Leak of 3nm A17 chip - single core 60% faster?!

[Analog Kid]

My realistic take on this, is that going from to 5nm to 3nm would enable the 60% improvement

How?

Apple will ship A17 with ~20% performance boost and efficiency improvement

Why?


I think you might be under the false impression that averaging the optimistic with the cynical somehow yields "real". The problem is that if neither of your inputs are tethered to rational assumptions then it's just garbage in, garbage out... If I pulled a claim of 120% improvement at 3nm out of thin air, would your expectation change to Apple delivering 30% year over year?

 

[Analog Kid]

Wrong. Our default position on anything we hear from hearsay should be to believe that it is unverified, not to always default to think that it is either false or true.

The impact any new evidence has on our prior beliefs should be weighted by our confidence in that new evidence. Given the source of the evidence here, the marginal likelihood that it carries any truth is vanishingly small. While Cromwell may beseech us to not consider something overwhelmingly unlikely to be true as false, as a practical matter I'm comfortable with the simplification when it comes to MaxTech because it saves me so much computation in the long run.

 

[PineappleCake]

If it hasn’t been made clear yet, I’ll type it in all caps for you…
IT’S NOT ABOUT BOOSTING PERFORMANCE TO THE MAX, IT’S ABOUT HAVING THE BEST PERFORMANCE PER W.
Even if hypothetically Apple could make the 15pro 60% faster, they wouldn’t.
They’d make it 20% faster… while consuming less power.

Yeah they it make it 70% faster and also more efficient like Nvidia did with their 4090.

 

[flashflood101]

Obviously fake.

 

[leman]

I think you might be under the false impression that averaging the optimistic with the cynical somehow yields "real". The problem is that if neither of your inputs are tethered to rational assumptions then it's just garbage in, garbage out... If I pulled a claim of 120% improvement at 3nm out of thin air, would your expectation change to Apple delivering 30% year over year?

Because it would make business sense and would fit the historical pattern. Post A14, there weren't any significant changes to basic P-core architecture. Apple is already overdue with a more substantial P-core overhaul. It makes sense for such an overhaul to land with the 3nm process, as this is how Apple has operated in the past. Also, they need a new hardware platform to deliver steady improvements for the next 2-3 years. So it makes sense to think about this new platform as technology that enables a certain performance range. Business sense dictates that the initial product should ship with a configuration that beats old hardware, but still leaves some potential for better performing products in the future. If the new tech can allow 50% better performance than the current tech, you don't want to ship that out immediately because it will cost you stagnating sales in the future. It's better to spread this advantage over a couple of product releases and work on something better in the meantime. That's how everyone operates, Apple included.

What all of this means is that the expected performance of A17 is dictated by business strategy and not technological progress. When you look at past history of iPhones (before A15), you'll see a steady almost exact 20% gen-to-gen improvement in single-core performance. Must be a crazy chance that the CPU engineers have been hitting those numbers like clockwork, right? But the mystery entirely disappears if you consider that these 20% is more or less an artificial number. The chips themselves were likely able to hit higher performance, there was just no need for that.

To wrap this up, I agree with @gpat that A17 will likely show 20% performance increase over A16 because it's a nice number and Apple strategic marketing division likes it. The performance potential of the chip itself is likely to be much higher. This of course assuming that the development proceeds without hiccups (which is not at all guaranteed). At any rate, Apple needs new CPU tech and if they don't deliver one with 3nm products, it's very bad news.

 

[Gudi]

Wrong. Our default position on anything we hear from hearsay should be to believe that it is unverified, not to always default to think that it is either false or true.

And there you have it. This forum has come down to discussing what a rumor even is? 🤷

 

[dmccloud]

No, not "Fake until proven true." This is a rumors site, and this remains a rumor until proven true. Calling it fake is some fool declaring that a rumor is untrue without any data one way or the other.

People can still call the rumor fake given freedom of expression and related things. Also, the credibility (or lack therof) of the source of said rumor can and should be taken into account when attempting to determine the accuracy of what was claimed. Given Max Tech's penchant for phishing for views and hits on their YT page along with the recent controversy surrounding their MBP 14" hit piece video (where they created an test with the sole purpose of proving their preconceived notions rather than testing any sort of real world usage, even among creative professionals), anything they say can and should be taken with a grain of salt even under the best circumstances.

 

[leman]

No, not "Fake until proven true." This is a rumors site, and this remains a rumor until proven true. Calling it fake is some fool declaring that a rumor is untrue without any data one way or the other.

It takes more than an unsubstantiated random claim to make a rumour. There has to be at least some credibility or circumstantial evidence for a claim to be interesting or relevant. We have people here with understanding of the industry, who have been following this kind of technology for years. And to them, this looks like random gibberish with no substantial basis. In fact, we are only discussing this because a popular YouTuber — who has been known known to make sensationalist, uneducated claims — has featured this particular piece of information on their channel.

Now, the second set of scores (3000/9000) look plausible, because they fit both the historical pattern and what's technically realistic. I have no idea whether these scores are authentic or whether they are just an educated guess, but I wouldn't be surprised if A17 ends up in that ballpark.

 

[joptimus]

By Intel fanatics, yes.

What a stupid statement. This is about technology and physics, not fanboyism - this should have no place here. Please point to a 60% single threaded performance increase between 2 consecutive generations in the last 20 years. I dare you.

 

[TechnoMonk]

No, not "Fake until proven true." This is a rumors site, and this remains a rumor until proven true. Calling it fake is some fool declaring that a rumor is untrue without any data one way or the other.

I would agree for some sources, but anything Max Tech puts out is fake/fraud or click bait until proven. They did lot of shady crap over last two years.

 

[Analog Kid]

I agree with @gpat

Do you agree with @gpat or do you just come to a similar number by different reasoning.

Is a 20% number be likely? Maybe, I'll get to that last, but I want to start with the post I responded to and why I can't agree with it-- maybe you can tell me why I should.

My realistic take on this, is that going from to 5nm to 3nm would enable the 60% improvement described by the leak, but Apple will ship A17 with ~20% performance boost and efficiency improvement, also because 3nm is here to stay for several years and they have to gauge the improvements for A18-A19 as well.
So an A17 benchmarking with 60% boost over 5nm would be fake news, but holding some truth nonetheless.

Here's what I see assumed in that comment:

• a process change enables a 60% performance increase
• the "leak" has merit
• Apple has built such a device, hence the "leak" and the reason we're discussing it.
• Having built it, Apple is hobbling it's performance to dribble it out over 3 generations
• 60/3 is 20, so let's assume that and the engineers can take a holiday
• any view holds some nugget of truth
• taking an absurd claim and backing it off somehow leads to realism.

I don't agree with any of those points.

Because it would make business sense and would fit the historical pattern. Post A14, there weren't any significant changes to basic P-core architecture. Apple is already overdue with a more substantial P-core overhaul. It makes sense for such an overhaul to land with the 3nm process, as this is how Apple has operated in the past. Also, they need a new hardware platform to deliver steady improvements for the next 2-3 years. So it makes sense to think about this new platform as technology that enables a certain performance range. Business sense dictates that the initial product should ship with a configuration that beats old hardware, but still leaves some potential for better performing products in the future. If the new tech can allow 50% better performance than the current tech, you don't want to ship that out immediately because it will cost you stagnating sales in the future. It's better to spread this advantage over a couple of product releases and work on something better in the meantime. That's how everyone operates, Apple included.

What all of this means is that the expected performance of A17 is dictated by business strategy and not technological progress. When you look at past history of iPhones (before A15), you'll see a steady almost exact 20% gen-to-gen improvement in single-core performance. Must be a crazy chance that the CPU engineers have been hitting those numbers like clockwork, right? But the mystery entirely disappears if you consider that these 20% is more or less an artificial number. The chips themselves were likely able to hit higher performance, there was just no need for that.

To wrap this up, I agree with @gpat that A17 will likely show 20% performance increase over A16 because it's a nice number and Apple strategic marketing division likes it. The performance potential of the chip itself is likely to be much higher. This of course assuming that the development proceeds without hiccups (which is not at all guaranteed). At any rate, Apple needs new CPU tech and if they don't deliver one with 3nm products, it's very bad news.

What you're saying is more sensible, though I still find 20% an arbitrary number.


Three generations of 5nm only gained about 45% in single core performance over 7nm (GB6). Could 3nm give us a 60% improvement? I suppose, but how? I don't think you can look back in time and cherry pick a rate of improvement. Everything was significantly less mature back then so you'd expect a higher rate of improvement. Apple can certainly pull another rabbit out and bend the curve again, but there's no reason to think the earlier rate was more indicative of what we should see next than the current rate is.

Or at least I don't see anything in this thread that explains it, I know you're in a ton of these conversations so maybe you know something I don't.


As you're suggesting, the performance improvements come from a combination of the underlying process and implementation technologies and from the architecture and design not from any one place. The generational improvements are incremental because the improvements are pursued incrementally. If Apple decided to jump straight to the end and make the A19 now, it would be a disaster. It would require significantly more engineering time, it would pile up a ton of risk, and it wouldn't benefit from the experience of the A17 and A18. So what a wise engineering team does is focus on a subset of potential areas for improvement and focus their attention on gains in a few areas at a time.

So sure, there might be a business reason to spread improvements out but there's also overwhelming technical reasons. If anything the business will want to push ahead faster because the sooner improvements are made the sooner they can be built upon. If 20% year over year was the trend for a while (though notably the A13 seemed to give a 30+% increase at it was using the same process as the A12) it's likely that that was the target improvement that the designers were given. It was an input, not an exercise in making a super chip and then rationing the annual updates.


And there's no reason I see to set 20% as an upper bound. I suspect this fancy VR headset is going to want a ton of performance at low power-- a lot will be in the GPU cores, but Apple may find that to be a reason to boost the overall system performance this generation to support the headset. I'm sure they see their AS line to be a competitive advantage in this space.

Exactly the opposite argument could be made based on the VR headset-- single core performance may not be a priority and all attention might be on the GPUs and coprocessors and no need to boost the Performance core at all.

All of which brings me back to my point-- an estimate is only realistic if it can be supported by realistic assumtions and there's certainly nothing in this supposed "leak" that should change our expectations in any way.

 

[Gudi]

If anything the business will want to push ahead faster because the sooner improvements are made the sooner they can be built upon.

Correct.

If 20% year over year was the trend for a while (…) it's likely that that was the target improvement that the designers were given.

Incorrect. Why do you backtrack from your initial thoughts? In interviews the Apple chip designers said, they "do not hold back". The M2 is as fast as they could build it right now. Likewise the M3 and A17 will do what’s possible on 3 nm. The only design targets are improved power efficiency and battery life.

 

[senttoschool]

My realistic take on this, is that going from to 5nm to 3nm would enable the 60% improvement described by the leak, but Apple will ship A17 with ~20% performance boost and efficiency improvement, also because 3nm is here to stay for several years and they have to gauge the improvements for A18-A19 as well.
So an A17 benchmarking with 60% boost over 5nm would be fake news, but holding some truth nonetheless.

There is no truth to it.

If they can get 60% in one generation, they would do it. They can't.

20% is about right, might be even a bit optimistic. A13 to A14 was a node jump from N7 to N5 and it got 20% boost in ST. N7 to N5 is expected to be a bigger jump than N5 to N3 because SRAM was still scaling fine. N5 to N3 will not provide any SRAM scaling. Thus, if Apple can get 20% boost from N5 to N3 without more SRAM, that's a huge win. I suspect that the A17 will be a physically bigger chip to add more SRAM.

 

[Analog Kid]

Correct.

Incorrect. Why do you backtrack from your initial thoughts? In interviews the Apple chip designers said, they "do not hold back". The M2 is as fast as they could build it right now. Likewise the M3 and A17 will do what’s possible on 3 nm. The only design targets are improved power efficiency and battery life.

Fair point. When you call it out like that it does look like I reversed myself within a few sentences...

I know there was an executive that made the point they don't hold back and dole out updates, but I kind of put that in the category of "of course that's what they say". I put a bit more weight on the pattern of updates looking like they've been going flat out generation to generation and taking what they can get.

I'm not sure how these things are done in Apple, exactly, and I'm not sure their marketing department plays the same role it does in other companies, but often there's a back and forth between marketing and engineering looking for an achievable design target that is sufficient to release. Marketing typically comes in wanting more than can be reasonably achieved and engineering is often concerned with predictability and controlling risk. What I meant by 20% is an input is that I'd expect the outcome of that the engineering team was able to find an achievable combination of improvements that they met the marketing threshold for release.

In any event, it's not a matter of holding back potential improvements to make sure you have something to follow with, it's about working within your available resources to maximize rate of improvement while minimizing technical and schedule risk.

 

[Zest28]

We won’t know the performance difference because Apple does a trade off of battery life vs performance.

Maybe Apple will focus on a lot more battery life and maybe increasing the speed by only 20%, rather than maximum performance gains.

 

[senttoschool]

What a stupid statement. This is about technology and physics, not fanboyism - this should have no place here. Please point to a 60% single threaded performance increase between 2 consecutive generations in the last 20 years. I dare you.

I can think of 3 instances in the last 20 years:

1. Apple A8 to A9 was a 70% jump in ST Geekbench5.

iOS Benchmarks - Geekbench Browser

2. Intel Netburst to Core 2 Duo was 60%+: https://www.anandtech.com/show/2045

3. AMD's Zen1 was 60%+ faster than Bulldozer: https://www.anandtech.com/show/1117...review-a-deep-dive-on-1800x-1700x-and-1700/17

Since you dared, what do I win? haha

 

[Gudi]

I know there was an executive that made the point they don't hold back and dole out updates, but I kind of put that in the category of "of course that's what they say".

I forgot to mention one important design target. They design most of their products for a specific price point and profit margin.

The M1 still used LPDDR4 RAM whereas the M1 Pro/Max/Ultra already used LPDDR5. Now the M2 also uses LPDDR5.

Apple will withhold certain features from their base models, if it makes them significantly cheaper to produce. But that’s mostly true for third party components. Everything they build themselves has already been accounted for as R&D costs.

So yes, marketing does inform which features go into which product. But not to keep them on a certain trajectory or force upsell. They do however strongly encourage upselling! 😂

 

[Analog Kid]

I can think of 3 instances in the last 20 years:

1. Apple A8 to A9 was a 70% jump in ST Geekbench5.

iOS Benchmarks - Geekbench Browser

2. Intel Netburst to Core 2 Duo was 60%+: https://www.anandtech.com/show/2045

3. AMD's Zen1 was 60%+ faster than Bulldozer: https://www.anandtech.com/show/1117...review-a-deep-dive-on-1800x-1700x-and-1700/17

Since you dared, what do I win? haha

I didn't think of going to Wayback-- that gives a few more generations of chips in the GB5 results. I wish there was a way to download tabular data from GB, but after some hand transcription:

Something introduced in the A13 generation was seriously upweighted by the GB6 benchmark.

The process geometry graph clusters the data by nominal node geometry (e.g. N5, N5P, N4P all nominally 5nm) against the previous geometry to show the initial improvement in the geometry step and the potential improvement when mature.

 

[Analog Kid]

They do however strongly encourage upselling!

I still argue that you don't encourage upselling by only stocking the bait in stores and making the switch an online BTO with a slower delivery date...

 

[joptimus]

I can think of 3 instances in the last 20 years:

1. Apple A8 to A9 was a 70% jump in ST Geekbench5.

iOS Benchmarks - Geekbench Browser

2. Intel Netburst to Core 2 Duo was 60%+: https://www.anandtech.com/show/2045

3. AMD's Zen1 was 60%+ faster than Bulldozer: https://www.anandtech.com/show/1117...review-a-deep-dive-on-1800x-1700x-and-1700/17

Since you dared, what do I win? haha

Yeah, but the starting point were designs with low IPC or at the end of their capabilities. The M2/A16 are much more mature starting points with limited improvement in manufacturing processes accompanying them. So not quite comparable to today.

 

[leman]

All of which brings me back to my point-- an estimate is only realistic if it can be supported by realistic assumtions and there's certainly nothing in this supposed "leak" that should change our expectations in any way.

I am 100% with you on this. And you are right that there are too many unknowns, which makes precise speculation difficult. What we can do is try to guess what kind of technological advancements new manufacturing process would enable (increasing complexity of design at same power consumption) and what business moves Apple might make.

The entire topic is too large and one can lose hours talking about details, so I'll just comment on a few things. And these are indeed comments rather than arguments because I don't find myself fundamentally disagreeing with things you say.

• Having built it, Apple is hobbling it's performance to dribble it out over 3 generations

Just a quick comment on this. This is standard business practice and something all companies do. Otherwise they would ship a very strong product and generate massive initial demand (likely without the means to satisfy this demand) and then suffer lacklustre sales for a while. That's not good business.

Three generations of 5nm only gained about 45% in single core performance over 7nm (GB6). Could 3nm give us a 60% improvement? I suppose, but how? I don't think you can look back in time and cherry pick a rate of improvement. Everything was significantly less mature back then so you'd expect a higher rate of improvement. Apple can certainly pull another rabbit out and bend the curve again, but there's no reason to think the earlier rate was more indicative of what we should see next than the current rate is.

I think this is a more complex question. Of course, it's not plausible to assume that merely by moving to 3nm such improvements become possible. But 3nm combined with a new design (wider and/or capable of higher frequencies) could do the trick. The thing is, if Apple wants to compete in the high end desktop/workstation segment, they need hardware capable of substantially better performance peaks than today. So yeah, I wouldn't be surprised if their upcoming designs end up 50-60% faster on some products, for those reasons.

Regarding the lower improvement rates with 5nm... if I remember the discussions correctly, 5nm is less of an improvement over previous generations than some other nodes, limiting what can be done somewhat. Also, Apple might well be experiencing diminishing returns with their designs — they have been steadily increasing the cache size and out-of-order execution capabilities of their processors, and they might have reached a peak. It's also interesting to note that the latest generation of x86 takes some hints from Apple (e.g. by implementing large caches). In other words, Apple might be running out usual tricks while others are quickly picking up some of their tricks.

But I think there is hope — from what I understand 3nm will come with some innovation (like flexible combination of performance-oriented and energy-efficient cells) that enables more design flexibility. And I am sure that Apple has more tricks up their sleeve (with the obvious one pursuing higher frequencies).

BTW, this is the main issue with this kind of speculation. We are dealing with way too may variables and factors.

So what a wise engineering team does is focus on a subset of potential areas for improvement and focus their attention on gains in a few areas at a time.

What Apple has been doing so far is interleaving the improvements in different areas. Previously they would ship a new cache system every two years and CPU backend improvements the other two years. Even even post A14, with slower performance gains they did tremendous work on E-cores, GPU, and most likely internal fabric (I have a suspicion that M2 Pro/Max is shipping with a new on-chip network based on patents and GPU scaling). So yeah, you area absolutely right.

And there's no reason I see to set 20% as an upper bound. I suspect this fancy VR headset is going to want a ton of performance at low power-- a lot will be in the GPU cores, but Apple may find that to be a reason to boost the overall system performance this generation to support the headset. I'm sure they see their AS line to be a competitive advantage in this space.

A quick comment on this — I think there is all reason to believe that the VR headset will be simply based on M2. VR is all about energy-(and bandwith-) efficient high-res rendering and Apple has been slowly building this kind of technology. A15 brings on-the-fly compression to render targets and Apple has had variable rate rasterization since A13. With these, they can probably achieve 6K-like quality while using 2K or lower actual render target.

 

[Jorbanead]

https://twitter.com/i/web/status/1635506320479784961

Going to point out that this 'leak' from a more trusted source is much more realistic of a year-over-year improvement.

This is way more realistic. This is still a big leap forward compared to the last several years, but is more in line with past upgrades.

Based on this info, I did some very quick (and probably wrong) math to see how this would scale with the M-series based on the gains from each year.

 

[Gudi]

This is standard business practice and something all companies do. Otherwise they would ship a very strong product and generate massive initial demand (likely without the means to satisfy this demand) and then suffer lacklustre sales for a while. That's not good business.

Unless, you’re Apple and proof those self-declared business experts wrong. Nobody really cares about quarterly earnings. Everybody knows the Apple Silicon transition was a tremendous success. And when next year’s sales are compared to this year’s sales, everything will look golden again.

Apple (AAPL) Q1 2023 Mac sales fell 28.7% from prior year​

 

[Jorbanead]

Unless, you’re Apple and proof those self-declared business experts wrong. Nobody really cares about quarterly earnings. Everybody knows the Apple Silicon transition was a tremendous success. And when next year’s sales are compared to this year’s sales, everything will look golden again.

Apple (AAPL) Q1 2023 Mac sales fell 28.7% from prior year​

I believe Leman was talking about a longer period of lackluster sales. Not just a quarter.

The argument being - pretend they could jump to 60% performance improvement this year in M3, but maybe that would mean the next two years (M4 and M5) they would have very little room to grow (this is all hypothetical, but just using this as an example). The headlines the next few years would be “Apple is done” and “Apple hit a wall” or whatever clickbait media comes up with, because Apple put all their eggs in one year. Mac sales would decline (over a few years) because people would think Apple made a mistake transitioning and maybe “intel was the right choice long term”.

Leman (I think) was mostly just saying, even if Apple has the power to push performance that high right now with M3, it’s not that odd for them to hold back a little bit (20% gains instead of 60%) to give them room to grow over the next few years. Having continued year or year growth that is steady is better than having one amazing year and two horrible years.

 

[Gudi]

The argument being - pretend they could jump to 60% performance improvement this year in M3, but maybe that would mean the next two years (M4 and M5) they would have very little room to grow.

That is of course exactly what will happen. I haven’t heard of any new fabs smaller than 3nm being build, so M4 just like M2 will be lackluster again.

The headlines the next few years would be “Apple is done” and “Apple hit a wall” or whatever clickbait media comes up with, because Apple put all their eggs in one year.

The thrill of Apple silicon is already gone​

Mac sales would decline (over a few years) because people would think Apple made a mistake transitioning and maybe “intel was the right choice long term”.

Yeah, nobody will ever buy a new Mac again, because the M1 was simply too good. Makes absolute sense, the appearance of yearly progress is all that matters, even above actual progress. *lol*