On October 29, Xinxin Hangtu launched the X7 Ultra, an autonomous driving chip for large models. CEO Li Zonglin said the company lets large models define the chip, reverse-engineering chip architecture from actual model needs to optimize compute, storage, and data flow paths. During chip planning, it works with algorithm teams on operators and model architecture to match future large-model iteration.
The X7 series Ultra chip shipped over 200,000 units in an eight-month production cycle, deployed in 10 mass-production models, with fixed-point orders for over 40 models. Back in 2022, Xinxin Hangtu and Momenta settled on this R&D approach, completing functional design and validation before tape-out and setting a 21-day rapid adaptation record.
The X7 Ultra delivers 700TOPS, measured on production autonomous driving large models, with quantized deployment output over 99.99% similar to FP32 models. It features six core technologies: a large-core NPU that consolidates hundreds of GB/s inter-core data exchange into TB/s intra-core high-speed interconnect with high-precision dedicated accelerators; software-hardware co-design for parallel capability, executing compute and weight loading simultaneously with measured NPU utilization of 98.7%; a rebuilt AI toolchain with whole-network fusion to optimize task scheduling and cut repeated memory reads/writes, reducing bandwidth consumption by 80% and doubling energy efficiency versus industry flagship solutions; the first domestically mass-produced Armv9 automotive CPU for system scheduling and complex control; heterogeneous units supporting lidar signal processing at up to 20 million points per second, doubling sensor preprocessing stability; and compliance with ISO 26262 ASIL-B, with an integrated MCU certified ASIL-D.
Turing Award winner Jack Dongarra noted that deep collaboration among algorithm, software, and chip teams before finalization better fits large-model-era development. With the X7 Ultra, the R7 Ultra large model can scale 5–10x in parameters, with safety performance up 5–10x or more. Momenta works with the world's top ten automakers; as of July, its mass-production solutions were in over 1.4 million vehicles, with 240+ fixed-point models and 120+ delivered models.
Founded in Suzhou in December 2023, Xinxin Hangtu was incubated by Momenta's chip team and focuses on automotive-grade AI autonomous driving chips. It has raised nearly RMB 5 billion from SAIC, Chery, CXMT, and Renault, state funds including Guochuang AI Fund, Shenzhen Capital Group, and CDB Capital, and investors like CDH, IDG, NIO Capital, and Shunwei.
ICgoodFind Summary: Xinxin Hangtu's X7 Ultra automotive-grade autonomous driving chip is officially launched, pairing high compute with software-hardware co-design to offer a new chip option for in-vehicle large-model deployment.
