Apple is falling behind, M4 Max nowhere close to RTX 5090.

[maflynn]

This article is a good summary, in particular the figure 3: https://www.arepa.com/media/zi2lbdvf/arepa-whitepaper-extending-life-expectancy.pdf

Thanks for the link, I reviewed the article, and there does seem to be empiracal evidense to support that temps 105c does not increase the failure rates.

I'll still hold onto the notion that for me, I do not want to see anything above 80c, but that's my personal bias.

 

[dmccloud]

It IS less efficient. You are getting a more powerful CPU with more power and heat. Nothing wrong with that; Intel does it too. But in our context, an increase in efficiency means doing more with less power (which is what Apple markets with their Apple Silicon processors).

If a new M series processor consumes 5% more power but provides a 10% increase in performance, it is by definition more efficient. The efficiency is measures in performance per watt, not how you're trying to calculate it.

 

[diamond.g]

Thanks for the link, I reviewed the article, and there does seem to be empiracal evidense to support that temps 105c does not increase the failure rates.

I'll still hold onto the notion that for me, I do not want to see anything above 80c, but that's my personal bias.

It is an interesting read. I was always under the impression that the temps matter more for the VRM and caps than anything else. Does anyone know much about how Apple lays out their power stages/delivery?

 

[leman]

I'll still hold onto the notion that for me, I do not want to see anything above 80c, but that's my personal bias.

That’s very fair. Of course, with an Apple system you have limited possibilities of controlling this.

By the way, policy of Puget Systems of Running CPUs below 85C also makes sense. They run servers, so likely thousands of CPUs in close proximity. Lowering the temperature somewhat will probably prevent a few faults per year and help them save on maintenance. Similar considerations are less relevant to the end consumer.

 

[cbum]

I'll still hold onto the notion that for me, I do not want to see anything above 80c, but that's my personal bias.

Aren't you and leman just pointing to different locations in the computer?
IMHO, solid data that CPU at 105C are ok and the general maxim that heat and heat/cooling cycles are bad for electronics can co-exist, no?

 

[novagamer]

Aren't you and leman just pointing to different locations in the computer?
IMHO, solid data that CPU at 105C are ok and the general maxim that heat and heat/cooling cycles are bad for electronics can co-exist, no?

Yes and to be fair, this has changed a lot over the last 25 years. If you come from the PC world many years ago there was no thermal throttling capability and you'd destroy your CPU if you did something like leave the heatsink off. Socketed AMD Athlons would literally pop.

Once thermal throttling was introduced and worked well, like with the Pentium 4 e.g. the Willamette, the ceiling was lower and sometimes throttling would start to take effect at ~75c and continue until you could basically run them passively at around 80-95c at significantly reduced performance. Conversely many newer PC CPUs don't throttle until nearly 100c.

Other modern products do not follow this paradigm at all and are often designed to run hotter as others have said. Part of the Radeon VII (the junction) ran at 107-110c as intended which even now seems ludicrous but those cards had no higher rates of failure.

It's somewhat understandable to have a baked-in preference, particularly if you were around for the older days of enthusiast PC building, but those old adages can't be reliably applied to modern technology.

The linked papers are interesting, but on the whole both leman and you are correct, hot/cold cycling is never good for electronics and the thermal characteristics / tolerances have changed fairly substantially over the years.

 

[M4pro]

Tuesday’s iPhone event played out in a best-case-scenario kind of way for Apple.

A19 Pro advances Apple GPU capabilities significantly.

Running large language models and gaming are now given the same importance as video editing in Apple’s marketing language.

“A19 Pro enables iPhone 17 Pro and iPhone 17 Pro Max to deliver up to 40 percent better sustained performance than the previous generation — ideal for gaming, video editing, and running large local language models.

…the 6-core GPU architecture includes Neural Accelerators built into each GPU core…

The GPU works in tandem with the new 16-core Neural Engine to power Al models, stunning graphics, and AAA gaming titles like Arknights: Endfield, enabling hardware-accelerated ray tracing and higher frame rates.”

 

[PaulD-UK]

Quote: “…the 6-core GPU architecture includes Neural Accelerators built into each GPU core…
The GPU works in tandem with the new 16-core Neural Engine to power Al models…”

How does that work?
Does each GPU core have 16/6=2.667 Neural Engine cores, or does each have 16 NE cores, for a total of 96 NE cores?

Or is there an old style 16 core NE somewhere else plus what’s in each GPU (which is an unspecified number)?

 

[M4pro]

There’s new "Neural Accelerators" (Apple’s marketing term) on each GPU core.

Plus an updated version of Apple’s seperate 16-core Neural Engine.

They handle different tasks.

 

[dmccloud]

Quote: “…the 6-core GPU architecture includes Neural Accelerators built into each GPU core…
The GPU works in tandem with the new 16-core Neural Engine to power Al models…”

How does that work?
Does each GPU core have 16/6=2.667 Neural Engine cores, or does each have 16 NE cores, for a total of 96 NE cores?

Or is there an old style 16 core NE somewhere else plus what’s in each GPU (which is an unspecified number)?

The Neural accelerators are not the same thing as the Neural Engine. ANE has its own section on the die, while the accelerators are paired 1:1 with each GPU core. So there are 16 ANE cores plus the 6 accelerators.

 

[leman]

The Neural accelerators are not the same thing as the Neural Engine. ANE has its own section on the die, while the accelerators are paired 1:1 with each GPU core. So there are 16 ANE cores plus the 6 accelerators.

Technically 24, since each GPU has four compute partitions. Of course, the term “neural accelerator” itself is more of a marketing gimmick.

 

[maflynn]

that the temps matter more for the VRM and caps than anything else.

I agree, and I've said as much in various threads, that while the CPU/GPU may be sturdy to handle 100c, the VRMs and. capacitors seem more fragile. SSDs seem to be falling a bit into that category as well, but I don't know the thermal threshold for them

 

[MayaUser]

Physics ..

 

[MayaUser]

 

[diamond.g]

Physics ..View attachment 2550507

Wonder if the Air is lower clocked than the others. GB is a bursty test so it should be thermally throttling that fast.

EDIT: shouldn't not should

 

[OptimusGrime]

Physics ..View attachment 2550507

Well…that’s the problem with a high variability test like GB. This score is much better

iPhone18,4 - Geekbench

Benchmark results for an iPhone18,4 with an ARM processor.

browser.geekbench.com

 

[M4pro]

iPhone benchmarks are a useful conversation starter. They’re kind of the pre-game before the party starts.

Of course no one knows anything about the M5 gen GPU until a Russian cartel steals a pallet of MacBook Pros from a warehouse and posts a technical brief to YouTube.

 

[leman]

Of course no one knows anything about the M5 gen GPU until a Russian cartel steals a pallet of MacBook Pros from a warehouse and posts a technical brief to YouTube.

It's the same GPU, so technical data and expected performance can be easily extrapolated from A19

 

[M4pro]

It's the same GPU, so technical data and expected performance can be easily extrapolated from A19

100 percent.

(And last year’s M4 MBP Russian palette pirate leak remains funny a year later)

 

[OptimusGrime]

Very nice RT improvement.

 

[dmccloud]

Technically 24, since each GPU has four compute partitions. Of course, the term “neural accelerator” itself is more of a marketing gimmick.

So not really much different from Nvidia's Tensor Cores...

 

[Homy]

Don't forget to provide the source to make it easier to check, especially when it's in Chinese:

 

[Boil]

Russian palette pirate

LOL, just thinking about that on a resume... ;^p

 

[leman]

Don't forget to provide the source to make it easier to check, especially when it's in Chinese:

That’s some very impressive performance AND efficiency improvements there! GPU gains are very good, but the CPU ones are really unexpected. I suppose all the work they’ve been doing with smarter execution really starts to pay off!

 

[OptimusGrime]

From the Geekerwan review. Big uplift in ray tracing and gaming performance.