A17 Pro interview

[pipo]

IGN had the opportunity to sit down and interview several members behind the development of the iPhone 15 Pro. Specifically, the interview dives into Apple's gaming ambitions for the iPhone 15 Pro and why the tech giant believes "it's going to be the best game console."

Apple Answers of All Our iPhone 15 Pro Questions: 'It's Going to be the Best Game Console' - IGN

The iPhone 15 is out on September 22 and it promises to deliver some unprecedented gaming capabilities, IGN sat down with Apple to dive deeper into the gaming features coming to the iPhone 15 Pro.

www.ign.com

Non-tech marketing push. But still, they're making an effort. The M3 (or next years version) and Apple TV strategy will be interesting too.

 

[Chiimaero]

Yes, not that many useful informations in there. I really miss the glorious "deep dive" articles from anandtech.

Also, some thoughts about Kaotik's off topic question (on the Immortalis thread) regarding the roots of the A17 Pro GPU. His question was "is it Img CXT based? or smth new?"
I find this question interesting, and sadly, there is less and less public documentation regarding mobile GPU these days.
One thing that may be a hint: the A17 Pro has Mesh Shader support. And I found no trace stating either CXT or DXT support them. The only claims I found are in fact in this very forum, with users questioning the feasability of Mesh Shader on a TBR arch.
Well, this question at least now has an answer: Apple did manage to implement the feature in their mobile GPU now. And I think Mesh Shader is a big enough change to back up Apple claiming A17 Pro GPU is their biggest GPU overhaul (this and RT support, though I don't know what implementation level Apple managed to achieve).

 

[Ailuros]

Yes, not that many useful informations in there. I really miss the glorious "deep dive" articles from anandtech.

Also, some thoughts about Kaotik's off topic question (on the Immortalis thread) regarding the roots of the A17 Pro GPU. His question was "is it Img CXT based? or smth new?"
I find this question interesting, and sadly, there is less and less public documentation regarding mobile GPU these days.
One thing that may be a hint: the A17 Pro has Mesh Shader support. And I found no trace stating either CXT or DXT support them. The only claims I found are in fact in this very forum, with users questioning the feasability of Mesh Shader on a TBR arch.
Well, this question at least now has an answer: Apple did manage to implement the feature in their mobile GPU now. And I think Mesh Shader is a big enough change to back up Apple claiming A17 Pro GPU is their biggest GPU overhaul (this and RT support, though I don't know what implementation level Apple managed to achieve).

Apple itself claims that the A17 Pro GPU has Apple designed shader architecture:

A17 Pro deep dive: A look at Apple's new iPhone 15 Pro chip

It's a powerhouse of a chip, packing some professional grade features worthy of the 'Pro' name.

www.zdnet.com

It's my understanding that since Apple introduced "custom designed GPUs" re-wrote almost entirely the ALUs. There are however still IMG specific extensions present in that GPU which tells me that some sort of the GPU backbone is left in there from the original IMG GPU IP and they probably just added a RAC unit for RT.

Since I followed the discussions you mention, if you want to revisit them, you should keep in mind that JohnH used to be one of the fathers of all GPU IP at IMG from SGX to Photon. There any random sw engineer trying to explain to a gentleman with as much experience and knowhow on TBDRs how his own architectures work, is usually a joke at best

 

[Chiimaero]

Apple itself claims that the A17 Pro GPU has Apple designed shader architecture:

A17 Pro deep dive: A look at Apple's new iPhone 15 Pro chip

It's a powerhouse of a chip, packing some professional grade features worthy of the 'Pro' name.

www.zdnet.com

It's my understanding that since Apple introduced "custom designed GPUs" re-wrote almost entirely the ALUs. There are however still IMG specific extensions present in that GPU which tells me that some sort of the GPU backbone is left in there from the original IMG GPU IP and they probably just added a RAC unit for RT.

That's the issue I have with Apple since the parted ways with IMG: we have no solid information about their GPUs. Them claiming they designed their GPU cannot be taken for granted as we have several hints showing the Apple GPU still have deep IMG roots (at least I found such hints up to the M1 GPU like here).
They obviously don't lie either, so, like you said, what they have now must be an IMG/Apple hybrid. Except we don't know what is the extend of the Apple part. I guess it was anectoctical back in the A9 GPU. Must be much more by now, but how much exactly?
I see Mesh Shader as an additional hint that the A17 Pro GPU is now more Apple than ever, and thus, less IMG than ever (since I think their is no existing IMG arch with Mesh Shader).
This leads me to think that the RT engine (I'd not call it a RAC since this is IMG naming) is maybe (just maybe) not based on the CTX RAC. We're left with so little info we can only guess...

Since I followed the discussions you mention, if you want to revisit them, you should keep in mind that JohnH used to be one of the fathers of all GPU IP at IMG from SGX to Photon. There any random sw engineer trying to explain to a gentleman with as much experience and knowhow on TBDRs how his own architectures work, is usually a joke at best

If I'm the random sw engineer (and btw, I'm precisely that, a random sw engineer ) , I feel a bit sorry you read my post this way as I did not mean to trigger any negative feeling. I mean I have zero pretentiousness to discard what was said there, be it from JohnH or any other less famous guy. I'm just noting that Apple claims to have Mesh Shader in their latest GPU which is (very likely) still TBDR.
This brings new factual data to the table regarding the usefulness of those paired techs. Does not mean old claims were false or poorly educated or anything like that! It just means that a few years later, Apple GPU arch may now be different enough for this to make sense. Or maybe newer implementations of Mesh Shader are more TBDR friendly. Or maybe Apple tried smth that will prove to be not that great in the long run.
If this can sparke some more insightful feedback from guys like JohnH or you, that's a plus in my book.

 

[Ailuros]

John is unfortunately not working at IMG for quite some time now and it's even more unfortunate that no IMG hw engineer is here these days to give any form of insight on their architecture (and by the way there was no negative vibe or feeling while reading your former post nor in my reply to it either).

For Mesh shaders my gut feeling tells me that it wasn't amongst IMG's engineering priorities so far; au contraire to variable rate shading (IMG calls it FSR I think) which is now present in DXT but Apple doesn't seem to support it in their latest GPUs or they do and no one has noticed so far? Somewhat unlikely.

For the record there had been also opinions I read in the past that VRS doesn't cope well with deferred renderers either. While I'm by far not an expert I miss the reasoning why a DR wouldn't digest mesh shaders well. It's not really a secret that deferred renderers aren't as "ideal" with geometry as immediate mode renderers are, but IMG has always found ways to workaround wherever their architecture wouldn't bode well, but that's something that isn't uncommon for IMRs either. Today's GPUs defer when it's a necessity and either way it's possible to combine the best of both worlds; first step for IMRs was early Z.

For Apple's GPU specifically I'd dare to believe them when they say that they have custom designed ALUs in their recent GPU generations. Too deep changes in the IP architecture itself would bring them more problems IMHO and after some point it would bear the question why they still pay IMG licensing fees. The rumor mill had it years before Apple parted from IMG, that Apple wasn't satisfied with IMG's roadmap. Didn't come as a surprise for me either back then when I heard it since QCOM's Adreno GPUs had reached an insane "GFLOPs/mm2 ratio" from which IMG's solutions were a few miles apart so to speak. No wonder Apple decided to revamp ALUs almost from ground up and most likely ended up with something that wouldn't have to hide even against IMG Albiorix ALU advancements.

For the RT implementation, the last official announcement was 3 years ago:

Imagination and Apple Sign New Agreement - Imagination

Apple signs new IP licence deal with Imagination Technologies

www.imaginationtech.com

London, UK; 2nd January 2020 – Imagination Technologies (“Imagination”) announces that it has replaced the multi-year, multi-use license agreement with Apple, first announced on February 6, 2014, with a new multi-year license agreement under which Apple has access to a wider range of Imagination’s intellectual property in exchange for license fees.

We still don't have a clue what they meant with that wider range of IP, but it very well could have included IMG RT implementation amongst other possible tidbits.

 

[Ailuros]

The Apple "Scary Fast" Event Live Blog (Starts at 5pm PT/00:00 UTC)

www.anandtech.com

So this brings the M3 up to par with the A17. And in PC terms, would be roughly DirectX12 Ultimate (feature level 12_2)

I don't think it'll be easy to find out, and not that it really matters in the "Apple world" but a GPU is either DX12 compliant or it isn't. It could even miss some DX11 requirements and no one would actually care.

 

[Arnold Beckenbauer]

It's still a smartphone.

 

[Chiimaero]

John is unfortunately not working at IMG for quite some time now and it's even more unfortunate that no IMG hw engineer is here these days to give any form of insight on their architecture (and by the way there was no negative vibe or feeling while reading your former post nor in my reply to it either).

For Mesh shaders my gut feeling tells me that it wasn't amongst IMG's engineering priorities so far; au contraire to variable rate shading (IMG calls it FSR I think) which is now present in DXT but Apple doesn't seem to support it in their latest GPUs or they do and no one has noticed so far? Somewhat unlikely.

For the record there had been also opinions I read in the past that VRS doesn't cope well with deferred renderers either. While I'm by far not an expert I miss the reasoning why a DR wouldn't digest mesh shaders well. It's not really a secret that deferred renderers aren't as "ideal" with geometry as immediate mode renderers are, but IMG has always found ways to workaround wherever their architecture wouldn't bode well, but that's something that isn't uncommon for IMRs either. Today's GPUs defer when it's a necessity and either way it's possible to combine the best of both worlds; first step for IMRs was early Z.

For Apple's GPU specifically I'd dare to believe them when they say that they have custom designed ALUs in their recent GPU generations. Too deep changes in the IP architecture itself would bring them more problems IMHO and after some point it would bear the question why they still pay IMG licensing fees. The rumor mill had it years before Apple parted from IMG, that Apple wasn't satisfied with IMG's roadmap. Didn't come as a surprise for me either back then when I heard it since QCOM's Adreno GPUs had reached an insane "GFLOPs/mm2 ratio" from which IMG's solutions were a few miles apart so to speak. No wonder Apple decided to revamp ALUs almost from ground up and most likely ended up with something that wouldn't have to hide even against IMG Albiorix ALU advancements.

For the RT implementation, the last official announcement was 3 years ago:

Imagination and Apple Sign New Agreement - Imagination

Apple signs new IP licence deal with Imagination Technologies

www.imaginationtech.com

We still don't have a clue what they meant with that wider range of IP, but it very well could have included IMG RT implementation amongst other possible tidbits.

Thx for those insights.
It seems that Apple GPUs do actually support VRS, at least according to this blog VRS on Adreno
Those recent mobile GPUs are packed with all the latest features now. It will be harder and harder to improve on them.

The Apple "Scary Fast" Event Live Blog (Starts at 5pm PT/00:00 UTC)

www.anandtech.com

I don't think it'll be easy to find out, and not that it really matters in the "Apple world" but a GPU is either DX12 compliant or it isn't. It could even miss some DX11 requirements and no one would actually care.

Yeah, DX12_2 or not, the GPU is incredibly feature rich and tailored for Apple use.
The most intriguing part regarding the M3 family is the cut down NPU. Given most of the innovation in the gaming tech world are tied to AI infused stuff lately, I find it weird that those chips come with reduced perfs on this front.
A17 Pro getting a very capable NPU enabling MetalFX "DLSS-like" upscaling techniques was a pleasant surprise.

It's still a smartphone.

It's definitely still a smartphone with its 6-8w budget. And we're definitely reaching some hard physical/economical barrier with process nodes. Optimization can still do wonders (looking at you 8G3 GPU monster). In fact I'm still puzzled as to how Qualcomm managed to deliver more perfs per watt given they operate with a supposedly worse process node AND larger quantities of RAM and faster RAM. This should translate into more wattage, but it seems that whatever the use case, the 8G3 is more power efficient than everything else, including A17 Pro, and 8G2, which was already quite an optimization marvel.
I guess optimization can only do that much though, and that most big vendors will have pretty much the same level of performance and features very soon. Boring times ahead... or maybe we'll witness some new wizardry!

 

[Pressure]

The most intriguing part regarding the M3 family is the cut down NPU. Given most of the innovation in the gaming tech world are tied to AI infused stuff lately, I find it weird that those chips come with reduced perfs on this front.
A17 Pro getting a very capable NPU enabling MetalFX "DLSS-like" upscaling techniques was a pleasant surprise.

The Neural Engine isn't cut down. In fact looking at the die shot it takes up more space on the M3 family compared to the A17 Pro.

The numbers most likely differ because they aren't referring to the same workload but we will know more soon enough.

 

[Chiimaero]

The Neural Engine isn't cut down. In fact looking at the die shot it takes up more space on the M3 family compared to the A17 Pro.

The numbers most likely differ because they aren't referring to the same workload but we will know more soon enough.

I'm honestly not that sure: A17 Pro vs M3
But I keep in mind those annotations are often not 100% accurate. On the pictures, M3 vs M3 Pro in particular seems to highlight less blocks (8 cores for the M3 vs 8 cores + some shared logic for M3 Pro)
And using different workloads to measure a given performance metric does not sound like a good idea. One possible explanation I read about this supposedly weaker NPU is that the larger M3 could dedicate more GPU power to AI tasks if needed. Well, I guess that's a valid point as long as what's in there is enough to deal with the more advanced MetalFX stuff they plan to implement in the 4-5 years to come.

Another note regarding GPU performance that is not often discussed is driver quality. In the PC space, this was once a huge topic of discussion with NVidia delivering supposedly the best drivers. Lately, we've seen Intel pushing updates bringing more and more performances for their Alchemist lineup.
Could driver quality explain how good the Adreno GPU sound right now vs the competition? In short, do we have an idea as to what is the quality of drivers from the major vendors? The only hints I've seen are from Samsung sadly, who does not update their drivers that often and who seem to lack conformance here or there. But what about Apple? ARM?

 

[pcchen]

I'm honestly not that sure: A17 Pro vs M3

I'm curious why do you think it's "cut down"? NPU is a very small part of M3 and I doubt there's much gain from reducing it.

 

[Chiimaero]

I'm curious why do you think it's "cut down"? NPU is a very small part of M3 and I doubt there's much gain from reducing it.

I was referring to Apple perfs numbers here. A17 Pro is 35Tops, M3 is 18Tops.
Like Pressure said, maybe it's just Apple messing around with their own convention here. But since my last post I've read Anandtech take on this, and they seem to think it simply a matter of INT8 vs INT16/FP16 so there may be no cut down...
Or maybe there is (sort of)... we know M3 GPU has the same arch as A17 Pro because of RT and Mes Shader support. But as Anantech stated, maybe the NPU is untouched, based on A16, and incapable of doubled INT8 (thus the lower figure relative to A17). By the way the CPU may be still based on A16 as well given the higher clock speed but similar GB6 score.
Another weird thing if that's the case... would mean M3 has 3nm and GPU from A17 but CPU and NPU from A16.

 

[Pressure]

I was referring to Apple perfs numbers here. A17 Pro is 35Tops, M3 is 18Tops.
Like Pressure said, maybe it's just Apple messing around with their own convention here. But since my last post I've read Anandtech take on this, and they seem to think it simply a matter of INT8 vs INT16/FP16 so there may be no cut down...
Or maybe there is (sort of)... we know M3 GPU has the same arch as A17 Pro because of RT and Mes Shader support. But as Anantech stated, maybe the NPU is untouched, based on A16, and incapable of doubled INT8 (thus the lower figure relative to A17). By the way the CPU may be still based on A16 as well given the higher clock speed but similar GB6 score.
Another weird thing if that's the case... would mean M3 has 3nm and GPU from A17 but CPU and NPU from A16.

No, it's definitely a wider CPU than what's in the A16. They even showed it in the event.

M3 is the same core design as the A17 Pro.

 

[Chiimaero]

No, it's definitely a wider CPU than what's in the A16. They even showed it in the event.

M3 is the same core design as the A17 Pro.

Most M3 articles I've read are not so clear about that.
M2 is based on A15 CPU cores (Avalanche/Blizzard). M3 has definitely better cores than M2, but it's still possible they are based on A16 cores (Everest/Sawtooth), and not on A17 Pro cores (which have weird generic code names PCORE_Coll/ECORE_Coll). Since we have no official word on that matter, I guess there is still the possibility that M3 has the same CPU core as A17 as you say.
I used the first M3 GB6 single core scores, which are pretty much in line with A16, not A17 as a clue.
But now it seems the first M3 Max GB6 score is quite a bit higher that base M3 (and in line with A17)? Base M3 and M3 Max having different scores makes no sense, so yeah, everything is possible it seems.

Anyway, M3 has the same GPU as the A17 Pro, that much is mostly a given. This should give us some more clues as to how capable the arch really is when compared to other state of the art efforts.
Because if we stick to the phone market, it seems the A17 Pro will in fact rock the less powerfull (and possible less power efficient) GPU of all flagship SoCs.
We already know that 8G3 and 8G2 are respectively way better and a bit better. Samsung claimed that their Exynos 2400 will beat Apple in the GPU department (with its rdna3 based solution), and first benchmarks of Dimensity 9300 (Mali Immortalis G720) also put it way ahead of the A17 Pro.
Weird times when the latest and greatest Apple product is just sub-par vs the competition... obviously only for the GPU here, their CPU cores are still undisputed for this generation. This may change with 8G4 though. And at this time, I guess all mobile vendors will reach some kind of performance plateau within 2 to 3 years.

 

[Pressure]

Most M3 articles I've read are not so clear about that.
M2 is based on A15 CPU cores (Avalanche/Blizzard). M3 has definitely better cores than M2, but it's still possible they are based on A16 cores (Everest/Sawtooth), and not on A17 Pro cores (which have weird generic code names PCORE_Coll/ECORE_Coll). Since we have no official word on that matter, I guess there is still the possibility that M3 has the same CPU core as A17 as you say.
I used the first M3 GB6 single core scores, which are pretty much in line with A16, not A17 as a clue.
But now it seems the first M3 Max GB6 score is quite a bit higher that base M3 (and in line with A17)? Base M3 and M3 Max having different scores makes no sense, so yeah, everything is possible it seems.

"The CPU cores feature improved branch prediction for greater efficiency, and the performance cores have a wider decode and execution engine, while the efficiency cores have a deeper execution engine, all of which boost overall performance and efficiency."

I mean, that sounds exactly like the A17 Pro, although I will concede that the A16 cores line up with the ~18% faster clocks of the M3 but I don't think Geekbench scales linear like that ...

That would however also mean that they ported the A16 "Everest" CPU core from TSMC N4P to N3B, which seems weird when they have the A17 Pro CPU core ready for that process.

We will know soon enough

 

[Chiimaero]

That would however also mean that they ported the A16 "Everest" CPU core from TSMC N4P to N3B, which seems weird when they have the A17 Pro CPU core ready for that process.

We will know soon enough

Precisely, but well, they likely ported the A16 NPU to N3B as well when they already had the updated A17 Pro NPU at hands. It will be very intersting to learn the weird schedule and decisions that led to M3 being such an hybrid beast.

 

[Pressure]

Precisely, but well, they likely ported the A16 NPU to N3B as well when they already had the updated A17 Pro NPU at hands. It will be very intersting to learn the weird schedule and decisions that led to M3 being such an hybrid beast.

Found the picture from the presentation.

 

[Chiimaero]

Found the picture from the presentation.

View attachment 9957

Indeed. Will wait for deep dive reviews, but as you say, it seems this M3 is more A17 Pro than A16, which makes perfect sense, but Apple failed to convey this in a clear and simple way. At this point it's very likely the NPU is also on par with A17 Pro I guess... That's a good thing as I always found it weird that the 'M' SoCs were a gen behind the 'A' line.

Back on the A17 Pro GPU, kind of: we now have benchmarks of Dimensity 9300, so first performance points for the latest Mali G720... and it fares even (slightly) better than 8G3!
First results here. I find it very interesting how close all those solutions end up to be when fabbed with the same process node. 8Gen3 and D9300 end up being very very similar.
So indeed, Adreno AND Mali are now ahead of Apple in the GPU efficiency race. We'll have to wait a bit more to see how RDNA3 performs in this ultra mobile space. Since it's fabbed with a different (and worse) process, I guess it will lag behing the two new champs. But maybe it will still be better than A17 Pro, which would definitely mean that those 3 vendors now have a better GPU arch than Apple (still custom IMG as it seems).

And last tidbits I find interesting about this GPU arch: we also got the first M3 Max GPU bench. And those are really close to mobile 4080 (which has a very similar bandwidth)... I guess the M3 Max GPU has more transistors, that part is hard to say (4080 mobile is ~35B), and we also lack the power consumption figures (which will likely be in favour of M3 Max). That's very interesting because it means the (Apple/IMG GPU) arch is good, but maybe having to scale from ultra mobile to powerful laptop/desktop machines induce some trade-offs.