Apple uses Amazon Web Services’ custom artificial intelligence chips to power its search services. The company is also evaluating if it can use Amazon’s latest AI chips to pretrain Apple Intelligence models.
Apple’s Benoit Dupin, senior director of machine learning and AI, made these revelations Tuesday in a surprise appearance at Amazon’s annual AWS re:Reinvent conference in Las Vegas.
Custom AWS AI chips deliver up to 50% efficiency improvement
It is rare for Apple to openly talk about and praise one of its suppliers or vendors. During his appearance, Dupin revealed that Apple has been using Amazon Web Services to power its cloud services, such as iCloud, Apple Maps and Siri, for more than a decade.
The executive said using Amazon’s custom chips led to a 40% efficiency gain. Due to these benefits, Apple is considering using AWS’ custom AI chips to pretrain its proprietary AI models.
Interestingly, AWS CEO Matt Garman even revealed to CNBC that Apple came to Amazon and asked how it could help them build their generative AI capabilities and the required infrastructure.
At the AWS Summit, Amazon announced the general availability of its Trainium2 chips for training and deploying large language models, aka LLMs. A single Trainium2-powered EC2 instance is equipped with 16 T2 chips. This enables one instance to deliver up to 20.8 petaflops of compute performance.
Dupin said that could bring big improvements when working with Apple Intelligence.
“In early stages of evaluating Trainium2 we expect early numbers up to 50% improvement in efficiency with pretaining,” Dupin said.
Apple Intelligence will continue using Apple silicon-powered Private Cloud Compute
Apple only plans to use Amazon’s custom AI chips to pretrain its AI models. It will not use them to power Apple Intelligence features. For those, Apple turns to its Private Cloud Compute infrastructure, which uses M2 Ultra-powered servers. The company plans to switch to faster and more efficient M4-powered variants next year.
The majority of companies use Nvidia’s GPU for AI training. Apple is one notable exception, using its own Apple silicon chips or other non-x86 SoCs to train its AI models.
