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

[Boil]

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.

3nm has many "stages" - N3B / N3E / N3S / N3P / N3X

Each of these improve upon the previous, and are scheduled for the next few years...

There are also reports of Apple & TSMC working up the 2nm process for future products...?

 

[Analog Kid]

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.

I'm thinking maybe my language hasn't been clear enough in distinguishing what I do think is standard practice from what may happen at some companies in some cases but which I'm fully convinced is not happening at Apple.

What is standard practice is incremental improvements coming from incremental development. A company has limited capacity to innovate because it has limited resources available to it (for Apple this is mostly due to the pool of qualified engineers), it is limited in its ability to communicate among a large team, it is dependent on technology improvements outside its control (N5P, N4P), and risk due to design changes compounds non-linearly. So an engineering team makes improvements in steps.

That is the model I see Apple pursuing, and they're taking larger steps than most.


What is not standard practice is making a technological advance and withholding it from the market so that it can be delivered in bits over the coming years. In this thread the (very suspect) claim is that Apple has already created a part capable of 60% improvement but will be delivering that 20% at a time for business reasons. I've personally never seen that done.

And it doesn't make sense from a business perspective either. It implies you've made a massive engineering investment but then aren't seeking a rapid return on that investment and it implies the engineering team has nothing left to do because there's 3 years of improvements already in the pipeline. It gives the competition more time to catch up and leaves the company less directly competitive at every release. And, while there's always pressure to have consistent quarter to quarter results, the time value of money means that revenue today is worth way more than revenue in the future so best to get it while you can.

The best business plan is take as much as you can get as quickly as you can get it and have faith that there's more to come.

The one situation where the latter approach might make sense is when technological improvement is received rather that developed. But even in that case, when Apple was only getting sporadic processor improvements from Intel, they chose to space out their product releases and roll system level improvements into the same releases rather than keep making unimpressive intermediate releases.

So maybe Apple releases a major P-core update that delivers massive improvement in one chunk and then has to wait for the next process generation for the next big leap. It's not like there aren't a million other parts of that chip that their engineers can turn their attention to in the mean time with the benefit of incremental process improvements along the way-- and other parts of the product (display, cameras, etc) and other products that can provide the intermediate sales.

I'm devoting more words to this than it deserves, but I'm not happy with how I'm explaining it. My point is that there's no good business justification to withhold product improvements. There are technical reasons to develop incrementally. I think people on the outside see the incremental development, don't understand the technical challenges, and take a cynical view that something is being kept from them.

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.

Do those applications really need single core performance though? Most of those applications need multicore (otherwise they could just run on one fast core in a Mini).

I lose track day to day what the latest benchmarks say, but does seem like if you live close enough to Hoover Dam you can get an Intel processor to win on a single core benchmark-- so there is a motivation for Apple to finally drive a stake through the heart of x86, but I can never convince myself that they even really look at the competition that way. Apple most often appears to set a course that they think is right and then largely compete with themselves and their internal expectations.

And I am sure that Apple has more tricks up their sleeve (with the obvious one pursuing higher frequencies).

No doubt they've got more to come. It's been really exciting to have true innovation in the processor space again.

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.

It'll be interesting to see what they do. If Apple succeeds in making a breakout device here, the world is going to have it in pieces quickly enough to find where they hid the magic.

 

[senttoschool]

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.

So you're adding more requirements to your original dare then? 😅

Anyways, I'm with you, my friend. I think the chance of a 60% increase in ST is close to zero. But your dare just wasn't well thought out.

 

[thenewperson]

So you're adding more requirements to your original dare then?

Yeah, the goalposts were moved quite a bit there lol

 

[leman]

I'm thinking maybe my language hasn't been clear enough in distinguishing what I do think is standard practice from what may happen at some companies in some cases but which I'm fully convinced is not happening at Apple.

What is standard practice is incremental improvements coming from incremental development. A company has limited capacity to innovate because it has limited resources available to it (for Apple this is mostly due to the pool of qualified engineers), it is limited in its ability to communicate among a large team, it is dependent on technology improvements outside its control (N5P, N4P), and risk due to design changes compounds non-linearly. So an engineering team makes improvements in steps.

That is the model I see Apple pursuing, and they're taking larger steps than most.


What is not standard practice is making a technological advance and withholding it from the market so that it can be delivered in bits over the coming years. In this thread the (very suspect) claim is that Apple has already created a part capable of 60% improvement but will be delivering that 20% at a time for business reasons. I've personally never seen that done.

And it doesn't make sense from a business perspective either. It implies you've made a massive engineering investment but then aren't seeking a rapid return on that investment and it implies the engineering team has nothing left to do because there's 3 years of improvements already in the pipeline. It gives the competition more time to catch up and leaves the company less directly competitive at every release. And, while there's always pressure to have consistent quarter to quarter results, the time value of money means that revenue today is worth way more than revenue in the future so best to get it while you can.

The best business plan is take as much as you can get as quickly as you can get it and have faith that there's more to come.

The one situation where the latter approach might make sense is when technological improvement is received rather that developed. But even in that case, when Apple was only getting sporadic processor improvements from Intel, they chose to space out their product releases and roll system level improvements into the same releases rather than keep making unimpressive intermediate releases.

So maybe Apple releases a major P-core update that delivers massive improvement in one chunk and then has to wait for the next process generation for the next big leap. It's not like there aren't a million other parts of that chip that their engineers can turn their attention to in the mean time with the benefit of incremental process improvements along the way-- and other parts of the product (display, cameras, etc) and other products that can provide the intermediate sales.

I'm devoting more words to this than it deserves, but I'm not happy with how I'm explaining it. My point is that there's no good business justification to withhold product improvements. There are technical reasons to develop incrementally. I think people on the outside see the incremental development, don't understand the technical challenges, and take a cynical view that something is being kept from them.

I agree with practically everything you say here. But I also thing that you are way too committed to a narrow (and less interesting) reading of "3nm will enable huge performance improvements". Of course it won't be the case that they deliver 50% improvements and then don't do anything for three years. I interpret this statement as "Apple hardware built on 3nm will reach X% improvements over time".

Do those applications really need single core performance though? Most of those applications need multicore (otherwise they could just run on one fast core in a Mini).

I lose track day to day what the latest benchmarks say, but does seem like if you live close enough to Hoover Dam you can get an Intel processor to win on a single core benchmark-- so there is a motivation for Apple to finally drive a stake through the heart of x86, but I can never convince myself that they even really look at the competition that way. Apple most often appears to set a course that they think is right and then largely compete with themselves and their internal expectations.

I'd say that single-core performance has more utility than multi-core performance, simply because it's more flexible. One core can run both serial and parallel workloads, while multiple cores cannot run serial workloads. That the narrative has shifted from single-core performance to multi-core performance has to do with the fact that most CPU makers has run into a design ceiling, where improving performance requires disproportionate amounts of energy. Sure, there are plenty of applications where independent multicore processing is exactly what you want (many server-like workloads, shared computing clusters, raytracing etc.), but for the common workstation use multiprocessing is often inherently limited by synchronisation overhead. There is a reason why Apple has asked Intel to develop a custom line of Xeons with higher single-core performance (which then became Xeon W).

And I think Apple has a unique opportunity here. They can deliver pretty much top-tier performance only consuming five-six watts of power, so it's possible that they still have ways to push that performance ceiling up without breaking the bank. For example, there were some rumours that the Mac Pro will feature M2 cores clocked at 4.2 ghz. If this is true, then M2 Ultra would offer competitive multi-core performance compared to the new Sapphire Rapids Xeon WS series (all but the top w9-3495X) while delivering considerably higher single-core performance.

 

[PauloSera]

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.

Same kind of comments were made about M1.

 

[leman]

Please point to a 60% single threaded performance increase between 2 consecutive generations in the last 20 years. I dare you.

Apple A8 to Apple A9

 

[Analog Kid]

I agree with practically everything you say here. But I also thing that you are way too committed to a narrow (and less interesting) reading of "3nm will enable huge performance improvements". Of course it won't be the case that they deliver 50% improvements and then don't do anything for three years. I interpret this statement as "Apple hardware built on 3nm will reach X% improvements over time".

To be fair, I'm only committed to that reading here because that was the supposed "new information" injected into the conversation by this thread. And because I think people with that misperception are susceptible to the "Apple's holding back for business reasons" argument.

I tried to show how incremental improvements are made in the graph I posted earlier showing multiple releases at the same nominal geometry and their improvement over the previous geometry.

I also think it's too narrow, by the way, to focus on single core performance as the metric to track here. If you are interested in gaming then the GPUs are probably much more interesting, if you're interested in these new machine learning applications then the Neural Engines are more interesting, if you're into content creation then you're probably watching the media processors.

The real advantage to Apple Silicon is all these coprocessors and how they can be custom shaped to best fit the applications enabled by the the OS and frameworks. Apple's architecture enables much more efficient use of silicon resources for real user benefit than we've ever seen before.

I'd say that single-core performance has more utility than multi-core performance, simply because it's more flexible. One core can run both serial and parallel workloads, while multiple cores cannot run serial workloads. That the narrative has shifted from single-core performance to multi-core performance has to do with the fact that most CPU makers has run into a design ceiling, where improving performance requires disproportionate amounts of energy. Sure, there are plenty of applications where independent multicore processing is exactly what you want (many server-like workloads, shared computing clusters, raytracing etc.), but for the common workstation use multiprocessing is often inherently limited by synchronisation overhead. There is a reason why Apple has asked Intel to develop a custom line of Xeons with higher single-core performance (which then became Xeon W).

And I think Apple has a unique opportunity here. They can deliver pretty much top-tier performance only consuming five-six watts of power, so it's possible that they still have ways to push that performance ceiling up without breaking the bank. For example, there were some rumours that the Mac Pro will feature M2 cores clocked at 4.2 ghz. If this is true, then M2 Ultra would offer competitive multi-core performance compared to the new Sapphire Rapids Xeon WS series (all but the top w9-3495X) while delivering considerably higher single-core performance.

If all you were interested in was fast single core performance then you'd never really need anything more than an iPad. If you ask what distinguishes a desktop/workstation from other computers, it's the multi-core performance. The only reason you need access to wall power and efficient active cooling is to handle all of the CPU and GPU cores churning away.

At this very moment, the M2 Air and even an iPhone14 Pro will out perform the M1 Studio Ultra in single core performance, but the Studio is still the right machine for the tasks it was made for.

Do users of desktops and workstations also run Word and Chrome? Sure. They'll benefit like any other consumer user by single thread performance. I hesitate to say single thread has more utility though, most useful computing these days isn't happening with someone at the keyboard. Even when it comes to something like a Mac Pro-- for people looking at that class of product, if required to choose between better multicore performance and better single core, they'd almost certainly choose multicore. That's what powers their renders and exports and model training.

If they could only watch YouTube as well as they could last year, I don't think that would drive their decision. If "Chrome tabs" is your benchmark of choice then you're not looking for a workstation.

 

[leman]

To be fair, I'm only committed to that reading here because that was the supposed "new information" injected into the conversation by this thread. And because I think people with that misperception are susceptible to the "Apple's holding back for business reasons" argument.

No, no, that's absolutely fair and reasonable. Limiting the spread of FUD is very important.

If all you were interested in was fast single core performance then you'd never really need anything more than an iPad. If you ask what distinguishes a desktop/workstation from other computers, it's the multi-core performance. The only reason you need access to wall power and efficient active cooling is to handle all of the CPU and GPU cores churning away.

At this very moment, the M2 Air and even an iPhone14 Pro will out perform the M1 Studio Ultra in single core performance, but the Studio is still the right machine for the tasks it was made for.

Do users of desktops and workstations also run Word and Chrome? Sure. They'll benefit like any other consumer user by single thread performance. I hesitate to say single thread has more utility though, most useful computing these days isn't happening with someone at the keyboard. Even when it comes to something like a Mac Pro-- for people looking at that class of product, if required to choose between better multicore performance and better single core, they'd almost certainly choose multicore. That's what powers their renders and exports and model training.

Workstation workloads tend to be hybrid in nature, some of which are more single-core, some of which rely on cooperative work sharing between processors, and some of which is trivially parallel (many work packages with no dependencies). A good workstation needs to be able to run any of these with good performance.

A hypothetical one-core processor running at 64Ghz will always outperform 32 processors running at 2Ghz each, because it can scale to any type of workload, while the 32-core system will only ever reach it's maximal potential doing something that has 32*N packages with very little synchronisation overhead. Of course, it's impossible to build a 64Ghz processor (at least with current technology), but this logic scales. Fewer fast processors will always be more flexible than more slow processors. And as I mentioned, Apple is very aware of it. They have always strived to maximise per-core performance on their workstation machines.

Of course, if you only care about CPU rendering, this reasoning doesn't apply (because you don't need the flexibility). But a workstation should cater to a wide group of professional users and for many of those (software developers, scientists, engineers etc.) processing flexibility matters a lot.

 

[Analog Kid]

A hypothetical one-core processor running at 64Ghz will always outperform 32 processors running at 2Ghz each, because it can scale to any type of workload

Agreed. Unfortunately, though, that's not quite the tradeoff we have available to us... We have a choice between two options both presenting sub-linear growth in performance in the general case (Safari snappiness), but one presenting near linear growth in broad classes of important use cases for workstation users (rendering, big data processing, machine learning, image and video encoding/compression, audio production, code compilation).

Fewer fast processors will always be more flexible than more slow processors. And as I mentioned, Apple is very aware of it. They have always strived to maximise per-core performance on their workstation machines.

I don't think they have. If they wanted to maximize their per-core performance on workstations, why can't anyone seem to get the Mac Studio Ultra fan to spin any faster than a dreamcatcher on a lazy Sunday?

Apple is clearly maximizing something, but I don't think it's per-core performance. I think what they're maximizing is essentially the ratio of user experience to development cost. They keep the denominator small by sharing development effort from the Watch to the Studio (and likely through to the Pro).

User experience is a bit more multi-dimensional. Quiet fans dates back to Job's earliest obsessions with silence, and it shares a design goal with long battery life: keeping power consumption down. Another dimension is performance. So if you have a budget of watts and square millimeters of silicon, where are you going to spend it?

You could chase benchmarks and try to flex in the face of Intel.

That doesn't seem to be the approach Apple is taking though, they seem to be deeply focused on the device experience and are leveraging their control of the software layer to optimize the system in a way that Intel or Microsoft can't dream of. They're pursuing a relatively balanced development of Pcores, Ecores, GPU, and Neural Engine development, adding special processors aimed at their customer base such as the Media Engines, and whatever that weird little matrix accelerator is called.

Apple looks at the problems it wants to solve and takes a very tightly integrated approach to solving them because they own the stack from top to bottom.

Yes, everyone benefits from single core performance, but not if it comes at the expense of something better.