Sept 2025 Apple iPhone Event

[Joelist]

Overall nice updates. Adding Neural units to the GPUs is a great upgrade.

 

[Jimmyjames]

No mention of RT improvements is a little disappointing.

 

[leman]

There is still an ane. No mention of performance improvements in it afaik. What were you thinking the gpu cores uplift would be?

It all depends on which precision they mean when advertising 4x improvement. FP32? Amazing! FP16? Meh… INT8? Lackluster…

I was hoping for 4-way FP16 dot product with 32-wide SIMD, which would match the performance of the 4070 RTX on the 40-core Max variant. But they mention 4x, which sounds more like 2-way dot product. My math could be off completely of course, it’s late and I’m tired from spending my day at the beach

 

[Jimmyjames]

It all depends on which precision they mean when advertising 4x improvement. FP32? Amazing! FP16? Meh… INT8? Lackluster…

I was hoping for 4-way FP16 dot product with 32-wide SIMD, which would match the performance of the 4070 RTX on the 40-core Max variant. But they mention 4x, which sounds more like 2-way dot product. My math could be off completely of course, it’s late and I’m tired from spending my day at the beach

Would Geekbench scores provide evidence of this or is there another test which would be better?

 

[Altaic]

Really, is that what they said? No more separate ANE? That is quite surprising.

edit: I am looking at the images and the ANE cores are still visibly present in the A19 Pro SoC depiction. So I’m inclined to interpret this in the most conservative way that the matmul performance of the new GPU cores is 4x of the old GPU cores. Which is a bit disappointing to be honest (and in line with my speculation thst the new mixed pipeline is used to support dot product computation).

Yeah, I just noticed that the ANE is still present as a separate unit. That said, why is 4x matmul disappointing to you? It’s pretty much what I was hoping for.

 

[leman]

Would Geekbench scores provide evidence of this or is there another test which would be better?

I don’t know if GB uses matmul APIs in their GPU tests. The difference could show up, if they do.

To test the new units, the simplest thing would be to write some very simple compute shaders using the new Metal tensor APIs.

Yeah, I just noticed that the ANE is still present as a separate unit. That said, why is 4x matmul disappointing to you? It’s pretty much what I was hoping for.

Well, the RTX 4070 offers 60 TFLOPs of FP16 matmul with FP32 accumulate and 120 TFLOPs with FP16 accumulate - and that is without scarcity. Would be great if the new Max Studio offered competitive performance.

 

[Altaic]

Well, the RTX 4070 offers 60 TFLOPs of FP16 matmul with FP32 accumulate and 120 TFLOPs with FP16 accumulate - and that is without scarcity. Would be great if the new Max Studio offered competitive performance.

IIRC the actual utilization for nvidia GPUs is quite low due to power or thermals or something and they only get close to their specs with trivial matrices (e.g. all 0s or all 1s). I can try to find the article again if you’d like.

 

[Jimmyjames]

IIRC the actual utilization for nvidia GPUs is quite low due to power or thermals or something and they only get close to their specs with trivial matrices (e.g. all 0s or all 1s). I can try to find the article again if you’d like.

I certainly would be interested to see it.

 

[Altaic]

I certainly would be interested to see it.

Found it (though I haven’t had a chance to reread it to check that I’m not misremembering): https://www.thonking.ai/p/strangely-matrix-multiplications

Edit: I should also mention that Horace He (the author of that blog) isn’t some internet rando. He’s been a gem in ML and worked on PyTorch for years, and is now a founding engineer at Thinking Machines which has gathered an extraordinarily talented crew. He posted about it here: https://www.thonking.ai/p/why-pytorch-is-an-amazing-place-to

Edit2: The penalty for real data is not nearly as bad as I’d remembered (my apologies). Still a very interesting read, though.

 

[Roller]

I see that the Air and 17 (non Pro) are limited to USB 2. IDK how much more it would have cost Apple to give them USB 3. But I suspect they think most transfers to and from the iPhone are done wirelessly these days, which is true in my case.

 

[rdrr]

New Airpods and Ultra 3 ordered. Not sure if the iPhone 17 Pro Max is for me, but my trade in gets me a $570 credit this year.

 

[Citysnaps]

I'm a little bummed the 17 Air doesn't have camera macro capability. Which now that I think about it, makes sense. I really wanted a much thinner phone, but I'll likely pass on this one.

 

[dada_dave]

Would Geekbench scores provide evidence of this or is there another test which would be better?

I don’t know if GB uses matmul APIs in their GPU tests. The difference could show up, if they do.

To test the new units, the simplest thing would be to write some very simple compute shaders using the new Metal tensor APIs.



Well, the RTX 4070 offers 60 TFLOPs of FP16 matmul with FP32 accumulate and 120 TFLOPs with FP16 accumulate - and that is without scarcity. Would be great if the new Max Studio offered competitive performance.

One would imagine that GB AI GPU should?

 

[dada_dave]

Found it (though I haven’t had a chance to reread it to check that I’m not misremembering): https://www.thonking.ai/p/strangely-matrix-multiplications

Edit: I should also mention that Horace He (the author of that blog) isn’t some internet rando. He’s been a gem in ML and worked on PyTorch for years, and is now a founding engineer at Thinking Machines which has gathered an extraordinarily talented crew. He posted about it here: https://www.thonking.ai/p/why-pytorch-is-an-amazing-place-to

Edit2: The penalty for real data is not nearly as bad as I’d remembered (my apologies). Still a very interesting read, though.

So flops per watt still important for GPUs even on the high end ... that could be interesting ...

 

[Altaic]

I'm a little bummed the 17 Air doesn't have camera macro capability. Which now that I think about it, makes sense. I really wanted a much thinner phone, but I'll likely pass on this one.

I’m really torn between the Pro Max and Air... No macro, but do you know what the minimum in-focus distance would be for the Air? I’m a bit weak with camera/optics stuff.

 

[Citysnaps]

I’m really torn between the Pro Max and Air... No macro, but do you know what the minimum in-focus distance would be for the Air?

That I don't know. But I *think* decent macro ability and close focusing needs a relatively wide angle lens.

The thinness of the 17 Air is what really grabbed me when months ago it became apparent Apple was going to make it. I like carrying my silicone-cased 16PM in the pocket of certain T-shirts I like to wear when I'm out and about making photos. But it's such a pain in the ass getting the phone in and out of the pocket. It's just too tight (not due to phone width, but rather thickness).

 

[Edd]

I’m quite surprised about the 17 ProMotion in addition to significant battery life improvement and base storage over the 16. Got me thinking about trading my 16 in but that would be a crazy move for me. I always keep a phone for at least 4 years.

 

[Jimmyjames]

Some of the first Geekbench scores:
CPU https://browser.geekbench.com/v6/cpu/13736665

GPU https://browser.geekbench.com/v6/compute/4765303

 

[dada_dave]

Some of the first Geekbench scores:
CPU https://browser.geekbench.com/v6/cpu/13736665
View attachment 36489
GPU https://browser.geekbench.com/v6/compute/4765303
View attachment 36488

If those are representative, that’s a substantial GPU improvement and an okay CPU upgrade over the A18 Pro.

 

[throAU]

If those are representative, that’s a substantial GPU improvement and an okay CPU upgrade over the A18 Pro.

That metal score is nuts.

Pretty sure my M4 Max 40 core GPU is under 200k for metal.