Sept 2025 Apple iPhone Event

[dada_dave]

But to my knowledge you cannot request device. You can only request which of several Metal devices counts as the Metal device in the group of 3 to choose from, CPU, ANE and MetalDevice

I’m not sure I follow the difference.

 

[leman]

I guess this means we are entering app upgrade season where every app will need to be updated.

How come?

 

[Herdfan]

How come?

Every new version of iOS seems to require updates to the apps.

I don’t mind updating apps, it’s just they tend to require it at the most in inopportune times.

 

[casperes1996]

I’m not sure I follow the difference.

There are three units to pick from when spawning a CoreML workload.

1) CPU
2) ANE
3) Metal Device (AKA GPU)

CoreML will dispatch work to any one of these as it sees fit. You may be in a Mac Pro with a Xeon CPU, no ANE, 1 Radeon Pro 580X and one Radeon RX 6800XT.

The API only allows you to pick which of the two GPUs CoreML*might* use. It might still instead use the CPU. This API also interacts with eGPUs on Intel Macs to pick whether the internal dGPU, iGPU or eGPU is used as the Metal Device. But it might still instead dispatch to CPU or on AS, ANE.

CPU and ANE are not Metal Devices, and the API I've seen specifically sets "preferredMetalDevice" to an object of type MTLDevice.

 

[dada_dave]

There are three units to pick from when spawning a CoreML workload.

1) CPU
2) ANE
3) Metal Device (AKA GPU)

CoreML will dispatch work to any one of these as it sees fit. You may be in a Mac Pro with a Xeon CPU, no ANE, 1 Radeon Pro 580X and one Radeon RX 6800XT.

The API only allows you to pick which of the two GPUs CoreML*might* use. It might still instead use the CPU. This API also interacts with eGPUs on Intel Macs to pick whether the internal dGPU, iGPU or eGPU is used as the Metal Device. But it might still instead dispatch to CPU or on AS, ANE.

CPU and ANE are not Metal Devices, and the API I've seen specifically sets "preferredMetalDevice" to an object of type MTLDevice.

Sorry by each device, I meant the simpler case of:

CPU (AMX) (technically host, at least per Nvidia terminology)
ANE
GPU (Metal device)

Rather than the more complicated case of multiple metal devices. I’ll use the term resource to describe them from now on. My presumption, perhaps wrong, is that GB in its AI benchmark is relying on CoreML to not redirect its workload from the requested resource for the models they choose to run. Because they definitely use the CoreML API and output results for the CPU, ANE, and GPU and as far as I can tell those results are consistent.

This is vaguely reminiscent of benchmarkers testing E-cores vs P-cores using Apple’s thread priority levels. Technically the OS could decide “nah going to run this on the completely different core”, but in practice it isn’t going to. Andrei from Anandtech relied heavily upon this “sensible” behavior of the OS/API.

 

[DT]

OK, looks like (all 256GB):

iPhone 17, Sage
iPhone Air, Sky Blue
iPhone Pro Max, Cosmic Orange

WOOT

 

[Jimmyjames]

I was tagged into this. It purports to show a 2.8x increase in FP16, 3.05x increase on int8 and a 60% uplift on FP32 on Geekbench AI between the 16 Plus and the Air. I'm a little suspicious giiven the version mismatch (1.4 vs 1.5). The notes say Windows on Arm cannot be compared between versions. There is a comparison between the 16 Pro Max and the Air which is only 2x for FP16, but the Pro Max has an extra core. Hmmm

 

[exoticspice1]

I was tagged into this. It purports to show a 2.8x increase in FP16, 3.05x increase on int8 and a 60% uplift on FP32 on Geekbench AI between the 16 Plus and the Air. I'm a little suspicious giiven the version mismatch (1.4 vs 1.5). The notes say Windows on Arm cannot be compared between versions. There is a comparison between the 16 Pro Max and the Air which is only 2x for FP16, but the Pro Max has an extra core. Hmmm

View attachment 36542

With Geekbench AI 1.5, we’ve focused on improving performance and compatibility with hardware. Here’s what’s new:

• Upgrade to ONNX Runtime 1.22.2 on Windows on ARM.

Due to these improvements, Geekbench AI 1.5 scores may be higher than Geekbench AI 1.4 scores on Windows on ARM devices, meaning Geekbench AI 1.5 scores aren’t strictly comparable with Geekbench AI 1.4 or earlier scores.
----------------
So we can still compare iOS GB AI1.4 to iOS GB AI1.5 then

 

[Alli]

All I know is for the first time in a while my alarm is set so I can pre-order in the morning.

 

[Jimmyjames]

With Geekbench AI 1.5, we’ve focused on improving performance and compatibility with hardware. Here’s what’s new:

• Upgrade to ONNX Runtime 1.22.2 on Windows on ARM.

Due to these improvements, Geekbench AI 1.5 scores may be higher than Geekbench AI 1.4 scores on Windows on ARM devices, meaning Geekbench AI 1.5 scores aren’t strictly comparable with Geekbench AI 1.4 or earlier scores.
----------------
So we can still compare iOS GB AI1.4 to iOS GB AI1.5 then

That was my hunch. Thanks.

 

[Altaic]

View attachment 36542

What the… that table-chart-chimera-thing is a monstrosity straight out of data science hell

 

[dada_dave]

Article complaining that Apple (and Samsung’s) marketing is misleading as the iPhone 17 Pro (and equivalent Samsung) don’t actually offer the optical zoom they claim but a digital zoom that is “optical quality” but this distinction isn’t actually mentioned in most of Apple’s marketing materials. For the most part they just say 8x optical zoom:

Apple's misleading iPhone 17 Pro camera marketing distracts customers from the real weaknesses

The Apple iPhone 17 Pro has received two significant camera upgrades - a 48 MP telephoto camera with a larger sensor and an 18 MP selfie camera with a square sensor. However, Apple has failed to address a major weakness of the previous model, while its marketing is more misleading than ever.

www.notebookcheck.net

 

[exoticspice1]

will Metal 4 take advantage of the new AI cores in A19? Or is it just for Core ML

 

[Altaic]

will Metal 4 take advantage of the new AI cores in A19? Or is it just for Core ML

Both will support the new GPU hardware.

 

[Andropov]

Article complaining that Apple (and Samsung’s) marketing is misleading as the iPhone 17 Pro (and equivalent Samsung) don’t actually offer the optical zoom they claim but a digital zoom that is “optical quality” but this distinction isn’t actually mentioned in most of Apple’s marketing materials. For the most part they just say 8x optical zoom:

Eh. I've complained about this myself in the past, as there's certainly nothing "optical" about the "8x" zoom. At the same time, people often misunderstand what camera specs mean, and saying that the new 8x option has "optical quality" is something people can understand while not being completely false. Apple increased the sensor size of the telephoto lens enough from the iPhone 16 Pro to the 17 Pro that a 50% crop of it (for the "8x" zoom) still uses a decent sensor area (11.75mm^2, which is almost as big as the entire telephoto sensor of the previous iPhone, which was 15.3mm^2). So I can see where Apple is coming from with the "optical quality" marketing.

 

[DT]

Oops.

Got a little f-ed over the inability to order 3 phones connected to carriers (via my Apple Card), so had to make 3 different orders, lots of re-auth over-and-over, the wife and daughter's phones slipped a week, no biggie, my big weird orange phone has a date range the following week.

 

[DT]

Side note: hopped anonymously (after a good year or more) on MR for a take on the new phones, good f-ing lord what a bunch of idiots

 

[Cmaier]

woke up at 5, and decided i really don’t need a new phone this year, even though i love orange.