news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

July 30 news, Monday, US time,appleThe company stated that itsAIThe core model of the system isGoogleThis is a sign that big tech companies are exploring other options besidesNvidiaGPUother alternatives.

Apple acknowledged in a newly published technical paper that it uses Google's Tensor Processing Unit (TPU) to train its artificial intelligence models. Separately, Apple on Monday launched a preview of Apple Intelligence to some users.

Nvidia's high-performance GPUs have long dominated the high-end AI model training market, and many technology companies including OpenAI, Microsoft, and Anthropic have adopted its GPUs to accelerate model training. However, in the past few years, Nvidia's GPUs have always been in short supply, so companies such as Google, Meta, Oracle, and Tesla are all developing their own chips to meet the needs of their respective AI systems and product development.

Meta CEO Mark Zuckerberg and Alphabet CEO Sundar Pichai both expressed their views last week, suggesting that their companies and others in the industry may be overinvesting in AI infrastructure, but also stressed that not doing so is a high business risk. Zuckerberg specifically pointed out that if they fall behind in this regard, they may lose their competitive advantage in key technology areas in the next 10 to 15 years.

In the 47-page technical paper, although Apple did not directly mention Google or Nvidia, it clearly stated that its Attentional Factorization Machines (AFM) model and AFM server were trained in a "cloud TPU cluster" environment, which indirectly indicates that Apple used the resources provided by cloud service providers to perform computing tasks.

In the paper, Apple emphasized: "The application of this system enables us to efficiently and scalably train AFM models, ranging from device-side AFM to server-side AFM and even larger models."

As of now, official representatives from Apple and Google have not responded to requests for comment.

Compared with many of its peers, Apple unveiled its AI strategy blueprint later, and after OpenAI launched ChatGPT at the end of 2022, other companies have quickly set off a warm pursuit of generative AI technology. This Monday, Apple officially launched Apple Intelligence, which debuted with a series of innovative features, such as Siri's new interface design, significantly improved natural language processing capabilities, and AI automatic summarization in the text domain.

In the coming year, Apple plans to launch more features based on generative artificial intelligence, including automatic generation of images and emoticons, and an enhanced version of Siri, which will be able to use users' personalized information to perform more complex and personalized tasks in various applications.

In a technical paper published on Monday, Apple revealed the specific details of the AFM model training on its device, which is completed on a separate "slice" containing 2048 of the latest TPU v5p chips. TPU v5p is the most advanced tensor processing unit currently available, which was first introduced in December last year. The AFM server training is even larger, which utilizes 8192 TPU v4 chips, which are carefully configured into eight slices and work together in the data center through a network to jointly support the powerful computing needs of the server.

According to Google's official information, the operating cost of its latest TPU is less than $2 per hour, but customers need to make a reservation three years in advance to ensure use. Since the advent of TPU designed for internal workloads in 2015, Google opened it to the public in 2017. Today, TPU has become one of the most mature and advanced custom chips in the field of artificial intelligence.

It is worth noting that despite having its own TPU, Google still maintains its position as Nvidia's top customer, using both Nvidia's GPUs and its own TPUs to train artificial intelligence systems, and providing access to Nvidia technology on its cloud platform.

Apple previously said that the inference process (using pre-trained artificial intelligence models to generate or predict content) will be partially performed on chips in its own data centers.

This is the second technical paper Apple has released recently about its AI systems, following a more extensive version in June, and further confirms that the company used TPUs in its development.