Apple’s split-level approach to AI appeals to me. I appreciate that Apple recognizes that some of its larger generative models require more compute space. It makes sense to treat the local and cloud planes as one persistent AI channel, a kind of permanent Bifrost between AI worlds.
Apple’s first foray into generative AI, which leaves new phones like the iPhone 15 behind, is less tolerable.
I’ve been thinking about Apple’s explanation in my head for days. Craig Federighi told an audience during a keynote chat after WWDC 2024 that the A17 Pro neural engine of the iPhone 15 Pro and iPhone Pro Max is “twice as powerful as the generation before it.” In other words: the A16 Bionic of the iPhone 15 is not fast enough and lacks the ‘architecture’.
Okay, but what does that actually mean? There is a 1 nanometer difference between the A16 Bionic (4nm) and the A17 Pro (3nm), but that usually has more to do with efficiency (battery life) than power.
Additionally, the A16 Bionic is capable of performing many machine learning tasks on-device, meaning it can’t operate without AI knowledge. But generative AI, which creates something new based on inferences and clues from data it can process through models of different sizes, is slightly different.
At least that’s what I think.
One thing we know is that even the A17 Pro and the M1 through M4 Apple Silicon chips also have their limits. That’s why Apple Intelligence sometimes sends the prompt to the Private Compute Cloud, where all of Apple’s major generative models reside.
I think this means it’s all about processing power, and a cloud-based server offers more than a mobile device. However, Apple said that these clouds are built around Apple Silicon. So are they more powerful? There’s a theory that gives me hope for the next generation of base iPhone models and their ability to support Apple Intelligence.
After the keynote, I heard some people suspect that the reason the iPhone 15 couldn’t support Apple Intelligence wasn’t because of the chip, but because of the memory.
Apple’s base model iPhone 15 has 6GB of RAM, which isn’t a terrible number and only 2GB less than that of the Pro and Max models. Many of the latest AI-enabled Android phones, like the Samsung Galaxy S24 Ultra and Google Pixel 8 Pro, pack 12GB of RAM and don’t even promise to keep the AI local.
I bet the iPhone 15 could support Apple Intelligence, but only the smallest generative models due to memory size. This would mean a very limited and perhaps unsatisfactory generative AI experience.
I still think there’s some achievement at play here too. The A17 Pro is a more powerful chip, but some of it comes from a better neural engine and some of it comes from a larger memory pool. The combination is enough to support a better Apple Intelligence experience.
If I’m right, every iPhone 16 model launching in September will ship with 8GB of RAM and an A18 or A18 Pro mobile CPU, and every one will support Apple Intelligence and that fresh new Siri.
The base iPhone 16 model may still not support as many generative models as the iPhone 16 Pro and Pro Max, but iPhone 16 owners may never notice because Apple doesn’t signal when Apple Intelligence is running locally or using Private Compute Cloud.
Frustration about the lack of Apple Intelligence for cheaper best iPhones (or the one you just bought last week) is justified, but if you want to trade up for a still-reasonably priced iPhone later this year without losing access to Apple Intelligence, possibly an iPhone 16 with 8GB of RAM to show you.