At the 4th AI-Defined Vehicle Forum 2026 hosted by Gasgoo on September 18, Kelly Liu, NXP Automotive Edge AI GC Regional Marketing Manager, delivered a speech titled "On-Vehicle AI Products and Implementation Scenarios." She provided a systematic overview of NXP's product portfolio, implementation scenarios, and toolchain capabilities in the field of on-vehicle edge AI.Liu noted that on-vehicle AI is expanding from large models in cockpits and ADAS to smaller models in zone controllers and MCUs. The period from 2028 to 2030 is expected to mark a peak in deployment. NXP's differentiation lies in its proprietary NPU IP, which is unified across its product portfolio, along with its low-loss toolchain, enabling it to meet the needs of vehicles ranging from high-end to economy segments. The three scenarios currently offering the greatest potential for implementation are BMS, intelligent chassis control, and virtual sensors, addressing the core demands of performance enhancement and cost reduction.In terms of product portfolio, NXP’s on-device AI products are arranged by process node from low to high, covering categories such as AI-capable MCUs, dedicated Audio DSPs, and UWB products. The entire series features proprietary embedded NPU IP—a unique full-stack proprietary support model in the industry—backed by a fully proprietary toolchain. The unified NPU IP architecture spans the full product lineup, from central computing platforms to the S32K5 MCU. This delivers acceleration efficiency 10 to 30 times that of the M7 core at the same clock frequency.Regarding central computing and AI Box solutions, NXP’s central computing platform supports task isolation across multiple domains with different functional safety levels, efficiently enabling cross-domain collaboration. The AI Box product targets existing economy-class cockpit platforms in the 120,000 to 150,000 yuan range, addressing the need for intelligent upgrades during model refreshes. It supports single-chip compute power of 40 TOPS, scalable to 160 TOPS, with power consumption of just 6.4W. The device supports two modes: independent Gigabit Ethernet connection to domain controllers or an external compute card attached to cockpit chips. For model adaptation, domestic adaptations for Qianwen and ModelBest are complete, while overseas adaptations for Llama and Gemini are finished, removing barriers for customer development.Turning to zone controller AI scenarios, Liu revealed that 10 of the top 14 OEMs by sales in China have concrete plans. They aim to deploy high-real-time small AI models in next-generation architectures between 2028 and 2030. The NXP S32K5 chip offers NPU compute power ranging from 51 to 400 TOPS. Its paired R52 core supports the direct deployment of floating-point models, avoiding precision loss without the need for quantization. Among the three key scenarios, the BMS AI solution replaces traditional equivalent circuits and Kalman filtering with AI models. It optimizes SOC and SOH accuracy from the 3% to 5% range to within 1%. Open-source pre-trained models are now available for rapid customer evaluation. The intelligent chassis AI solution uses a lightweight road classification model with only 186,000 parameters. It has a single inference time of 2.2 milliseconds, meeting the 10-millisecond control loop requirements of the chassis. The virtual sensor solution focuses on cost reduction and adding new measurable dimensions, with cooperative projects for tire pressure and automatic headlight sensor replacement nearing mass production.On the toolchain front, NXP’s proprietary toolchain primarily handles the deployment and adaptation of models trained on servers or PCs to the device side, while also achieving model compression. Through exclusive strategies like mixed-precision quantization, it controls the accuracy loss of converting 32-bit floating-point to 8-bit integer data within 0.1%—far surpassing the performance of open-source toolchains. Liu stated that NXP will leverage its full-line proprietary NPU IP and low-loss toolchain as core competencies. The goal is to drive the large-scale implementation of on-vehicle AI across a broader range of scenarios.
+22 others
