Sept 2025 Apple iPhone Event

[leman]

The more I think about it, there more I see open questions. Why 3x matrix compute on Air but 4x on Pro, aren’t they using the same chip? Why no statement is made about the base A19? Why isn’t GB6 showing better results?

Maynard on MR had a similar speculation as me - that the additional FP16 pipe is used to accelerate matrix multiplications. This would result in 3x performance though. The math just doesn’t add up.

 

[exoticspice1]

Why 3x matrix compute on Air but 4x on Pro, aren’t they using the same chip?

simple, the Air has one less core. each gpu core has a Neural Accelerator

 

[Jimmyjames]

simple, the Air has one less core. each gpu core has a Neural Accelerator

That was my initial thinking but I suppose the question is why would 20% more neural cores yield 33% more performance?

 

[diamond.g]

That was my initial thinking but I suppose the question is why would 20% more neural cores yield 33% more performance?

Higher GPU clocks on the Pro cause better thermals?

 

[Jimmyjames]

The more I think about it, there more I see open questions. Why 3x matrix compute on Air but 4x on Pro, aren’t they using the same chip? Why no statement is made about the base A19? Why isn’t GB6 showing better results?

Maynard on MR had a similar speculation as me - that the additional FP16 pipe is used to accelerate matrix multiplications. This would result in 3x performance though. The math just doesn’t add up.

What results would you expect?

 

[leman]

simple, the Air has one less core. each gpu core has a Neural Accelerator

Higher GPU clocks on the Pro cause better thermals?

Ok, I suppose 10% clock reduction could explain it.

 

[leman]

What results would you expect?

I don’t expect anything, I just want to understand

 

[Jimmyjames]

I don’t expect anything, I just want to understand

You said “Why isn’t GB6 showing better results?” Better implies the results are less than you expected no?

 

[leman]

You said “Why isn’t GB6 showing better results?” Better implies the results are less than you expected no?

Ah, that’s what you mean. GB6 ML only shows 2x improvement and not 3x or 4x as advertised. So I see the double FP16 rate on general-purpose shaders but not the matrix-related improvements. This strikes me as odd, since I’d expect CoreML models running on the GPU to take advantage of the new functionality. After all, it’s Apple’s own framework.

 

[Jimmyjames]

Ah, that’s what you mean. GB6 ML only shows 2x improvement and not 3x or 4x as advertised. So I see the double FP16 rate on general-purpose shaders but not the matrix-related improvements. This strikes me as odd, since I’d expect CoreML models running on the GPU to take advantage of the new functionality. After all, it’s Apple’s own framework.

Ahh ok understood.

I hope it’s not like Apple mentioning 2x faster ray tracing that didn’t translate to 2x improvements in any ray tracing application.