Apple's 'It's Glowtime' event on September 9th is not just about a new iPhone. It's about the first hardware designed from the ground up to run Apple Intelligence, which changes everything for developers building on iOS.
When Apple announced Apple Intelligence at WWDC in June, people were skeptical. On-device AI, private cloud compute, an upgraded Siri - it all sounded good, but would the hardware deliver?
I have seen many attempts at on-device AI fail due to insufficient processing power or memory constraints. For instance, the early attempts at using Core ML for image classification on iPhone 8 were plagued by slow performance and high power consumption. However, with the A18 chip, Apple has made significant strides in addressing these issues, with the Neural Engine being 60 percent faster than A17.
The iPhone 16 answers that question. Apple designed the A18 chip specifically for AI workloads. The Neural Engine in A18 is 60 percent faster than A17. Apple also added a dedicated hardware accelerator for the type of matrix operations that large language models need.
For example, using the Accelerate framework, developers can achieve speeds of up to 2.5 TFLOPS on the A18 chip, which is a significant improvement over the 1.8 TFLOPS on the A17 chip. This increased processing power enables more complex AI models to run on-device, reducing the need for cloud calls and improving overall app performance.
The iPhone 15 Pro could run Apple Intelligence, but with limitations. The iPhone 16, on the other hand, was purpose-built for it across the entire lineup, not just the Pro models.
Apple's approach to on-device AI is different from Google and Microsoft. With Apple Intelligence, you don't call a model directly - you call system APIs that happen to use AI under the hood. This is both a constraint and an advantage. The trade-off is that developers have to rely on Apple's pre-trained models, which may not always be the best fit for their specific use case, but in return, they get the benefits of Apple's rigorous testing and quality assurance processes.
The constraint is that you can't swap in your own model or fine-tune for your use case. The advantage is that your app inherits the quality and privacy guarantees of Apple's system without any infrastructure cost. Using tools like Create ML, developers can still customize their AI models, but they have to do so within the constraints of Apple's ecosystem.
When a task is too complex for on-device processing, Apple routes it to their cloud infrastructure running on Apple Silicon servers. The privacy claim is that Apple can't see the data and it isn't retained after the request. Cryptographic attestation lets the device verify it's talking to genuine Apple hardware.
In my experience, the key to making this work is to carefully evaluate the trade-offs between on-device processing and cloud calls. For instance, using a library like Core Data, developers can cache frequently accessed data on-device, reducing the need for cloud calls and improving overall app performance. However, this requires careful consideration of data consistency and synchronization issues.
The upgraded Siri has context awareness across apps. It can send a message to someone you were just emailing or add a restaurant from a recent Safari search to your calendar. This works through on-device context, not a cloud call. For developers, there are new App Intents APIs that let you make your app's actions available to this upgraded Siri.
Apple Intelligence is rolling out in US English first, with other languages following in 2025. The features are opt-in initially. Expect a slower adoption curve than a typical iOS feature.