Rumored pro chip

[thingstoponder]

I don’t think it’s fair to say the M1 was designed for the lowest common denominator. I suspect it was designed for the Macs and just happens to work fine in the ipad. I don’t think they said “gee, would be nice to put this on the chip for macs, but since the iPad Pro can’t make use of it never mind.”

I don’t think that’s the case at all. The M1 is basically an a14X. Same big core count, same small core count, same GPU core count, and same neural engine core count as the a14. Apple could have never transitioned the Mac and this same chip would exist solely for the iPad Pro.

 

[thingstoponder]

Continuing on from my previous post...


The 2545 Geekbench number seems stupid and ridiculous and fantasyland stuff, but when you look at it as simple inconsequential numbers, it’s weirdly easily doable if we look at what Apple has been able to achieve over the last decade with their A series chips.

View attachment 1779815

M1 has a current single core of 1700 or so...a cumulative improvement in all or some using a combination of...

IPC improvements of 10 to 15 percent...?

TDP/Frequency increase of 10 percent from 3.2GHz to 3.5GHz...?

Higher Memory Bandwidth going from LPDDR4x to LPDDR5 of 50 percent ...?

***( LPDDR5 or HBME2)

TSMC 5nm updated PDK for maybe another 7.5 to 12.5 percent ...?



Apple M2 high performance core
1700 x 1.5 = 2550.


I’m estimating a 50 percent improvement by adding up all the smaller improvements everywhere, but maybe these are fantasy numbers for a single core. Even from Apple.

Maybe when 3nm exists which is another year away. M3 generation. M2 generation will be 5nm and be iterative.

 

[cmaier]

I don’t think that’s the case at all. The M1 is basically an a14X. Same big core count, same small core count, same GPU core count, and same neural engine core count as the a14. Apple could have never transitioned the Mac and this same chip would exist solely for the iPad Pro.

That’s simply false. First, there is no a14x, so claiming the m1 has the “same core counts” is meaningless. M1 has more cores, more cache, more physical address space, more memory channels, and higher frequency than the A14. Whereas an A14X would have the same physical core design as an A14 (based on every prior X variant), the M1 clearly has completely different physical core designs than the A14.

It’s so weird when people who have never designed a CPU get all hung up on the fact that two die have the same core microarchitecture and then claim that they are “the same.:

 

[Joelist]

I’m not an expert, only read about this stuff here and there but apparently a popular way to bun things in mobile is by power rather than by performance (which is what we know from desktop). Basically, as I understand it, one defines a performance limit and kind of tries to match it, even if it means that some chips will end up using more power and some end up using less power. The end effect of this is that some phones have slightly better battery life and some slightly worse, but it’s not something a user is likely to notice. Besides, mobile chips tend to have more complex power delivery (because efficiency matters more), and can dynamically tweak the needs of individual components better. So if you have a slightly inferior chip whose GPU needs more power to operate, yiu can usually supply that power without much problem.

For M1 the idea is that Apple takes chips where the GPU needs too much power to operate and disables one core to bring the power consumption to the expected numbers. I think this makes more sense than the defective core theory because M1 GPU cores are very small it’s it’s quite unlikely that a defect hits a GPU core and not something else.



Yes, but we still need to be a bit careful here, because this advantage does not apply to every situation. There are still pipeline delays and data dependencies, and M1 low frequency can become a measurable disadvantage when running code with low ILP.

Hi leman!

While it is plausible a program with a very low ILP could be somewhat disadvantaged, most programs today are not like that. In fact the trends for development for a while now have been towards more parallelism which plays right into Apple Silicon's hands. This was even true back when Core 2 hit the streets (which was why I included the little recall from that history). Also Apple has their thumb on the own OS and Apps and also provides their dev tools, and in both cases you can bet they are going for as high an ILP as they possibly can.

 

[thingstoponder]

That’s simply false. First, there is no a14x, so claiming the m1 has the “same core counts” is meaningless.

The M1 has the same core counts as the a12x/a12z. You know that's what I meant.

If the M1 wasn't designed for the iPad Pro at all and it was just a happy accident that it ended up there then what do you think they would have done for the iPad Pro? Make a separate 4+4+8 core die? That would make zero sense as it would be the same chip. Make a smaller die that has less cores than the a12z? They would never do that. Put an a14 in it? The iPad Air already has it. Apple isn't just winging it. They knew from day one designing the M1 that it would be in the iPad Pro as well as half their Mac lineup. They probably had that planned for years.

M1 and A14X are marketing names. Yes the A14X doesn't exist because Apple decided not to call it that, but if the Mac transition never happened then this same chip would be called A14X.

M1 has more cores, more cache, more physical address space, more memory channels, and higher frequency than the A14.

Just like the X variant of A-series chips always did.

Whereas an A14X would have the same physical core design as an A14 (based on every prior X variant), the M1 clearly has completely different physical core designs than the A14.

No it doesn't? Do you have a source? Its the same architecture. They're not going to waste resources with two architectures. Why do you think the a14 chip supports hardware virtualization for the first time now? Because it's the same architecture and making two different architectures just for Mac features is a waste of time and money.

It’s so weird when people who have never designed a CPU get all hung up on the fact that two die have the same core microarchitecture and then claim that they are “the same.:

And you have designed a CPU?

 

[cmaier]

No it doesn't? Do you have a source? Its the same architecture. They're not going to waste resources with two architectures. Why do you think the a14 chip supports hardware virtualization for the first time now? Because it's the same architecture and making two different architectures just for Mac features is a waste of time and money.



And you have designed a CPU?

1) yes it does. it’s easy to see just by looking at die photos. I never said it was a different architecture - I said it was a different physical design. And I don’t know why you are talking about architecture - the issue is microarchitecture. Obviously it has the same architecture - these are all Arm.

2) yes, i have designed many CPUs. PowerPC x704, UltraSparc V, AMD K6-II, AMD K6-III, AMD Opteron, AMD Athlon 64, etc. etc.

 

[thingstoponder]

1) yes it does. it’s easy to see just by looking at die photos. I never said it was a different architecture - I said it was a different physical design. And I don’t know why you are talking about architecture - the issue is microarchitecture. Obviously it has the same architecture - these are all Arm.

2) yes, i have designed many CPUs. PowerPC x704, UltraSparc V, AMD K6-II, AMD K6-III, AMD Opteron, AMD Athlon 64, etc. etc.

The instruction set is Arm. By architecture I mean the generation of cores aka Firestorm, Lightning, Twister, etc.

Yes the layout is a little different, that doesn’t make it a different microarchitecture. X chips always had more cache and bandwidth and memory channels to feed the GPU. The difference between the M1 and the A14 isn’t any different than the A12 and A12X/Z.

The code name of the die is even in the same family as the a14 and previous X-series iPad chips. People have even dug up new chip code names that have a completely separate naming scheme. These are the dedicated Mac chips.

 

[dgdosen]

That’s assuming the cores are maxed out all the time when doing work.

You have to remember this chip was designed for the lowest common denominator, the iPad Pro. The rumored 10 core chip is the first chip designed only for the Mac. It’s possible that if the M1 was designed only for the Mac it would have 2 efficiency cores.

Either way I’m sure Apple knows what they’re doing with this whole silicon design thing. I suspect there will be no issues with 2 cores.

I keep going back to the reports that Apple swooped in to pick up additional capacity at TSMC when the Huawei embargo left TSMC a hole in their 5nm schedule. Back at that time, maybe Apple only had the M1 ready to go. If that was the case, Apple would have a glut of M1s (not a bad problem to have) and would have to figure out where to put them. I have a feeling we'll see more M1 devices before it's over...

 

[cmaier]

The instruction set is Arm. By architecture I mean the generation of cores aka Firestorm, Lightning, Twister, etc.

Yes the layout is a little different, that doesn’t make it a different microarchitecture. X chips always had more cache and bandwidth and memory channels to feed the GPU. The difference between the M1 and the A14 isn’t any different than the A12 and A12X/Z.

The code name of the die is even in the same family as the a14 and previous X-series iPad chips. People have even dug up new chip code names that have a completely separate naming scheme. These are the dedicated Mac chips.

What you call “architecture” is “microarchitecture.” And, as I said, the physical design is different. I never claimed that the microarchitecture is different. But A12X and a12z have the same core physical designs as A12, for example. Whereas M1 and A14 have completely different core physical designs (they are not “a little different.”)

You constantly pointing out that they have the same microarchitecture is pointless - I said that was the case from the get-go. What I also said is, unlike previous x-chips, the M1 has a different physical design than the A14.

 

[reallynotnick]

We already have the info from Gurman. 8 performance cores, 2 efficiency cores, 16 or 32 GPU cores.

Gurman seemed to say these are two different chips but everyone seems to think it’s one chip that’s binned. I think it’s two chips. I think the 14” gets 16 GPU cores abd 16” gets 32.

I agree it has to be two separate chips, that's way too aggressive of a cut for binning. If it were 24 or more GPU cores I could see it as binning. I wouldn't be surprised if we saw binned versions of those chips though with like 14 and 28 GPU cores or something that. That then would give the 14" and 16" two different chips to choose from.

 

[CWallace]

I don’t think it’s fair to say the M1 was designed for the lowest common denominator. I suspect it was designed for the Macs and just happens to work fine in the ipad. I don’t think they said “gee, would be nice to put this on the chip for macs, but since the iPad Pro can’t make use of it never mind.”

As you have (tried to) note(d), if the M1 never existed, the A14X would have had 4 Firestorm cores, 4 Icestorm cores and 8 GPU cores as the M1 does, but it would have physically been different from the M1.

I do believe Apple always planned to use M1 in the iPad Pro and future iPad Pros will also use the M-series SoC. While I do not believe Apple is trying to "merge" the two lines into a common product family, the iPad Pro is branching off from the iPhone more and more and I believe it benefits from the "Mac" parts Apple is putting in the M1 that were not in previous A-X SoCs.

We already have the info from Gurman. 8 performance cores, 2 efficiency cores, 16 or 32 GPU cores.

Gurman seemed to say these are two different chips but everyone seems to think it’s one chip that’s binned. I think it’s two chips. I think the 14” gets 16 GPU cores abd 16” gets 32.

I think it’s definitely based off the M1/a14 generation also. People who think they’re going to debut a new generation of architectural cores with this chip just have no clue what they’re talking about. That chip would be at least 9-12 months away, not right on the horizon for a possible WWDC unveiling.

A15 should already be in early production to support the iPhone 12s/13 so Apple could conceivably have a new M-series SoC using the same various cores. But I find it more likely it will be more Firestorm cores, more GPU cores and more Neural Engine cores.

 

[theorist9]

Someone correct me if I'm wrong, but it seems the barrrier Apple would have to overcome in switching from 4 efficiency cores to 2 isn't performance per se, it's maintaining close to linear scaling of power while doubling performance.

Performance isn't a barrier--Apple can easily make a new core that has 2x the performance of an Icestorm core. After all, the M1's Firestorm cores have about 4x the performance of its Icestorm cores, so if performance were the only concern, they could have replaced the 4 Icestorms with 1 Firestorm, and designated that for the background/utility/low priority tasks.

Instead the barrier is mainitaining power/performance while scaling to higher power. In the M1, to achieve the needed power/performace for its efficiency cores, Apple was limited to the current Icestorm's performance. However, if they're able to use 2 cores instead of 4, it means they've managed to double the performance while maintaining (close to) linear scaling of power consumption

 

[cmaier]

Someone correct me if I'm wrong, but it seems the barrrier Apple would have to overcome in switching from 4 efficiency cores to 2 isn't performance per se, it's maintaining close to linear scaling of power while doubling performance.

Performance isn't a barrier--Apple can easily make a new core that has 2x the performance of an Icestorm core. After all, the M1's Firestorm cores have about 4x the performance of its Icestorm cores, so if performance were the only concern, they could have replaced the 4 Icestorms with 1 Firestorm, and designated that for the background/utility/low priority tasks.

Instead the barrier is mainitaining power/performance while scaling to higher power. In the M1, to achieve the needed needed power/performace for its efficiency cores, Apple was limited to the current Icestorm's performance. However, if they're able to use 2 cores instead of four, it means they've managed to double the performance while maintaiing (close to) linear scaling of power consumption.

Exactly. That’s why I think these can’t be ice storms but have to be something new.

 

[thingstoponder]

What you call “architecture” is “microarchitecture.” And, as I said, the physical design is different. I never claimed that the microarchitecture is different. But A12X and a12z have the same core physical designs as A12, for example. Whereas M1 and A14 have completely different core physical designs (they are not “a little different.”)

You constantly pointing out that they have the same microarchitecture is pointless - I said that was the case from the get-go. What I also said is, unlike previous x-chips, the M1 has a different physical design than the A14.

I know what microarchitecture is. People often just say architecture. Yes the M1 and a14 have different layouts. They’re different chips. Do you have some proof that it is different in a way that the a12 and a12x/z wasn’t? I’m having trouble finding an a12x die shot that compares it directly to an a12.

My main point is that the M1 is not a ground up Mac chip that they decided to throw in an iPad last minute. It’s an iPad chip that they decided to target the Mac with as well. If it was a ground up Mac chip it wouldn’t have limitations like only 1 display.

If there was any doubt then just look at the internal code names.

A12 - t8020
A12X/Z - T8027
A13 - t8030
A14 - t8101
M1 - t8103

https://twitter.com/i/web/status/1357015237623431168

It’s the same family of chips. The engineers making the chips see it as a direct continuation of the a12x. M1 is just a consumer marketing name.

Do you think they were cooking up an iPad Pro specific A14X chip that they later canceled because the M1 was sufficient? You still haven’t answered this. I’d love to know how they would make an a14X that would be any different than the M1. Maybe it would have less cores than the A12X/Z. Seems unlikely to me.

 

[thingstoponder]

and I believe it benefits from the "Mac" parts Apple is putting in the M1 that were not in previous A-X SoCs.

There aren’t any special Mac parts from what I can tell. Maybe Thunderbolt if you count that. I’ve heard a lot of people assume that the M1 has dedicated accelerators for Final Cut or other Apple software but I don’t see any proof of this. It’s just a fast chip and the media engine in their GPU is fantastic.

A15 should already be in early production to support the iPhone 12s/13 so Apple could conceivably have a new M-series SoC using the same various cores. But I find it more likely it will be more Firestorm cores, more GPU cores and more Neural Engine cores.

I agree that they have early versions in their labs but that’s a long ways away from being a consumer product. A bigger Mac chip on a new generation of IP isn’t coming before the a15 or M2, especially less than a year before the a14 came out.

 

[Rickroller]

Someone correct me if I'm wrong, but it seems the barrrier Apple would have to overcome in switching from 4 efficiency cores to 2 isn't performance per se, it's maintaining close to linear scaling of power while doubling performance.

Performance isn't a barrier--Apple can easily make a new core that has 2x the performance of an Icestorm core. After all, the M1's Firestorm cores have about 4x the performance of its Icestorm cores, so if performance were the only concern, they could have replaced the 4 Icestorms with 1 Firestorm, and designated that for the background/utility/low priority tasks.

Instead the barrier is mainitaining power/performance while scaling to higher power. In the M1, to achieve the needed power/performace for its efficiency cores, Apple was limited to the current Icestorm's performance. However, if they're able to use 2 cores instead of 4, it means they've managed to double the performance while maintaining (close to) linear scaling of power consumption

Wow! When you put it that way...then imagine four of these littler beasties in the next low power chips?

 

[Rickroller]

I’m wondering how the input from Apples Pro team is factoring in the decisions they make in regards to the hardware they’re designing/designed. I think the pandemic has had to have some influence on what a mobile pro device should be able to do performance wise. Maybe the reason for beefing up the GPU much more than the 16 Core that was expected for the M1X?

 

[cmaier]

I’m wondering how the input from Apples Pro team is factoring in the decisions they make in regards to the hardware they’re designing/designed. I think the pandemic has had to have some influence on what a mobile pro device should be able to do performance wise. Maybe the reason for beefing up the GPU much more than the 16 Core that was expected for the M1X?

They sort of gave away that they were planning to do this way back at WWDC last year, in the session where they talked about treating the M1 GPU as if it were discrete. And they’ve been working on a stand-alone GPU die for at least 2 years.

 

[Joelist]

So far the next chip seems like it will be pretty beefy. We already have notions of what they are doing on CPU and GPU. It makes one wonder what is in store for ML, the other assorted blocks, controllers, cache and so on. Also whether or not they elect to go to an even faster RAM type and so on.

 

[dgdosen]

While the estimated die sizes for these new chips is small enough to fit within TSMC's packaging, does anyone think Apple and TSMC might still be working together to become a 'model' of using TSMCs 3D stacking (silicon on silicon)?

Might there be easy footprint/perf/power gains (and challenges) from going more 3D?

 

[Rickroller]

While the estimated die sizes for these new chips is small enough to fit within TSMC's packaging, does anyone think Apple and TSMC might still be working together to become a 'model' of using TSMCs 3D stacking (silicon on silicon)?

Might there be easy footprint/perf/power gains (and challenges) from going more 3D?

this is what I about possible 3D earlier


“I’ve focused on the M2 high performance core, but I’m guessing/ assuming the same performance uplift will apply to the high efficiency cores as well. That means for the new M2 8+2 SIP, you would technically have 3 M1 efficiency cores in the package, which I guess is enough for the intended target products.



My guess is also that there is one basic SIP to rule them all, let’s call it ‘Jade C-Die’ containing the 8P+2E + 32GPU cores. Chop some GPU cores for a tighter fitting 14 inch premium suit or lower priced 16 inch tracksuit.



As for how Apple will get to the 2 or 4 Jade punch combo, we might need to consider 3D glasses.



I think Apple will introduce a 3D stacked die.



Some links about Apple 3D and patents like these...”



https://coreteks.tech/articles/inde...-density-and-flexibility-through-3d-stacking/



https://www.macrumors.com/2018/06/12/apple-3d-chip-packaging-patents/



https://semiaccurate.com/2020/12/07/apple-gets-creative-with-their-upcoming-cpus/

 

[cmaier]

this is what I about possible 3D earlier


“I’ve focused on the M2 high performance core, but I’m guessing/ assuming the same performance uplift will apply to the high efficiency cores as well. That means for the new M2 8+2 SIP, you would technically have 3 M1 efficiency cores in the package, which I guess is enough for the intended target products.



My guess is also that there is one basic SIP to rule them all, let’s call it ‘Jade C-Die’ containing the 8P+2E + 32GPU cores. Chop some GPU cores for a tighter fitting 14 inch premium suit or lower priced 16 inch tracksuit.



As for how Apple will get to the 2 or 4 Jade punch combo, we might need to consider 3D glasses.



I think Apple will introduce a 3D stacked die.



Some links about Apple 3D and patents like these...”



https://coreteks.tech/articles/inde...-density-and-flexibility-through-3d-stacking/



https://www.macrumors.com/2018/06/12/apple-3d-chip-packaging-patents/



https://semiaccurate.com/2020/12/07/apple-gets-creative-with-their-upcoming-cpus/

When you don’t need to use an MCM, don’t use an MCM. I remain skeptical. It’s possible that they may do it for reasons of supply chain flexibility or yields, but the reticle can handle die >800mm squared, which should be big enough for everything rumored so far. Guess we’ll see.

(I invented a 3D chip stacking technique in 1996, so I’m all for 3D chip stacking, but in this case I don’t yet see the need)

 

[Rickroller]

When you don’t need to use an MCM, don’t use an MCM. I remain skeptical. It’s possible that they may do it for reasons of supply chain flexibility or yields, but the reticle can handle die >800mm squared, which should be big enough for everything rumored so far. Guess we’ll see.

(I invented a 3D chip stacking technique in 1996, so I’m all for 3D chip stacking, but in this case I don’t yet see the need)

That’s why it’s great when people with your knowledge can help steer a complex conversation about technology.

So a question for you would when be the time for 3D stacking...?

what if I were to ask you to think up a chip/system that extends past the MBP iMacs and Mac Pro, to some other areas of their business that would require a much larger order of magnitude of these same chips...? The only one that’s easily discernible is to host their own stuff internally...They can give you just about unlimited R&D budget, and won’t baulk at crazy new ideas. Maybe for a start you may just need a target of cost parity with whoever is hosting their servers now...?

I can’t even begin to imagine that contract and numbers on it?

sorry for the jump around in thoughts, but I’d like to think it’s an effort to create something so much more ambitious than an M1X 8+4+16 for such relatively low volume products.

 

[theorist9]

The thing is, we still don't know the real limit of Firestorm. Apple seems to do something different here, as they don't bin the chips by max stable frequency, only by power (presumably). It is entirely possible that already Firestorm is able to run at 3.5-3.7ghz and power consumption of 10 watt per core instead of 5 watts per core. So a prosumer Apple Silicon might end up bringing a bigger boost than what would be expected from the process improvement alone.

If Apple doubles the TDP of the big cores in their highest-performing chips, that would made it even more unlikely we'd see those chips in the iMac—you'd need the thermals of the Mac Pro.

But yeah, I think that 20% in single core is a reasonable estimate. That would firmly put Apple on top of any current or expected CPU (overclocked models notwithstanding) until end of 2022 at least.

According to the Bloomberg article, some increase in per-core speed will be seen in the new Air ("Staten" chip). That would be embarrassing for AMD and Intel to have an Air beat their fastest stock chips for single-core speed.

 

[jeanlain]

M1 and A14 have completely different core physical designs (they are not “a little different.”)

Why do you mean by "core physical design". Both SoCs have Firestorm and Icestorm cores, right?