Thread: iPhone 15 / Apple Watch 9 Event

To be fair some of the thermal results appear toasty if accurate. But yeah.
Indeed.

My impression is that the new u-arch appears to be more aggressively performance-focused than the 5nm products. Instead of trying to improve CPU performance at the same power draw, they go for maintaining perf/watt and opening up the upper range. We could be seeing a desktop-focused design here. I am less worried about the increase in power consumption at peak, as the phone appears to manage the total energy levels just fine and overall system performance looks much better with comparable or lower power draw in sustained scenario (e.g. the gaming benchmark).
Well, we have been talking about the scalability of the clock speed for Apple's core designs for a while here, which isn't as important on a phone but is for Macs or other devices that have the thermal envelope to sustain higher clocks. The Mac Studio has a massive cooling system, for instance. And I guess if you have a chip that can get up to higher clock speeds, it's not unreasonable to allow the clocks to go up for a brief period of time on the iPhone, for short bursty loads. Which might be the reason the clocks go down to 3.46GHz after several minutes of benchmarks. In any case, the power figures, if accurate, are massive. 14W for a CPU benchmark alone?

I mentioned this at the other place, but this figure from the video:

Screenshot 2023-09-20 at 09.53.14 copy.png


Looks pretty decent to me. At the same performance as the A16 Bionic, the A17 Pro is able to shave off a couple watts of power (11.3 -> 9.4 W, approximately). And the A17 Pro is also able to scale much further (just look how far to the right the A17 Pro can go), which is what everyone wanted for the M line anyway. Maybe we're just getting too tangled up in the fact that the A17 Pro can get to ridiculous (for a phone) power consumption figures, but the performance curve has improved at any power draw level.
 
Last edited:
Another not-insignificant fact: it seems, from the video, that the A17 Pro efficiency cores are a much greater improvement over the A16 than the A16 ware over the A15. Let's not forget that the vast majority of the time the iPhone is not running it's cores at full speed, 100% usage, but quite the opposite. So the overall efficiency of the SoC is hardly the performance per watt figure at 3.78GHz.

Screenshot 2023-09-20 at 10.08.31 copy.png
Screenshot 2023-09-20 at 10.09.16 copy.png


Interestingly enough, despite this improvements, the A17 Pro multicore score improved less than the single score, while an improvement of the E cores should boost the multicore score higher than what the single core improvement indicates. Maybe Apple is limiting the power draw when all cores are running?
 
In any case, the power figures, if accurate, are massive. 14W for a CPU benchmark alone?

Yeah, I dint quite understand this one. I suppose this is multi-core with aggressive cooling? If so, it merely tells us that Apple doesn’t throttle P-cores under load, that’s it. This figure is kind of meaningless without seeing the sane setup for other devices.

Another not-insignificant fact: it seems, from the video, that the A17 Pro efficiency cores are a much greater improvement over the A16 than the A16 ware over the A15.

This is approaching Akser Lake’s E-cores in some benchmarks, which is very impressive indeed.

Interestingly enough, despite this improvements, the A17 Pro multicore score improved less than the single score, while an improvement of the E cores should boost the multicore score higher than what the single core improvement indicates. Maybe Apple is limiting the power draw when all cores are running?

Reminds me of this: https://eclecticlight.co/2022/01/03/power-frequency-management-how-m1-e-cores-win/

Could very well be that Apple is limiting the frequency of E-cores under full-CPU load.
 
Yeah, I dint quite understand this one. I suppose this is multi-core with aggressive cooling? If so, it merely tells us that Apple doesn’t throttle P-cores under load, that’s it. This figure is kind of meaningless without seeing the sane setup for other devices.
I think you've seen my reply on the other site, but I have no idea how some of the data points were measured:
- How was that extremely low multicore score for the A17 Pro measured? It can't be Low Power Mode alone, it's 30% worse than the Low Power Mode score of the A16 or A15.
- How was the A17 Pro point in the middle measured? Can't be a regular run, it barely matches the A16 score.
- How was the highest-scoring A17 Pro measured? If it's a regular run, wow, it uses a massive amount of power. But everything in that graph is so weird, so I don't know whether to trust it.

This is approaching Akser Lake’s E-cores in some benchmarks, which is very impressive indeed.
I just learned today that on Low Power Mode the iPhone shuts down the P cores entirely, and runs on the E cores alone. That's impressive. I wonder if this is eventually going to lead to a 3-tier layout. After all, if the E cores alone are enough to run everything in the OS, wouldn't it make sense to have slower, but even more efficient cores to run background / low priority tasks?

EDIT: Not sure cores are actually disabled now, I’ve only seen that statement in a few places, so now I think maybe it’s not true after all. Idk.

Could very well be that Apple is limiting the frequency of E-cores under full-CPU load.
It could very well be that, particularly if those power figures are accurate.
 
Last edited:
How about waiting until the iPhone 15 Pro Max and the rest are actually out in the wild? These supposed performance evaluations before release are always extremely suspect.
 
How about waiting until the iPhone 15 Pro Max and the rest are actually out in the wild? These supposed performance evaluations before release are always extremely suspect.
While geekerwan might’ve rushed their review to do everything they did before the embargo lifted, they have an actual review unit from Apple. They are a Chinese reviewer.
 
IIRC gfxbench scores show the 740 behind Apple as well as geekbench. I haven’t seen any 3dmark tests for the a17. I forget it geekwerwan did those.
I think gfxbench is in the geekerwan video and from a couple of other sites I likewise found the 740 outpaced the A17, but by a smaller margin that you’d think for the difference in TFLOPs. Geekbench is hard to compare because of the Metal vs OpenCL vs Vulkan API results. However earlier I said that the GB OpenCL results within themselves were all over the map which is my distinct memory, but looking at them again and they seem pretty rational so I may have been simply wrong on that front or confusing those problems with trying to compare Vulkan vs Metal scores? Not sure. Regardless it may be possible to do some sort of conversion between the various scores but it’s difficult and even if the OpenCL results are in general okay, the drivers for any given device may not be well optimized for essentially a legacy API.

The Adreno architecture is unknown but a very old Anandtech article (Anand is on the byline) states that it is somewhere between TBDR and IMR. But that was over a decade ago and even the article is less than clear about what they meant exactly. It should be noted Nvidia does some kind of hybrid these days too - while i don’t think they do all the pre-rendering work of a full TBDR solution, since Maxwell apparently they do tile the rendering process more strongly than previously.
 
Last edited:
I think gfxbench is in the geekerwan video and from a couple of other sites I likewise found the 740 outpaced the A17, but by a smaller margin that you’d think for the difference in TFLOPs. Geekbench is hard to compare because of the Metal vs OpenCL vs Vulkan API results. However earlier I said that the GB OpenCL results within themselves were all over the map which is my distinct memory, but looking at them again and they seem pretty rational so I may have been simply wrong on that front or confusing those problems with trying to compare Vulkan vs Metal scores? Not sure. Regardless it may be possible to do some sort of conversion between the various scores but it’s difficult and even if the OpenCL results are in general okay, the drivers for any given device may not be well optimized for essentially a legacy API.
I’m a little cautious about the geekerwan review at the moment given their admission of rushing and some discrepancies. I just checked the gfxbench site and it showed the A17 beating the Adreno 730 in Aztec off screen and 4K Aztec so…I don’t know. I couldn’t find Adreno 740 benchmarks.
 
I’m a little cautious about the geekerwan review at the moment given their admission of rushing and some discrepancies. I just checked the gfxbench site and it showed the A17 beating the Adreno 730 in Aztec off screen and 4K Aztec so…I don’t know. I couldn’t find Adreno 740 benchmarks.
The 740 was a big uplift over the previous generation 730. Unfortunately you’re right that they don’t have the 740 results in the gfxbench database for some reason but I found gfxbench results for the 740 on other sites and they outpace the A17.
 
The Adreno architecture is unknown but a very old Anandtech article (Anand is on the byline) states that it is somewhere between TBDR and IMR.

From what I understand Adreno and Immortalis are TBR GPUs: they bin triangles into small tiles and draw the image tile by tile, but they don't to the deferred step like IMG/Apple. So basically forward rendering within a small tile. This simplifies the internal architecture (don't need to deal with the TBDR complexities) but still gets most of the benefits of cache and texture locality.
 
The 740 was a big uplift over the previous generation 730. Unfortunately you’re right that they don’t have the 740 results in the gfxbench database for some reason but I found gfxbench results for the 740 on other sites and they outpace the A17.
Interesting. Do you have any figures? I wonder if they use more power or are they just a better design?
 
From what I understand Adreno and Immortalis are TBR GPUs: they bin triangles into small tiles and draw the image tile by tile, but they don't to the deferred step like IMG/Apple. So basically forward rendering within a small tile. This simplifies the internal architecture (don't need to deal with the TBDR complexities) but still gets most of the benefits of cache and texture locality.
That’s supposedly what Nvidia does now as well. The only reason I’m hesitant is that the TBR description of the Qualcomm GPU is from over a decade ago. I couldn’t find anything recent.
 
Interesting. Do you have any figures? I wonder if they use more power or are they just a better design?
Power draw is what’s less clear to me too, Geekerwan is the first power draw estimates I’ve seen for it and according to them it is substantially better while using less power on a worse node - unironically big if true.

I gotta do something right now but I’ll try to collect some of the results I found scattered over the web.
 
I was reminded earlier of the review that Andrei did of the A15, and more specifically it’s gpu. He said something interesting about comparing game performance between Android phones and Apple ones.
1695227307451.png

I’d be interested to know if anyone is aware if things are more equal now? Are details and visual fidelity the same between Android and Apple gaming titles? I don’t believe this would effect benchmarks at all, but it could be worth considering when games are used as a means of comparison.
 
Power draw is what’s less clear to me too, Geekerwan is the first power draw estimates I’ve seen for it and according to them it is substantially better while using less power on a worse node - unironically big if true.

I gotta do something right now but I’ll try to collect some of the results I found scattered over the web.
I’ve been looking for benchmarks and as we said, it’s tough to find some (3DMark) and easier to find others.

I read this review earlier : https://www.tomsguide.com/features/iphone-15-pro-benchmarks

As I thought, there is a healthy lead in cpu scores, both single and multi core (76% and 41%) which I find hard to believe 8gen3 could make up for in one iteration (supposedly it gets 7400 in GB 6, a near 50% increase!). Back to the subject of gpus, here we see it’s much closer
1695227928091.png


We see the A17 beats the 8gen2 etc Android phones in the 1440p “Wildlife Unlimited” but is 1 or 2 fps behind then 8gen2 Android devices in the 4k “Wildlife Extreme” test.

If we compare some other scores, we see a similar pattern. GFXBench
In Aztec ruins offscreen, the A17 is ahead of the Adreno 730 by a large margin (60fps to 43fps).
Manhattan 1440p off-screen A17 = 109fps Adreno 730 = 89fps
Manhattan 1080p off-screen: A17 = 201fps, Adreno 730 = 140fps
T-Rex 1080p off-screen: A17 = 473fps, Adreno 730 = 381fps

BUT...

4K Aztec: A17 = 26fps, Adreno 730 = 20fps is what’s listed on the site, but I’ve seen higher than that. Close if not higher than the A17 iirc.

So what is it about 4K tests that causes the A17 to struggle vs the 730, but win at 1080/1440p? Design the gpu, the lower amount of ram?
 
I’ve been looking for benchmarks and as we said, it’s tough to find some (3DMark) and easier to find others.

I read this review earlier : https://www.tomsguide.com/features/iphone-15-pro-benchmarks

As I thought, there is a healthy lead in cpu scores, both single and multi core (76% and 41%) which I find hard to believe 8gen3 could make up for in one iteration (supposedly it gets 7400 in GB 6, a near 50% increase!). Back to the subject of gpus, here we see it’s much closer
View attachment 26027

We see the A17 beats the 8gen2 etc Android phones in the 1440p “Wildlife Unlimited” but is 1 or 2 fps behind then 8gen2 Android devices in the 4k “Wildlife Extreme” test.

If we compare some other scores, we see a similar pattern. GFXBench
In Aztec ruins offscreen, the A17 is ahead of the Adreno 730 by a large margin (60fps to 43fps).
Manhattan 1440p off-screen A17 = 109fps Adreno 730 = 89fps
Manhattan 1080p off-screen: A17 = 201fps, Adreno 730 = 140fps
T-Rex 1080p off-screen: A17 = 473fps, Adreno 730 = 381fps

BUT...

4K Aztec: A17 = 26fps, Adreno 730 = 20fps is what’s listed on the site, but I’ve seen higher than that. Close if not higher than the A17 iirc.

So what is it about 4K tests that causes the A17 to struggle vs the 730, but win at 1080/1440p? Design the gpu, the lower amount of ram?
For the 740 here’s what I found:


Basically matching or beating the A17 in Aztec Ruins at both high(1440p)/4K resolutions - for older tests with even lower resolutions the A17 pulls ahead. Unclear why lower resolution favors the iPhone here, you’re possibly right about RAM but I haven’t checked.

Interestingly they do have power consumption on Aztec normal for the Adreno 740 (all the way at the bottom), but the most recent iPhone chip you can compare to is the A15 (which is better).
 
For the 740 here’s what I found:


Basically matching or beating the A17 in Aztec Ruins at both high(1440p)/4K resolutions - for older tests with even lower resolutions the A17 pulls ahead. Unclear why lower resolution favors the iPhone here, you’re possibly right about RAM but I haven’t checked.

Interestingly they do have power consumption on Aztec normal for the Adreno 740 (all the way at the bottom), but the most recent iPhone chip you can compare to is the A15 (which is better).
Many thanks. That’s great. I find their layout a little confusing but that’s probably just me! Overall 29fps in 4k Aztec is not as good as I thought it would be given the boasting. It is higher than the A17 though so credit where it’s due. I’ll need to read it more closely.
 
Many thanks. That’s great. I find their layout a little confusing but that’s probably just me! Overall 29fps in 4k Aztec is not as good as I thought it would be given the boasting. It is higher than the A17 though so credit where it’s due. I’ll need to read it more closely.
It’s not a great layout. Oh the Geekbench results are wrong on that page - it quoted its Vulkan score as mid 3K - 1/2 the rate of Adreno 730. Obviously incorrect. I saw other sites say it is around 10K.
 
It’s not a great layout. Oh the Geekbench results are wrong on that page - it quoted its Vulkan score as mid 3K - 1/2 the rate of Adreno 730. Obviously incorrect. I saw other sites say it is closer to 10K.
Ok thanks. That geekbench still doesn’t seem that great.
 
Back
Top