precision vector calculations. The stated goal is to preserve model accuracy without sacrificing efficiency.
The promises: more parameters, less memory consumption
XHEART claims that the X7 Ultra can handle up to ten times the model parameters of previous solutions, thanks to joint work on architecture and the software toolchain. According to the company, this is the condition for larger models to run directly on board the vehicle.
The figures provided by the company are ambitious: a compute utilization rate of 98.7%, memory bandwidth consumption 80% lower than industry-leading products and doubled energy efficiency. According to XHEART, the chip also integrates the first mass-produced Armv9 CPU for vehicles in China, with a 20% improvement in performance and 10% in efficiency.
The most significant technical aspect is the large-cluster architecture, which the company describes as an improvement in cooperation among multiple cores and in L3 memory access performance, together with the integration of the SVE2 vector engine: an evolution of the previous NEON that offers better digital signal processing capabilities.
What remains to be verified
All performance data cited come from XHEART statements reported by the specialist press and have not yet been measured by independent third parties. Comparisons with “industry-leading products” do not indicate which chips were used as benchmarks, nor how the tests were conducted.
It also remains to be clarified which automakers have adopted the X7 Ultra and on what production timelines, as well as whether the tenfold parameter limit translates into models that can actually run on board. Finally, the presentation contains no indication of the chip’s price or commercial availability. For now, the information reflects the company’s vision rather than verified performance.
