Currently, the AI industry has some key names, such as NVIDIA. The company grew enormously thanks to other big names using its GPUs for AI model training, becoming the market leader. However, it seems that NVIDIA’s position does not fully convince Apple, who preferred to resort to Google’s AI chips to train its own LLMs.
Technical paper offers details on Apple’s AI training
Apple came a little late to the AI segment, but it does not want to be left behind. At the latest WWDC24, the company announced a plethora of AI-powered features for its mobile devices and Macs. Most of these features will arrive starting next year, and even those that were supposed to arrive this September were delayed. However, something like this was to be expected considering the arduous development behind high-end AI models.
That said, Apple published a technical paper offering details on the development of its LLMs. These will be the basis of the Apple Intelligence platform that will support the company’s own models and others from third parties. Apple’s first external partner in this regard is OpenAI, collaborating to integrate ChatGPT functionality into iOS for cloud-based tasks. The company might also offer Google Gemini support in the future.
Apple turns to Google chips for AI training with cloud-focus
Anyway, the paper reveals that Apple turned to Google’s TPU platforms instead of NVIDIA GPUs. This contrasts with the approach taken by other major companies like OpenAI and Microsoft. More specifically, the paper mentions “Cloud TPU clusters” as the driver of Apple’s AI training. The company refers to its own set of LLMs as “Apple Foundation Model (AFM).” Basically, the Cupertino giant rented cloud processing power from Google instead of acquiring NVIDIA GPUs.
The company claims that the cloud approach was the best for them. According to the paper, “This system allows us to train the AFM models efficiently and scalably, including AFM-on-device, AFM-server, and larger models.”
Apple is also looking to make its AI models have a low “harmful” response rate. This includes handling sensitive issues, factual errors, performance in mathematical operations, and more. According to the paper, AFM currently shows a “harmful output violation rate of 6.3%.” For reference, Apple detected a rate of 28.8% in the GPT-4 model. With respect to on-device processing, AFM exhibits an error rate of 7.5%, while Meta’s Llama-3-8B model shows a 21.8% rate.