More than a month after the official announcement of "Apple Intelligence," Apple's artificial intelligence (AI) system has finally landed on terminal devices. On July 29th, Eastern Time, the company released the first iPhone AI version of Apple Intelligence. The new software is currently only released in the developer beta version of iOS 18.1 and is only available to registered developers who have paid $99 per year.
The features of the beta version mainly revolve around writing tools, Siri, photo albums, and other aspects. This update has not yet integrated ChatGPT functionality, and Apple stated that this feature and more updates will be officially launched next year.
On the same day, Apple released a research paper on Apple's intelligence, revealing that its new AI system relies on Google designed TPU (Tensor Processing Unit) instead of the widely used NVIDIA GPU.
The Daily Economic News reporter noticed that the price of Google TPU is relatively competitive, with the latest model of the product costing less than $2 per chip per hour (training) during the three-year pre order period. Moreover, compared to Nvidia's independent chips, the advantage of Google TPU lies in its particularly high chip level interconnectivity. Apple's technical paper indicates that TPU architecture can develop larger and more complex AI models.
On its first day of popularity, Apple's smart products exploded with three new modules
With the launch of iOS 18.1 Beta, registered developers can now experience some of the features of Apple AI from today on. It is reported that this update only supports M-series and A17 Pro chips, which means it is only compatible with iPhone 15 Pro and iPhone 15 Pro Max on iPhone, and similar versions are also available on iPad and Mac. In addition to hardware requirements, the developer's regional setting must be in the United States, and the device and Siri language must be in English.
The Apple Smart launched this time only has some functions online, mainly focusing on modules such as text generation, Siri, and photo albums. However, this update has not yet integrated ChatGPT functionality, and Apple stated that this feature and more updates will be officially launched next year.
Ruoming Pang, the head of Apple's Basic Big Model team, emphasized that these basic models are not chatbots, but support a wide range of features, including writing assistance, tool usage, and code.
X platform screenshot
As an important component of Apple AI, the application scope of text generation function is not limited to official Apple applications. As long as the standard input text system is used, this function can also be used for text summarization, proofreading, and rewriting in third-party applications. In addition, combined with the audio transcription function already available in the voice memo, the text generation system can also generate recordings.
Regarding the AI features provided in the latest beta version of iOS 18.1, X netizens quickly posted feedback. A netizen claimed that "Apple Intelligence helped me find a better way to express what I really want to say." In his video, Apple Intelligence modified the content it was preparing to send in the conversation to be more tactful.
X platform screenshot
In addition, Siri's comprehensive overhaul is also stunning. Some netizens praised Apple's new Siri interface and said that the new version of Siri has "absolute breakthrough".
X platform screenshot
The update of the album allows users to search for specific photos or even specific moments in videos using natural language, and create photo memories.
X platform screenshot
On July 29th, Apple also released a technical report on the Apple Intelligent Model and stated that the Apple Cloud Large Model achieved better results than GPT-4 in tasks such as instruction following and text summarization.
Apple's latest technical documentation
Why did Apple give up Nvidia GPU?
Apple has previously stated that ChatGPT is not a mandatory feature in Apple AI, and its main functionality is driven by its own large models. And another astonishing fact is that according to Apple's latest technology report, Nvidia's hardware infrastructure content is zero in the training of Apple's basic model.
The training of Apple's basic model is carried out through its AXLearn framework based on JAX, using Google TPU hardware. 8192 TPUv4 chips are used on the cloud side, and 2048 TPUv5p chips are used on the end side.
Why is this surprising? NVIDIA GPUs have always been the preferred choice for training large-scale AI models, with advantages in computing power and parallel processing capabilities, occupying approximately 80% of the market share. However, according to foreign media analysis, the growing demand for Nvidia GPUs has led to tight supply and price increases, forcing tech giants to explore alternative solutions.
According to reports, Nvidia's latest Blackwell GB200 is priced between approximately 30000 to 40000 US dollars. According to market reports, chips based on the Blackwell architecture are currently the fastest AI chips in the world and the most expensive chips to date. However, demand remains strong and this trend will continue until 2025. It is reported that Nvidia's CoWoS (a semiconductor packaging technology) packaging orders will reach 340000 next year.
Compared to the independent chips sold by Nvidia, Google TPU is based on the Google Cloud platform, and Apple does not need to invest heavily in hardware infrastructure to obtain a large amount of computing resources. Google TPU is custom designed for machine learning tasks and has been part of Google's internal infrastructure since 2015. It was publicly available through Google Cloud in 2017.
In contrast, Google TPU also offers a more competitive price. According to foreign media, during the three-year pre order period, the latest model of the product has a (training) cost of less than $2 per chip per hour. Apple's technical paper indicates that TPU architecture can develop larger and more complex AI models.
A Silicon Valley engineer who declined to be named previously told the Daily Economic News reporter that the advantage of Google TPU over Nvidia GPU is its particularly high chip level interconnectivity. Therefore, some analysts believe that this competitive pricing, combined with the architectural advantages of TPU, makes it a substitute for Nvidia GPUs.
