Discussion Apple Silicon SoC thread

[amrnuke]

This looks more like AMD advocacy, and the opposite of reality.

The 5950X isn't designed for more, it does significantly less, than an A14.

No GPU.
No Media encoder.
No AI engine.
No Flash controller.

The A14 SoC includes all that and does WAY more than the 5950x, which is just a CPU.

I think from both of our standpoints, it would be silly to continue to discuss these things without first discussing how AMD and Apple regulate power supply to unused components. For instance, AMD's IF is active and has the capacity to handle 32 threads of information, but it is still active even when only 1 core is used. Additionally, in GB5, it may not test the full capacity of the AI engine, but the AI engine may activate during certain parts of the test. So depending on what is tested and how power regulation has occurred, we may or may not get an accurate estimate on these tests of actual power consumption. I'd wager, though, that the 5950X has a lot more limits because it also needs the framework for scaling out to 16 cores, SMT, PCIe lanes, etc.

 

[shady28]

With all the hype surrounding M1 due to some geekbench scores and vague marketing and Apple fanboy hype, it's bound to become disappointing. Yes it's a great product, but probably less so than it now seems. We will see soon I guess.

From what I see right now, the M1 looks incrementally better than either Zen 3 or Tiger Lake in *raw performance*.

People are making a big deal about IPC which factors in frequency, but in the end neither the frequency nor IPC matters, only the raw performance. No one knows if M1 can even go beyond 3.2Ghz or so, it may be a by product of its architecture roots that it is at its max frequency at 3.1 - 3.2 and will never exceed that.

If that is the case, it's impressive for low power use cases but does little to move the bar in terms of absolute performance past Zen 3 / Tiger Lake. In fact, it's likely that 8 core Tiger Lake H (high power 35-45W laptop chip) will match M1 in single thread and pretty much 100% certain that it will demolish it in multi-thread given that it's already matched by the low power 4-core TGL. A Zen 3 laptop part (when it comes out) would likely do the same.

Honestly most of the advantages that M1 has right now can probably be attributed to the 5nm process node they are using.

 

[Eug]

Speaking of numbers of cores, how far up do you think they'll scale?

It's a lock they'll create 12 core chips for MacBook Pros and iMacs, but presumably that will consist of 8 performance cores and 4 efficiency cores. I'm also thinking they'll have a Mac Pro chip with 12 performance cores, but would it make sense to remove the efficiency cores? And what about beyond that? Would it make sense to create a dual-CPU Mac Pro with 2 x 12 performance cores, for a total of 24 performance cores and no efficiency cores?

 

[gdansk]

Who said Apple will scale clock speeds fast?

Also, who said that Apple will scale core counts fast?

Personally I think Apple won't raise clock speeds that much, and I also personally think that Apple will only scale core counts up to a certain level that fits their business model. Apple is not in the server or HPC business.

Presumably they will want a Mac Pro, right? They'll need at least more cores to compete there. In the end, AMD has a large addressable market in servers which means they can repurpose those designs for workstations. Where as Apple has apparently no purpose in designing workstation chips expect what they can sell in the Mac Pro and maybe iMac Pro.

If they don't clock higher than there is no great obsoletion of the PC space. Without higher clocks and core counts, it isn't really the revolution to which I was looking forward.

 

[IvanKaramazov]

Yes we are discussing core performance. Core performance of the 5950X, 5900X, 5800X, and 5600X are all higher than the Firestorm core.

You keep saying that core performance of all these Zen 3 parts are higher than the Firestorm core, but what exactly is your metric for saying so? I'm not arguing that they're not better, necessarily, just that we have very little data yet, and for better or worse what we have is Geekbench for M1, and additionally Spec2006 for A14. Geekbench 5.3 did indeed see a slight uplift for the Zen 3 parts, but even if you only accept the highest scores, it looks to me like the Zen 3 cores average around the upper 1700s, versus a total average of 1740 for the M1. It's around a 2% difference. A14 v Zen 3 in Spec2006 shows a slightly larger gap, but doesn't account for the ST uplift the M1 has over the 14, which if my math is correct likely puts them, again, within about 2% difference. Hardly much of a grand lead for Zen 3 there. Why exactly should Apple even want higher clocks, if they can achieve parity at far lower ones?

It is entirely possible that Apple is pushing the M1 to its top possible frequency, and that no further ST performance is to be had there by simply improving thermals (e.g. putting it in a desktop form factor). In fact I suspect that's true. But presumably if AMD could improve ST performance on Zen 3 without negatively hampering something else, they would do so as well. So we're left with a situation where the mobile M1 and desktop Zen 3 cores are essentially at parity, while the M1 core uses way less than 50% power at load. If anything, our current data suggests that Arm and x86 are roughly equal in per-core performance potential, despite Arm apparently having a massive power efficiency advantage. In the desktop, the power efficiency advantage only matters if Apple can build a 24 core version of this or something (and maybe they can! We just don't know yet). But in mobile, it's hard to see any argument for x86.

Also, worth noting that the Zen mobile chips have always exhibited slower ST performance than their analogous desktop equivalent, so I will be very (pleasantly) surprised if Cezanne comes anywhere close to M1's ST performance. I fully expect it to surpass it in MT, of course, but I still think it will be slower than the 8x4 Bloomberg core, whenever that materializes, which will likely be the more apples to apples comparison based on TDP.

 

[Eug]

Presumably they will want a Mac Pro, right? They'll need at least more cores to compete there. In the end, AMD has a large addressable market in servers which means they can repurpose those designs for workstations. Where as Apple has apparently no purpose in designing workstation chips expect what they can sell in the Mac Pro and maybe iMac Pro.

If they don't clock higher than there is no great obsoletion of the PC space. Without higher clocks and core counts, it isn't really the revolution to which I was looking forward.

That's why I asked if it makes any sense to create a dual-CPU Mac Pro. See my post right above yours.

ie. 12 perf core for low end Mac Pro with no eff cores. 2 x 12 perf core for high end Mac Pro.

 

[guidryp]

Speaking of numbers of cores, how far up do you think they'll scale?

It's a lock they'll create 12 core chips for MacBook Pros and iMacs, but presumably that will consist of 8 performance cores and 4 efficiency cores. I'm also thinking they'll have a Mac Pro chip with 12 performance cores, but would it make sense to remove the efficiency cores? And what about beyond that? Would it make sense to create a dual-CPU Mac Pro with 2 x 12 performance cores, for a total of 24 performance cores and no efficiency cores?

I agree. Midrange will likely be with 8 Performance and 4 Efficiency cores, since it still has to go in Laptops, 2021 will likely bring the new midrange chip that fills in higher end Macbooks and regular iMacs. Leaving just Mac Pro and iMac pro at the high end.

High end is VERY murky.

Multiple options:

Massive monolith with 16+ performance cores.

Some kind of Chiplet design. So the can Scale core count like AMD does.

Meshing multiple full function SoC in some novel way to make use of combined CPU/GPU/AI cores.

Main SoC used with optional GPU and CPU chiplets to improve peformance as needed.

There are so many ways the could go with the high end, it's really hard to figure where they go, and this will probably be the last part to arrive, probably in 2022.

 

[jeanlain]

but in the end neither the frequency nor IPC matters, only the raw performance

Perf/Watt is important in most use cases. This is where the M1 dominates.

 

[IvanKaramazov]

If that is the case, it's impressive for low power use cases but does little to move the bar in terms of absolute performance past Zen 3 / Tiger Lake. In fact, it's likely that 8 core Tiger Lake H (high power 35-45W laptop chip) will match M1 in single thread and pretty much 100% certain that it will demolish it in multi-thread given that it's already matched by the low power 4-core TGL. A Zen 3 laptop part (when it comes out) would likely do the same.

Honestly most of the advantages that M1 has right now can probably be attributed to the 5nm process node they are using.

I think you're probably right to guess the Tiger Lake H will match M1 in ST performance. I would be surprised if Zen 3 mobile does. The M1 doesn't really compete with the H-series chips though, or really even with Zen chips running at 25w. Rumors from manufacturing have for almost a year now been saying an 8x4 variant chip is coming, and frankly it would be surprising if it wasn't. Apple will transition their 16MBP and iMacs at some point, and they're not likely to just pop an M1 in.

My own reading of the tea leaves is that the M1X chip, or whatever it's called, will probably match TGL H in ST and significantly surpass it in MT, while it will probably match a theoretical 5900H in MT while surpassing it in ST. And I suspect it will achieve this "best of both worlds" performance while remaining more efficient than Intel or AMD's offerings. I'm less sure, but still suspect, that the difference in efficiency will be more than can be simply explained by the ~15% advantage brought by process node.

 

[name99]

According to Jobs autobiography, he was inclined to Intel but the project manager strongly opposed it and chose ARM for mobile platform. Perphas initially Jobs preferred Atom for iPad/iPhone and PWRfficienct PA6T for Mac but in the end both didn't take place.It is said that in 2008 Jobs arranged PA semi team to design SoC for iPhone.

Yeah, I'm not convinced it would have changed anything ultimately.
Intel would have Intel'd, shipping barely adequate mobile SoCs, refusing everything Apple wanted to do that was more ambitious.
Look at how the IMG thing played out. We don't (and maybe never will) know the exact details, but the large contours of the story seem to be that Apple wanted IMG to be ever more ambitious and sophisticated, and IMG was apparently unwilling to run that fast (maybe they weren't confident they could do the job? maybe they thought they couldn't sell the results to anyone except Apple?)

My guess is with Intel things would have played out essentially the same way; maybe a year or two of Intel chips before Apple said "to hell with this" and moved on to ARM anyway; if anything the end result might have been (less sophisticated) Apple cores as soon as the A4 or so.

 

[name99]

Got the iPhone 12 Pro Max on Friday. Man, this thing flies! I knew it was much faster on paper, esp. considering my 7 Plus is so old now, but I’m a little surprised it still feels so much faster even just for basic OS navigation and light mobile websites. And it’s nice to see my iPhone doesn’t appear to throttle significantly. FWIW, this is what I got the one and only time I’ve run Geekbench on this phone.

Makes a difference, doesn't it? Even if you don't care about benchmarks, you can feel the speed difference.
I agree that no normal person upgrades devices every year (and anyone who constructs their reviews or analysis around this behavior is a mendacious idiot), but the two year difference is noticeable, and by three years you really want to upgrade if you can afford it.

 

[amrnuke]

You keep saying that core performance of all these Zen 3 parts are higher than the Firestorm core, but what exactly is your metric for saying so? I'm not arguing that they're not better, necessarily, just that we have very little data yet, and for better or worse what we have is Geekbench for M1, and additionally Spec2006 for A14. Geekbench 5.3 did indeed see a slight uplift for the Zen 3 parts, but even if you only accept the highest scores, it looks to me like the Zen 3 cores average around the upper 1700s, versus a total average of 1740 for the M1. It's around a 2% difference. A14 v Zen 3 in Spec2006 shows a slightly larger gap, but doesn't account for the ST uplift the M1 has over the 14, which if my math is correct likely puts them, again, within about 2% difference. Hardly much of a grand lead for Zen 3 there. Why exactly should Apple even want higher clocks, if they can achieve parity at far lower ones?

It is entirely possible that Apple is pushing the M1 to its top possible frequency, and that no further ST performance is to be had there by simply improving thermals (e.g. putting it in a desktop form factor). In fact I suspect that's true. But presumably if AMD could improve ST performance on Zen 3 without negatively hampering something else, they would do so as well. So we're left with a situation where the mobile M1 and desktop Zen 3 cores are essentially at parity, while the M1 core uses way less than 50% power at load. If anything, our current data suggests that Arm and x86 are roughly equal in per-core performance potential, despite Arm apparently having a massive power efficiency advantage. In the desktop, the power efficiency advantage only matters if Apple can build a 24 core version of this or something (and maybe they can! We just don't know yet). But in mobile, it's hard to see any argument for x86.

Also, worth noting that the Zen mobile chips have always exhibited slower ST performance than their analogous desktop equivalent, so I will be very (pleasantly) surprised if Cezanne comes anywhere close to M1's ST performance. I fully expect it to surpass it in MT, of course, but I still think it will be slower than the 8x4 Bloomberg core, whenever that materializes, which will likely be the more apples to apples comparison based on TDP.

For Zen3 leading Firestorm, that comes from Anandtech's M1 article showing A14 performance in SPEC int and fp workloads (granted, with A14's int score artificially inflated by using a new version of LLVM compiler in Xcode 12). Even with that advantage, Zen3 still surpasses Firestorm by 8.2% int / 15.8% fp (5950X) or 4.5% int / 8.8% fp (5600X). This is despite being two steps behind in process (A12 was N7, A13 was N7P, A14 is N5, while Zen3 is N7). If you assume 15% performance gain from moving from N7 -> N5 (or 30% power reduction) it's clear that with respect to process-independent performance and uarch refinement, they're not that dissimilar. If AMD the same results as now, and used 30% less power, I think we'd all be very impressed. Well, that's possible with just moving to N5 and taking the pure efficiency benefit.

One non-sequitor that I think is interesting here is scalability and efficiency. We know that when fully loaded in Anandtech's per-core power loading charts, the 5950X's total package draw is 119W, or 7.44 total package watts per core. We do not have a clue what Firestorm scaling would be and what MT performance would look like in e.g. CB23. We also don't have a clue what frequency scaling would be at that level. I just think there's a lot to digest here, and even a multichiplet architecture isn't as simple as it's made out to be.

I am hopeful - Firestorm is a damn fine core. It's going to make Apple a lot of money, and it has the potential to save users a lot of money (as the die area of the core is so tiny, and by cutting out Intel as a middle-man, Apple should save money). But ultimately I think what it does is advance things one step further for all of us, pushing the bleeding edge into the consumer space. Which is awesome!

 

[shady28]

Perf/Watt is important in most use cases. This is where the M1 dominates.

That does matter for a laptop only insofar as it translates to fast + long battery life, but no one actually asks 'what is the performance per watt'. They want to know how fast is it, and how long does the battery last. After all, modern PCs and laptops spend most of their time at idle, and this isn't a rendering farm CPU.

M1's issue will be scaling up to desktop performance levels.

Here's an comparison of the submitted MacBook Pro M1 vs MSI Prestige14 (Tiger Lake) and recent 5800X Zen 3 scores on GeekBench. So all 3 of these represent best in class I believe, 8 Core (4+4) / 8 thread M1, 4 Core / 8 Thread Tiger Lake, 8 Core / 16 Thread Zen 3.

 

[Carfax83]

What Zen3 part has sixteen cores on a die?

On the same piece of silicone then. Technically you're right, as Zen 3 has 8 cores per CCX so the 5950x has two CCXs connected to each other by the I/O die.

 

[name99]

So you expect Apple to leave their working high margin strategy behind and allow competing subscription based services to fight for the same TAM?

How does it work out well for Apple if consumer buys cheap Apple hardware and then opts for subscriptions from Spotify, Netflix, Disney, HBO, Dropbox, Microsoft, Google, Amazon? You're talking about universal app frameworks, but that's just the technical solution. The real hot potato is the appstore tax (on both iOS and Android), and we're nowhere near the point of having the facts straight on what will happen with this in the next years. All we know so far is the giants won't stand for that 30% fee, and without that fee Apple would essentially facilitate the success of the competition. (they build the low margin hardware, the low margin marketplace, and combined competiton gets big chunk of subscriptions)

Subscription model does wonders when your customers are captive or when you get a big slice of the pie no matter who sells the subscription. If any of that changes you better make sure your cloud services are THE best.

You are refusing to listen to what is being said because you already think you know the answers.

Apple doesn't sell iPhones, and Apple doesn't sell subscriptions. Apple sells "personal computing". That's a selection (growing wider every year) of pieces that are different, but all work together, to provide whatever computing you need, from watch to speaker to phone to TV screen to car to laptop, including not just HW but remote storage or content if those are what you want to add to your bundle.

The importance of these new ways of writing apps is not some crazy paranoia about app store taxes and where Apple will (or will not) make money; it's that

(a) these new ways of writing apps are exactly what you want for apps that are used on everything from a watch to a car head unit screen to a TV screen -- they are fairly easily portable across UI models in a way that was not true for a 1995-style mac or windows app.

(b) that same portability means that they're also not locked into the Windows world -- which in turn opens up Apple to be able to pick up a lot more of the PC market based both on many more apps not being Windows exclusive, and having right out the door a suite of useful stuff in the form of iOS/iPad apps -- not yet UI optimized for new macs, but good enough.

 

[IvanKaramazov]

For Zen3 leading Firestorm, that comes from Anandtech's M1 article showing A14 performance in SPEC int and fp workloads (granted, with A14's int score artificially inflated by using a new version of LLVM compiler in Xcode 12). Even with that advantage, Zen3 still surpasses Firestorm by 8.2% int / 15.8% fp (5950X) or 4.5% int / 8.8% fp (5600X). This is despite being two steps behind in process (A12 was N7, A13 was N7P, A14 is N5, while Zen3 is N7). If you assume 15% performance gain from moving from N7 -> N5 (or 30% power reduction) it's clear that with respect to uarch, they're not that dissimilar. If AMD had the SPEC performance lead, had proven scalability to multiple cores, and used 30% less power, I think we'd all be quite impressed. Well, that's possible with just moving to N5 and taking the pure efficiency benefit.

One non-sequitor that I think is interesting here is scalability and efficiency. We know that when fully loaded in Anandtech's per-core power loading charts, the total package draw is 119W, average freq drops to 3.775 GHz, and total package draw divided by cores is 7.44 total package watts per core. We do not have a clue what Firestorm scaling would be. We also don't have a clue what frequency scaling would be at that level. I just think there's a lot to digest here, and even a multichiplet architecture isn't as simple as it's made out to be.

100% agree there. Hopefully we'll have answers soon.

For what it's worth, I too am skeptical that the M1 cores can be scaled up to 5950X type MT workloads. It remains to be seen, of course, but I wouldn't be surprised if Apple's desktops never quite catch up with the best from AMD or whoever. On the other hand, their absolute advantage in mobile (and by that I mean up to an including AIO forms like the iMac) should be concerning to the x86 world, because if the advantage is large enough that Microsoft and OEMs push groups like Qualcomm and Nuvia to supply Arm chips for laptops, it's very possible that these x86 chips only have an advantage in the extremely narrow high-power desktop market. I'm not sure you can maintain the levels of R&D budget and innovation if you start to bleed the mobile market.

 

[SarahKerrigan]

On the same piece of silicone then. Technically you're right, as Zen 3 has 8 cores per CCX so the 5950x has two CCXs connected to each other by the I/O die.

It's not on the same piece of silicon (not silicone.) It's an MCM, with multiple chips in one package.

 

[IvanKaramazov]

M1's issue will be scaling up to desktop performance levels.

I think scaling up to very, very high-end desktop parts will be an issue, yes, such as the 5950x for example. And the jury is obviously still out on how Apple tries to achieve that, as others have been discussing here (chiplets, huge monolith, etc.).

However, regarding those scores you posted, I will eat my hat if an 8x4 variant of the M1 isn't already in the general MT ballpark of the 5800X, and at less than 50w TDP.

EDIT - And to put this in context, I've been trying and failing to buy a 5800X for a gaming PC build. Zen 3 is a marvel, and I'm suspicious that anyone will be doing their PC gaming primary on an Apple chip in any possible timeline for the next decade. But I will still be surprised if my future M-series 16" MBP is not fully at parity with my (hopefully near-future) gaming desktop for raw CPU performance.

 

[guidryp]

Are you a semiconductor engineer by chance? That's not what I've heard. Adding more cores dramatically increases the complexity of the uncore design, because having a big multicore CPU isn't beneficial if the CPU doesn't scale well due to poor core to core communication or not being fed enough bandwidth.

Then you heard wrong.

You don't have be an engineer to know that increasing IPC is much more difficult than putting more cores on a piece of Silicon. Companies have placed 64 or more ARM cores on a Server chip with negligible history of CPU Design (Amazon) and done a credible job of it. Multiple cores is a well known and well solved issue. There is absolutly nothing stopping Apple from putting any reasonable number of cores on a chip to power Mac Pros. Though they may choose a multi-chip option when going beyond 16 cores for the Mac Pro.

OTOH, despite all the companies selling ARM based designs, it's really only Apple that has succeeded in producing designs with significantly higher IPC than the designs that ARM licences. Adding more cores is trivia compared to actually designing new cores that advance the state of the art on IPC.

 

[amrnuke]

And process doesn't really change IPC either.

I wanted to circle back to this, because someone said you're wrong. You are right. Process node does not increase IPC. It just gives you more real estate to add more logic/cache, the liberty to clock higher, the ability to use less power. But it doesn't directly affect IPC (cf. Zen -> Zen+, on which there was a 3% IPC gain from revised cache, but nothing from process).

 

[name99]

If you want to see advances in technology, you should not put your bets on the awful x86 ISA. Whatever you can do with x86 you can do better (with less power and/or higher performance) with ARM - indeed i do not believe that Apple engineers are inherently more capable than Intels or AMDs. The engineering teams working within the solution space given by the instruction set architecture specification.

Apple have been doing this for seven (or 8, or 10 years, depending on whether you want to start with the shipping of the A4, the A6, or the A7). But if we start with A7, over seven years they have
- boosted the single threaded speed of their cores by 6x (A7 to A14)
- shipped one, then two, now three different cores per year (large, small, tiny)
- expanded to world class GPU and NPU
- moved from 1B transistors (A7) to 16B (M1)
- moved from one manufacturer to another (SS to TSMC) sometimes using both
without a single serious delay or upset along the way.

I'm sorry, but 'i do not believe that Apple engineers are inherently more capable than Intels or AMDs" is false. You don't have to be an Apple fan, you just have to know engineering, to know that this is pretty much the most impressive performance in human history. (Though good for TSMC, who have been a stellar partner!)

To look at this record and say "well, they haven't yet shown they can do a good job scaling to more cores, or to a server" is not caution, it's sticking your head in the sand. It is a deliberate choice to be ignorant and blind.

Most of the people commenting so confidently about Apple here have no clue about APple history. Their vague awareness of actual Apple engineering (as opposed to stupid jokes by fellow idiots) began maybe a year ago. They're saying what they're saying with zero knowledge of this history, but absolute conviction that they know the truth, the whole truth.

 

[Eug]

Makes a difference, doesn't it? Even if you don't care about benchmarks, you can feel the speed difference.
I agree that no normal person upgrades devices every year (and anyone who constructs their reviews or analysis around this behavior is a mendacious idiot), but the two year difference is noticeable, and by three years you really want to upgrade if you can afford it.

Yes, although I think I may try to keep onto the 4-year cycle for my wife and myself for the time being. I did notice a bit of lag at the 3 year mark with my 7 Plus once I upgraded to iOS 13, but be honest I think it was more related to RAM than CPU speed. Cuz for navigation and such it was fine, but every so often I'd just get this 1-2 second lag of nothing happening, which I think may be due to insufficient RAM, as it would be reloading something into memory.

OTOH, my wife with her XR from 2 years ago notices no lag at all, despite the fact it has the same 3 GB RAM, so maybe I'm wrong, or else the faster SoC hides this better. However, my wife usually doesn't notice these things until it becomes really obvious so I'm not sure.

I wonder though if the same will be true 3 years from now, since I now have 6 GB iPhone, when the bulk of Apple's iPhones sold only have 4 GB RAM. After a certain point there are diminishing returns.

I note the on the Mac side, upgrading from 1 GB to 2 GB was huge, and upgrading from 2 GB to 4 GB was also huge, whereas upgrading from 4 GB to 8 GB was a more moderate improvement for light users, and upgrading from 8 GB to 16 GB is a much less of a noticeable improvement for basic mainstream usage. I have a 16 GB 2017 12" MacBook, but in retrospect I could have just gotten an 8 GB model, giving me an excuse to upgrade to a 16 GB Apple Arm MacBook model sooner.

tl;dr:

Using recent history as a guide, your 3-year upgrade cycle advice for iPhones makes sense in terms of performance, but it may be due just as much to RAM as opposed to just CPU speed. Then again, many mainstream users like my wife would be OK with a 4 year-old iPhone in terms of speed.

 

[amrnuke]

100% agree there. Hopefully we'll have answers soon.

For what it's worth, I too am skeptical that the M1 cores can be scaled up to 5950X type MT workloads. It remains to be seen, of course, but I wouldn't be surprised if Apple's desktops never quite catch up with the best from AMD or whoever. On the other hand, their absolute advantage in mobile (and by that I mean up to an including AIO forms like the iMac) should be concerning to the x86 world, because if the advantage is large enough that Microsoft and OEMs push groups like Qualcomm and Nuvia to supply Arm chips for laptops, it's very possible that these x86 chips only have an advantage in the extremely narrow high-power desktop market. I'm not sure you can maintain the levels of R&D budget and innovation if you start to bleed the mobile market.

I agree that the high end market is going to be tough to break into for Arm licensees, who are fairly far behind AMD and Apple even by IPC estimates, let alone raw performance. Apple aren't likely to be too interested in licensing their cores for the HEDT/server/HPC market, or for that matter, entering those markets on their own. We'll see what SD 875 (?Arm X1?) can bring.

Apple have been riding the high end of the frequency/voltage train since A12 and I really don't see how we can just assume they'll be able to clock all that much higher with an unrestricted power. The A14 appears to have sorted out some power issues that the A13 had. But as an example, for the A12, check out this graph from AT's deep dive on A13. If you trend out from 2.76W -> 3.85W, and 2.43 GHz -> 2.5 GHz, and disregard exponential growth and just take a straight line, you see that with Apple's uarch, Arm ISA, and TSMC N7, if you want to eek out 3 GHz on N7 with all that, you're looking at 11.6W, even assuming linear and not exponentially-increasing power draw which is wrong, but underestimates power consumption. But it starts to become really clear that it's not rudimentary for Apple (or anyone, for that matter) to design a chip that handles frequency increases with any kind of grace. My impression is that for A13 and A14 Apple are likely riding the high end of this F/V curve and similar statements likely apply to those chips. A key question is whether this inflection point is a characteristic of the ISA on any given node, or whether it's a characteristic of the microarchitecture on any given node, or how closely those two are intertwined.

What I find really interesting is that the 5600X achieves 96.6% of the 5950X's SPECint2006 score at 54.5% of the per-core power consumption and 40.8% of the total package power consumption. It truly makes me wonder what will happen in the mobile space with Zen3.

If you plot total package power vs SPECint2006 score for the A14, 5600X, 5800X, 5900X, 5950X and slap a trendline on there, performance scaling down to lower frequencies/power draws for Zen3 seems to offer some promise in relation to A14... Granted with such few data points it's hard to make any significant judgment on that front.

 

[IvanKaramazov]

I agree that the high end market is going to be tough to break into for Arm licensees, who are fairly far behind AMD and Apple even by IPC estimates, let alone raw performance. Apple aren't likely to be too interested in licensing their cores for the HEDT/server/HPC market, or for that matter, entering those markets on their own. We'll see what SD 875 (?Arm X1?) can bring.

Apple have been riding the high end of the frequency/voltage train since A12 and I really don't see how we can just assume they'll be able to clock all that much higher with an unrestricted power. The A14 appears to have sorted out some power issues that the A13 had. But as an example, for the A12, check out this graph from AT's deep dive on A13. If you trend out from 2.76W -> 3.85W, and 2.43 GHz -> 2.5 GHz, and disregard exponential growth and just take a straight line, you see that with Apple's uarch, Arm ISA, and TSMC N7, if you want to eek out 3 GHz on N7 with all that, you're looking at 11.6W, even assuming linear and not exponentially-increasing power draw which is wrong, but underestimates power consumption. But it starts to become really clear that it's not rudimentary for Apple (or anyone, for that matter) to design a chip that handles frequency increases with any kind of grace. My impression is that for A13 and A14 Apple are likely riding the high end of this F/V curve and similar statements likely apply to those chips. A key question is whether this inflection point is a characteristic of the ISA on any given node, or whether it's a characteristic of the microarchitecture on any given node, or how closely those two are intertwined.

What I find really interesting is that the 5600X achieves 96.6% of the 5950X's SPECint2006 score at 54.5% of the per-core power consumption and 40.8% of the total package power consumption. It truly makes me wonder what will happen in the mobile space with Zen3.

If you plot total package power vs SPECint2006 score for the A14, 5600X, 5800X, 5900X, 5950X and slap a trendline on there, performance scaling down to lower frequencies/power draws for Zen3 seems to offer some promise in relation to A14... Granted with such few data points it's hard to make any significant judgment on that front.

Yes to all that. I actually think the Cezanne debut, whenever it happens (4800u really only came out a couple of months ago, and still is hard to come by in shipping laptops) will be a very interesting comparison, and probably shape some of the narrative around the Apple chips.

 

[amrnuke]

Yes to all that. I actually think the Cezanne debut, whenever it happens (4800u really only came out a couple of months ago, and still is hard to come by in shipping laptops) will be a very interesting comparison, and probably shape some of the narrative around the Apple chips.

Apple are the team to beat on mobile and likely the team to beat on mid-entry level laptops (pending reviews). I hope there's competition with Cezanne and Arm's X1, and so on. It's looking like a tough hill to climb.