Smart cockpit large model, hybrid computing power still relies entirely on device-side output

In addition to the traditional LLM model, many large models in vertical fields are also developing rapidly, especially in the automotive field. More and more car companies are beginning to consider introducing AI large models in the car. Nowadays, due to the limitation of vehicle-side computing power, large models are difficult to run on smart cockpits, and the large models of cockpits that run throughout the network are lacking in user experience. However, as the new generation of smart car computing hardware has entered the next stage, smart cockpits are now beginning to be rolled out on new models.

Ideal Mind GPT

At the end of last year, Ideal released the latest OTA 5.0 version, taking the lead in upgrading the AD MAX 3.0 intelligent driving for the Max model. However, this time the OTA also includes the latest smart space SS 3.0, which officially introduced Mind GPT, a multimodal cognitive large model for smart cockpits, for the Max, Pro and Air models.

Ideal built the original base model of Mind GPT from scratch and adopted the self-developed TaskFormer neural network architecture. Mind GPT enabled Ideal's voice assistant "Ideal Classmate" to obtain a comprehensive upgrade in hearing and self-driving capabilities, and customized 111 fields and more than 1,000 exclusive capabilities based on its key scenarios, such as understanding semantics, multilingual communication, fuzzy perception, and itinerary planning.

From the Chinese large model ranking C-Eval website, it can be seen that Mind GPT, which was released on December 4, won the first place in closed-source models at the time. Although it was subsequently surpassed by Honor, HiSilicon and vivo's large models, Mind GPT currently still ranks fourth with a score of 89.7.

It can be seen that Ideal Mind GPT does not have high requirements for hardware. From the Mind GPT reasoning engine shared by Ideal, its reasoning load is allocated to the cloud GPT and the end-side NPU, thereby reducing the computing power requirements of the smart cockpit, so that more old models can also enjoy the complete experience of the large model.

NIO NOMI GPT

NIO claims that NOMI GPT is the industry's first multimodal perception large model deployed purely on the end-side. With this large model, the new NOMI can identify thousands of objects without the need for networking, and respond quickly. It can also complete the color combination of the atmosphere light based on non-color magnetic beads. Because it is purely end-side processing, NOMI GPT ensures that the owner's privacy data does not leave the car, improving information security.

The reason why it can be purely end-side is naturally inseparable from the support of computing hardware. All 2024 NIO models will adopt the new central computing platform ADAM, and the smart cabin will be equipped with Qualcomm's latest Snapdragon 8295 smart cockpit chip. With the 5nm process and architecture optimization, the 8295 has made a huge improvement in performance compared to the previous generation 8155. In addition to the CPU and GPU computing power for general computing and graphics processing, which have doubled and tripled respectively, the NPU computing power used for AI computing has also increased by 8 times, reaching 30TOPS.

In addition, NIO's ADAM has achieved cabin-driver integration through system-level innovation, which can not only expand computing power, but also realize cross-domain computing power sharing. In this way, multi-mode large-model end-side deployments such as NOMI GPT can call on the computing power from the intelligent driving chip Orin X, thereby improving the efficiency of CPU, GPU and NPU.

Written in the end

It can be seen that Ideal and NIO have adopted different strategies for the smart cockpit big model. Ideal's Mind GPT focuses on allowing all L series car owners to enjoy the smart space of the big model, while NIO pursues the ultimate hardware computing power and leaves all reasoning calculations on the end side. Both methods have their own advantages and disadvantages. In the future, as more car companies follow up on the smart cockpit big model, we will see which technical route is more popular.