Apple Designing Its Own AI Chip For Data Centres

Apple, renowned for its custom processors used in the Mac, iPhone, iPad, Apple Watch, AirPods, and most recently, the Apple Vision Pro, is reportedly working on developing server-grade chipsets for its own data centres to run cloud-based AI applications. According to The Wall Street Journal, the project, named "Apple Chips In Data Centre", has been in progress for several years, though there is no specific timeline for when Apple will transition from traditional silicon provided by vendors such as Intel and Nvidia.

Apple is preparing to launch its own generative AI applications and services at WWDC in June, and is exploring partnerships with both OpenAI and Google for foundational models that will function via the cloud on the iPhone. Simultaneously, the company is fine-tuning its own software, such as Siri, to operate on custom-designed LLMs locally on its own silicon, providing an additional layer of privacy.

The upcoming M4 chipset is also expected to feature a new NPU for AI applications. This could debut as soon as this week with new iPads. Apple is reportedly collaborating with TSMC to design chipsets for data centres, likely focusing on developing chipsets for running AI models for inference rather than training them, as Nvidia's chipsets are widely used for the latter.

Apple is looking for the next big growth avenue with its sales being down this quarter. It has fallen behind on generative AI and it has been forced to shelve its decade-long electric self-driving car project. It has even moved engineering talent towards its generative AI overtures. It is also reportedly developing a robotic product with generative AI.

Apple's competitors in the smartphone industry have already announced generative AI integration, with Samsung introducing features for its flagship Galaxy S24 and S23 models, as well as its foldable phones from the previous year. Google has also incorporated on-device generative AI features into its Pixel 8 smartphones, and OnePlus has announced the addition of generative AI features utilising Google's Gemini models.

Tim Cook has strongly hinted that Apple will make its mark in the era of generative AI, and research papers have shown that Apple's software engineers have been diligently working on creating their own large language models that operate on-device.

To catch up, Apple has invested billions of dollars in acquiring startups like Darwin AI, developed its internal search engine called Ajax, and is working on a revamped version of Siri.

However, LLMs are currently limited by on-device processing and require significant computing power in the cloud, for which Apple relies on processors from Intel and Nvidia. The company also needs its own foundational models for the cloud. Apple has a sophisticated in-house silicon team led by Senior VP Johny Srouji and has been designing chips since 2009, following its acquisition of PA Semi.

Among the major cloud players, Google's TPUs are the most mature internal silicon solution. Amazon has its Graviton and Tranium processors, while Microsoft has recently launched its MAIA chips, which will be manufactured by Intel. OpenAI is also exploring the creation of a chip project, as mentioned by Sam Altman. It is evident that the industry is attempting to move away from Nvidia's hardware, which all major cloud players currently rely on more than their own custom solutions.