Neural Engine patent growth
- Patent analysis traced Apple's Neural Engine growth from roughly 0.6 TOPS to about 38 TOPS over seven years. - PatSnap reviewed more than 29 patents to chart the Neural Engine's scaling and architectural shifts. - The patent trajectory shows aggressive on-device ML scaling, with implications for power, memory, and integration trade-offs (x.com).
A Neural Engine is the part of Apple’s chip built for artificial intelligence math, and patent analysis says its speed rose from about 0.6 to 38 trillion operations per second in seven years. (machinelearning.apple.com) (apple.com) PatSnap said it reviewed more than 29 Apple patents filed from 2018 to 2025 to map how the Neural Engine changed from a small add-on for Face ID into a core block in Apple’s A-series and M-series chips. (patsnap.com) Apple’s own machine learning team said the first Apple Neural Engine shipped in the A11 chip in 2017 and delivered 0.6 teraflops in half-precision format for features including Face ID and Memoji. (machinelearning.apple.com) By May 2024, Apple said the M4 chip’s Neural Engine could reach 38 trillion operations per second, and called it 60 times faster than the first Neural Engine in A11 Bionic. (apple.com) A neural engine is a dedicated calculator for machine learning, the kind of software that recognizes faces, transcribes speech, or predicts the next word, and Apple has pushed more of that work onto the device instead of sending it to a server. (machinelearning.apple.com) (apple.com) Apple said in 2022 that optimizing Transformer models for the Neural Engine could make one example model run up to 10 times faster while using 14 times less memory, a sign that raw speed alone was not the only design target. (machinelearning.apple.com) That trade-off now sits at the center of Apple’s artificial intelligence strategy. Apple says Apple Intelligence handles many requests on device, and sends more complex jobs to Private Cloud Compute, a server system that runs on Apple silicon. (apple.com) (support.apple.com) PatSnap’s timeline breaks the chip effort into three phases: foundation from 2015 to 2017, neural acceleration from 2017 to 2020, and unified architecture from 2020 to 2026, when Apple tied central processor, graphics processor, memory, and Neural Engine more tightly together. (patsnap.com) That tighter integration is visible in Apple’s recent chip launches. Apple said in October 2024 that M4 Pro and M4 Max paired the faster Neural Engine in the M4 family with higher unified memory bandwidth for artificial intelligence workloads on Mac. (apple.com) Apple’s latest software push leans on that hardware base. In June 2025, the company said developers would get direct access to the on-device large language model behind Apple Intelligence, extending the Neural Engine’s role from Apple features to third-party apps. (apple.com)
