According to Zhichuang Finance APP, Changjiang Securities released a research report stating that Apple (AAPL.US) launched a personal intelligent assistant, Apple Intelligence, at the global developers conference in 2024. The assistant contains multiple powerful generative models that can quickly and efficiently handle users' daily tasks and immediately adapt to their current activities. In addition, Apple released the basic model AFM, which empowers the underlying operating system. Apple Intelligence is powered by the AFM basic model. In terms of computing power, the AFM model is supported by Google's (GOOG. US) TPU chip, the performance of which is equivalent to that of Nvidia's (NVDA. US) flagship chip.

Apple released Apple Intelligence, an intelligent assistant.

Apple released the basic model AFM, which empowers the underlying operating system. They also launched a personal intelligent assistant, Apple Intelligence, which contains multiple powerful generative models that can quickly and efficiently handle users' daily tasks and immediately adapt to their current activities. Apple Intelligence can write and refine text, prioritize and summarize notifications, and create interesting images for conversations with family and friends, as well as simplify cross-application interactions with in-app operations. Apple Intelligence will be available on IOS18, IOS18ipad, and MacOS18 operating systems. Apple Intelligence is currently only available on the iOS18.1Beta version for registered developers to try out, and the subscription price is $99 per year. Ordinary users still need to wait in line.

Apple Intelligence is powered by the AFM basic model.

The AFM basic model mainly includes two parts: the terminal-side model and the cloud-side model. The terminal-side model is designed for specific scenarios of terminal-side applications, can only process language-related single-modal tasks, and can be localized on devices such as iPhones, iPads, and Macs, with a model parameter amount of 3 billion. The cloud-side model is designed for private cloud application scenarios, has multi-modal capabilities, and has higher generalization capabilities to handle more generic tasks. These two basic models are part of the generative model family created by Apple. In addition to the above two models, Apple Intelligence also includes an encoding model and a diffusion model. The encoding model is based on the AFM language model and is used to inject intelligent features into Xcode; the diffusion model helps users express themselves visually, such as in the Messages app.

The performance of the AFM cloud-side model is comparable to that of GPT-3.5 and slightly inferior to that of GPT-4.

In the model performance evaluation phase, Apple designed 1393 tasks to compare the performance of the AFM model with other mainstream models. The comparative results show that the performance of the AFM cloud-side model surpasses that of the Mixtral-8x22 hybrid expert model, the GPT-3.5 model, and is slightly inferior to the GPT-4 and LLaMA-3-70B models; in the terminal-side model, the performance of the AFM terminal-side model is similar to that of the mainstream terminal-side models on the market. The results of human evaluation show that the performance of the AFM terminal-side model exceeds that of mainstream models such as Gemma-7B, Phi-3-mini, Mistral-7B, Gemma-2B, and is slightly inferior to the LLaMA-3-8B model. These results prove the excellent performance of the AFM terminal-side model, which is expected to have high practicality on devices such as iPhone and iPad.

Heterogeneous chips may become a new direction for AI computing power development.

In terms of computing power, the AFM model is supported by Google's TPU chip. Google provided computing power support for this training. The cloud-side AFM-server model was trained with 8192 TPU V4 chips. In the training phase, Apple divided the 8192 chips into 8 groups, each group consisting of 1024 chips connected in series to form a basic unit, and the groups remain parallel to each other. The training data and iterations are only completed within the group; the terminal-side AFM-on-device model was trained with 2048 TPU V5p chips.

TPU (Tensor Processing Unit) is an ASIC chip designed specifically for processing tensor operations. TPU achieves efficient computation through pulsation array mechanism. Compared with GPUs, TPU does not need to access memory frequently, reducing the interaction with storage and greatly improving computing efficiency. Therefore, the effective utilization rate of TPU's computing power is higher than that of GPUs; the computing power utilization rate of GPUs is usually 20%-40%, while the computing power utilization rate of TPUs often exceeds 50%.

Risk warning

1. AI technology is not developing as expected;

2. The downstream application demand is not as expected.