Apple only “major” device maker on 3nm in 2023?

Cmaier

Site Master
Staff Member
Site Donor
Posts
5,215
Reaction score
8,264
What exactly is different? "FinFlex" is still FinFET, just with gate-by-gate control over the 2-1 / 2-2 / 3-2 configuration. It is Samsung that is attempting GAAFET on N3.
Good point. I forgot that this node isn’t GAA.
 

theorist9

Site Champ
Posts
603
Reaction score
548
But it feels weird because you’re already proposing a clock speed increase that doesn’t exist yet for an M2 chip that doesn’t exist so you’re resisting adding more clock speed for the M3 version that doesn’t yet exist :) despite that being one of the primary advantages of moving to a new node (or lower power for the same clock speed but this is the Mac Pro, so not the directions they’d go in).
Nothing weird about it at all ;). I just needed a plausible upclock figure for the M3, and chose 4.2 GHz because it's a value that's already been rumored for the Mac Pro by someone who gave pretty good predictions about the Studio. The fact that he was rumoring it for the M2 rather than the M3 didn't matter for my purposes since, as I said, this was just a Fermi calcuation.
 

dada_dave

Elite Member
Posts
2,070
Reaction score
2,047
Nothing weird about it at all ;). I just needed a plausible upclock figure for the M3, and chose 4.2 GHz because it's a value that's already been rumored for the Mac Pro by someone who gave pretty good predictions about the Studio. The fact that he was rumoring it for the M2 rather than the M3 didn't matter for my purposes since, as I said, this was just a Fermi calcuation.
Ah but for the Fermi calculation the primary limiting factor should be the power calculation - how much extra power would it take for Apple to hit that score on the M3 SC and is that reasonable for the device in question given Apple’s design principles and goals? So you should choose a wattage to fix the core at, not a GHz.
 

theorist9

Site Champ
Posts
603
Reaction score
548
Ah but for the Fermi calculation the primary limiting factor should be the power calculation - how much extra power would it take for Apple to hit that score on the M3 SC and is that reasonable for the device in question given Apple’s design principles and goals? So you should choose a wattage to fix the core at, not a GHz.
Nope. "Fermi calculation" just means "back-of-the-envelope calculation". So the calculation should suit my purposes, which was to see if it was plausible one could get to that score with an M3, which it did. I.e., I wanted to start with the score as a target, and estimate what power that needed. There's nothing about a Fermi calc that requires you to start with the power as a target instead. If you personally want to take that different approach, feel free to do so.
 
Last edited:

dada_dave

Elite Member
Posts
2,070
Reaction score
2,047
Nope. "Fermi calculation" just means "back-of-the-envelope calculation".
I know what Fermi calculation means.

So the calculation should suit my purposes, which was to see if it was plausible one could get to that score with an M3, which it did. I.e., I wanted to start with the score as a target, and estimate what power that needed. There's nothing about a Fermi calc that requires you to start with the power as a target instead. If you personally want to take that different approach, feel free to do so.
You misunderstood my post. I apologize if I was unclear. Yes the score is the target value. But you’re trying to estimate the power needed to attain that score and instead you got hung up on a fixed GHz and then decided it was not plausible because such a score required too much IPC uplift to be plausible. What I’m trying to say is that for a given score, the GHz is only relevant so far as how much power it takes to get there. You constrained yourself at 4.2 GHz when you should constrain yourself either by Watts or IPC.

What you wrote here is the right approach:

I wanted to start with the score as a target, and estimate what power that needed.

So here’s how I would’ve done it using your ^4 approximation:

A) assume a 15% IPC increase - very reasonable, even beatable (of course caveat being IPC is also a function of clock speed, but that’s a nuance we’ll ignore)

Score jumps up by 1.15. So clocks have to increase by 1.6/1.15 = 1.4 roughly. Now we plug that into the power: almost 4x power (really 3.7) of a Mac Pro core compared to a iPhone core. AMD it especially Intel? Sure. Apple? Less likely. Possible for a MacPro but still feels unlikely. Aha! But! The process node gets us 15% of that clock speed increase for free. So the actual increase in power needed is [1.6/(1.15*1.15)]^4 = 2.14x iPhone core power. Much more plausible for an Apple designed chip, even a workstation one.

B) You could go the other way too which is similar to your approach, but instead of a set GHz, it’s a set power - let’s say you allow for a 2x iPhone core power. Then, to attain a 60% score increase would require a 17% IPC uplift … not 31%. 2= [1.6/(1.17x1.15)]^4

That’s what I’m trying to say you’re effectively missing a factor in your calculations that lead you to the wrong conclusion that the IPC uplift is unlikely - which 31% would be!

Of course the whole thing is moot since as @Colstan says there’s no way this A17 score was a real test. But as a fun thought experiment you were correct that if Apple does plan on increasing clock speeds to a Mac Pro chip, this score would be plausibly attainable by an M3 MacPro.
 
Last edited:

theorist9

Site Champ
Posts
603
Reaction score
548
But you’re trying to estimate the power needed to attain that score and instead you got hung up on a fixed GHz and then decided it was not plausible because such a score required too much IPC uplift to be plausible....

That’s what I’m trying to say you’re effectively missing a factor in your calculations that lead you to the wrong conclusion that the IPC uplift is unlikely - which 31% would be!
Nope, I didn't get hung up on a fixed GHz (inded, you'll see from the quote below that it's exactly the opposite), nor did I miss a factor, nor did I conclude IPC uplift is unlikely. It seems you've managed to misinterpret a large fraction of what I wrote.

Again, I choose 4.2 GHz as a target, and intially said you'd need an additional 30% from IPC + process to get to the 3986 score. But then I added a correction saying you wouldn't have IPC + process, you'd only have IPC, because process is subsumed into IPC and clock (initially you argued the point with me but finally agreed).
EDIT: Upon further consideration, an ~30% generational increase in IPC may be unlikely. But that just means that, if this result is real, Apple was experimenting with even higher clocks than 4.2 GHz.
[Emphasis added] I.e., I realized that, since a 30% IPC uplift was unlikely, you'd probably need a bit more than 4.2 GHz to reach the target score. Do I really have to get into parsing English with you to explain that acknowleding you'd need higher clocks is the opposite of getting "hung up on a fixed GHz"?? Further, in no way did I say that such was implausible. I've no idea where you're getting that from.

Indeed, I'd considered redoing my calculation with higher clocks and a lower IPC, but it didn't seem worth it, since this whole thing was such a rough estimate anyways. After all, I'm not even sure what the exponent is supposed to be.

Either way, I figured my conclusion stands: By limiting the higher clock (whatever it is) to 2–4 cores, and possibly adding an algorithm that limits their use to when the GPU isn't being stressed then, based on my rough upper bound calculation, you can probably deal with the added TDP in Macs besides the MP (Studio, Mini, 16" MBP) (in the MP you'd of course be even less constrained).

Nor was I missing a factor for efficiency improvment due to the M3's (expected) improved N3 process, since that was already subsumed into my starting point, which was what the M3's WPC would be at 3.66 GHz:
Suppose that the M3 P-cores consume 5 WPC (watts per core) when running at 3.66 GHz
Yes, I could have done a more detailed calculation that started with the M2 and the tried to add an efficiency improvement for going from M2 and M3. But I didn't want to. It was simpler to just start with an estimate for the M3's TDP, and proceed from there. You say this:
I know what Fermi calculation means.
...but you don't seem to appreciate its spirit. The point is to do a very rough, simple calculation, requiring as little input as possible. That's what I did, and yet you're quibbling with me about small details that are not needed for a very coarse-grained analysis. You seem to think I missed something—got things wrong—because I didn't include those details. No. Excluding those details was a choice. I feel like I'm back at MacRumors!
 
Last edited:

dada_dave

Elite Member
Posts
2,070
Reaction score
2,047
If I misunderstood your posts I apologize but there seems to be that on both sides.
Nope, I didn't get hung up on a fixed GHz (inded, you'll see from the quote below that it's exactly the opposite), nor did I miss a factor, nor did I conclude IPC uplift is unlikely. It seems you've managed to misinterpret a large fraction of what I wrote.

Again, I choose 4.2 GHz as a target, and intially said you'd need an additional 30% from IPC + process to get to the 3986 score. But then I added a correction saying you wouldn't have IPC + process, you'd only have IPC, because process is subsumed into IPC and clock (initially you argued the point with me but finally agreed).

No. I argued the point that process and IPC improvements nets you performance gains without increasing power. Thus they need to be separated from clocks speed increases that require extra power of you’re going to calculate how much power it would take to get a given score. Yes process improvement in this instance also means clock speed increases but it separate but it doesn’t require extra juice to be fed into the processor.

You argued that you chose 4.2 as a target because that’s what the equivalent M2 chip would have. So I took your later quote to mean “experimenting with higher clocks” to mean higher power when in fact the M3 variant would have a higher clock anyway than 4.2 if that’s what the M2 is to be clocked at. This may be part of our mutual misunderstanding.

[Emphasis added] I.e., I realized that, since a 30% IPC uplift was unlikely, you'd probably need a bit more than 4.2 GHz to reach the target score. Do I really have to get into parsing English with you to explain that acknowleding you'd need higher clocks is the opposite of getting "hung up on a fixed GHz"?? Further, in no way did I say that such was implausible. I've no idea where you're getting that from.

You just said you considered a 30% IPC lift unlikely. Which I agree with. That’s what I was saying you were saying is implausible. And as I showed in my last post if you don’t put the 15% improvement from the process, the result of the power calculations looks a lot like an AMD/Intel chip core which I said was implausible even if technically feasible.

Indeed, I'd considered redoing my calculation with higher clocks and a lower IPC, but it didn't seem worth it, since this whole thing was such a rough estimate anyways. After all, I'm not even sure what the exponent is supposed to be.

Either way, I figured my conclusion stands: By limiting the higher clock (whatever it is) to 2–4 cores, and possibly adding an algorithm that limits their use to when the GPU isn't being stressed then, based on my rough upper bound calculation, you can probably deal with the added TDP in Macs besides the MP (Studio, Mini, 16" MBP) (in the MP you'd of course be even less constrained).

Nor was I missing a factor for efficiency improvment due to the M3's (expected) improved N3 process, since that was already subsumed into my starting point, which was what the M3's WPC would be at 3.66 GHz:

Yes, I could have done a more detailed calculation that started with the M2 and the tried to add an efficiency improvement for going from M2 and M3. But I didn't want to. It was simpler to just start with an estimate for the M3's TDP, and proceed from there.

This where I think the disagreement is the M3 TDP or perhaps should say thought the disagreement was since this was not clear from your posts or calculations. It seemed to me that you were starting from the M2 power of 5W at 3.66 GHz and extrapolating that to 4.2GHz whereas the M3 would be expected to start with either lower power or higher clocks. I believe this is where the confusion lies. I didn’t see where the process improvement was in your estimate. Which as I said above and in the previous post actually changes the result of the Fermi calculation from plausible to implausible if it isn’t there.

You say this:

...but you don't seem to appreciate its spirit. The point is to do a very rough, simple calculation, requiring as little input as possible. That's what I did, and yet you're quibbling with me about small details that are not needed for a very coarse-grained analysis. You seem to think I missed something—got things wrong—because I didn't include those details. No. Excluding those details was a choice.

The version I did is still a simple estimate where the purpose is to estimate the power cost to get a higher score. Again I felt explicitly including process was less a detail and more a key point that if left out qualitatively changed the result - which is important for a Fermi calculation. However, you’re saying you didn’t leave it out and I just missed it. If so, I apologize but I didn’t and, to be honest, still don’t see it.

I feel like I'm back at MacRumors!

There’s no need to be insulting. We’re having what I thought was an interesting technical disagreement. Again if I’ve misunderstood your posts I apologize. If you don’t want to continue this conversation I understand.
 
Last edited:
Top Bottom
1 2