News Intel and AMD forge x86 ecosystem advisory group that aims to ensure a unified ISA moving forward

[TheHerald]

OK I did some rough estimations based on alderlake. A 12900k is a total of 4.2b transistors. Removing the igpu displays etc and just focusing on the cores + ringbus, that takes 54% of the die, which means 2.26b transistors.

The m3 is 25b transistors in total. The cores are around 22% of the die (including the amx accelerator). That's 5.5b transistors, and even if we exclude the amx it still is bigger than alderlake


So the m3 cores are 2.5 times larger than the 12900k. At same power, they be 2.5 times faster. I'm leaving the power draw out of this comparison (for now, we will get back to it) just for simplicity. Is there any cpu workload appropriate for a comparison between the 2? Checking cbr23 it looks like a 12900k is 2.7 times faster, so each x86 transistor performance 7 times as much as an arm transistor. There is no way in hell that this gap will close by dropping power. Even at 35w the 12900k is around 40% faster, while being freaking tiny in comparison

So what the heck am I missing?

 

[bit_user]

But that's still irrelevant. If the m3 core is 55 times as big then it makes sense that it's faster and more efficient.

Except Apple M3's P-core is only 2.49 mm^2, while Lunar Lake's Lion Cove P-core is 4.53 mm^2 or 81.9% bigger. Both on the same node: TSMC N3B It really doesn't get much more apples-to-apples than that!

To put it in your terms, the M3's P-core is actually 0.550 times as big, which I find quite a coincidence!

Source:

https://www.reddit.com/r/hardware/comments/1fuuucj/lunar_lake_die_shot/

Now, let's see if you can be honest, for once, and acknowledge the truth.

OK I did some rough estimations based on alderlake.

Aren't you Mr. iso-comparisons? Why are you comparing to a core on a completely different node? You could also compare the power-to-weight ratio of a riding lawnmower to a sports car and it would be about as useful.

 

[TheHerald]

Except Apple M3's P-core is only 2.49 mm^2, while Lunar Lake's Lion Cove P-core is 4.53 mm^2 or 81.9% bigger. Both on the same node: TSMC N3B It really doesn't get much more apples-to-apples than that!

To put it in your terms, the M3's P-core is actually 0.550 times as big, which I find quite a coincidence!

Source:

​

https://www.reddit.com/r/hardware/comments/1fuuucj/lunar_lake_die_shot/

​

Now, let's see if you can be honest, for once, and acknowledge the truth.


Aren't you Mr. iso-comparisons? Why are you comparing to a core on a completely different node? You could also compare the power-to-weight ratio of a riding lawnmower to a sports car and it would be about as useful.

Well I made the estimations. Let's see is you can be honest for once and acknowledge the truth.

 

[bit_user]

Well I made the estimations. Let's see is you can be honest for once and acknowledge the truth.

What the heck truth are you even talking about?

Like I said, your thought experiment was completely irrelevant. You can go back and look performance vs. transistor count for all the Intel cores in history and you'll see a disturbing trend where IPC increases a lot slower than transistor count. That's just the way it is, and it's why you ought to focus on comparing cores on as similar a node as possible.

This shows just how faithless you really are. You asked a question, got an answer, and now you're throwing a tantrum like a baby, because it's not the answer you wanted.

 

[TheHerald]

What the heck truth are you even talking about?

Like I said, your thought experiment was completely irrelevant. You can go back and look performance vs. transistor count for all the Intel cores in history and you'll see a disturbing trend where IPC increases a lot slower than transistor count. That's just the way it is, and it's why you ought to focus on comparing cores on as similar a node as possible.

This shows just how faithless you really are. You asked a question, got an answer, and now you're throwing a tantrum like a baby, because it's not the answer you wanted.

I did some calculations comparing alder vs M3. Alderlake wiped the floor with m3 in transistor count / performance, which you clearly disliked so now you are throwing a tantrum. Oh well. It's fine, you can ignore the evidence.

If we start adding ecores and zen5c to the comparison, oh boy.

 

[bit_user]

I did some calculations comparing alder vs M3. Alderlake wiped the floor with m3 in transistor count / performance,

Which is irrelevant.

You're just making up bizarre comparisons, because you don't like what the straight-forward, obvious, and most consequential ones reveal. Maybe you should've done some of your own digging, before you opened this Pandora's Box, but now it's too late. You can't paper this over with phony statistics.

 

[TheHerald]

Which is irrelevant.

You're just making up bizarre comparisons, because you don't like what the straight-forward, obvious, and most consequential ones reveal. Maybe you should've done some of your own digging, before you opened this Pandora's Box, but now it's too late. You can't paper this over with phony statistics.

How is it irrelevant? If anything your comparison is irrelevant. Let's just say for the sake of argument the numbers you posted are absolutely 100% accurate. Doesn't mean anything, cause you not comparing x86 to arm, you are comparing Lunar lake specifically. Lunar Lake might be the worst cpu in terms of Trans / perf so you obviously can't reach conclusions between arm and x86 by using the worst.

I used alderlake which isn't even particular good in Trans / perf and it still dismantled the m3. Imagine if I actually just used the ecores or zen 5c, it would be devastating

 

[bit_user]

Let's just say for the sake of argument the numbers you posted are absolutely 100% accurate. Doesn't mean anything, cause you not comparing x86 to arm, you are comparing Lunar lake specifically.

Yes. It's just Lunar Lake, which contains Intel's newest and most efficient CPU cores vs. Apple M3. That's all I ever said it was. I never claimed it proved anything about conclusive about x86 vs. ARM.

Lunar Lake might be the worst cpu in terms of Trans / perf

Why this funny new obsession of yours with performance per transistor? Like I said, if you go back through Intel's history, you're going to find them getting steadily worse, on this metric.

I used alderlake which isn't even particular good in Trans / perf and it still dismantled the m3.

If you compared it worth an Apple core on the same or older node, the Apple would trounce it. It's pointless. It's just a ridiculous comparison you dreamed up to try and change the subject.

 

[TheHerald]

Yes. It's just Lunar Lake, which contains Intel's newest and most efficient CPU cores vs. Apple M3. That's all I ever said it was. I never claimed it proved anything about conclusive about x86 vs. ARM.


Why this funny new obsession of yours with performance per transistor? Like I said, if you go back through Intel's history, you're going to find them getting steadily worse, on this metric.


If you compared it worth an Apple core on the same or older node, the Apple would trounce it. It's pointless. It's just a ridiculous comparison you dreamed up to try and change the subject.

It is indeed a ridiculous comparison. X86 is so much faster at same trans counter (7 times, lol) that there is no reason to even compare them. They are on a different sport.

But sure, I'll give you the benefit of the doubt. I'll compare M2 with zen 5c or ecores. Is that okay with you? Will you accept the results afterwards or you'll come up with a new excuse?

 

[bit_user]

X86 is so much faster at same trans counter (7 times, lol)

No.

I'll compare M2 with zen 5c or ecores. Is that okay with you?

What are you hoping to see that hasn't already been demonstrated? How is it any more predictive to go further back in time?

Will you accept the results afterwards or you'll come up with a new excuse?

I think M2 was TSMC N5P, so even comparing to Zen 4 isn't as good as the comparison between M3 and Lunar Lake.

Pretty much the next closest equivalent would probably be Ryzen Z1 Extreme. That has 4x Zen 4 + 4x Zen 4C cores and is made on TSMC N4. According to the link I posed above, its full size cores are each 3.84 mm², compared to the M2's P-core at 2.76 mm². Performance and efficiency data on it is a little less easy to come by.

Probably the next best point of comparison after that or Strix Point would be Meteor Lake?

But, Gracemont? That's even weirder and more pointless than comparing it to Golden Cove.

Seriously, why do you have to keep picking this scab at all? Just lick your wounds and walk away. There are other threads you can haunt.

 

[TheHerald]

No.


What are you hoping to see that hasn't already been demonstrated? How is it any more predictive to go further back in time?


I think M2 was TSMC N5P, so even comparing to Zen 4 isn't as good as the comparison between M3 and Lunar Lake.

Pretty much the next closest equivalent would probably be Ryzen Z1 Extreme. That has 4x Zen 4 + 4x Zen 4C cores and is made on TSMC N4. According to the link I posed above, its full size cores are each 3.84 mm², compared to the M2's P-core at 2.76 mm². Performance and efficiency data on it is a little less easy to come by.

Probably the next best point of comparison after that or Strix Point would be Meteor Lake?

But, Gracemont? That's even weirder and more pointless than comparing it to Golden Cove.

Seriously, why do you have to keep picking this scab at all? Just lick your wounds and walk away. There are other threads you can haunt.

You are the one licking wounds man, it's already established that alderlake , which is particularly bad in trans / performance absolutely beats the m3 to a pulp. You literally are the one making up excuses for why the m3 gets annihilated and then pretend like you won the argument, lol


The funny part is that your excuse makes no sense. Somehow a node shrink from Intel 7 to TSMC 4 causes a CPU to require 7 times as many transistors for the same performance? That's your argument pretty much, lol. Then why do companies want to shrink their nodes? They would be looking to enlarge them if that was the case. Can you even pack 7 times as many transistors on TSMC 4 compared to Intel 7? Most likely no, not even close.

Anyways, conclusion is, m3 needs 2.7 times the transistors to end up at 1/3 the performance of your average, 3 year old x86 chip. Arm is great, for sure.

 

[waltc3]

Wanted to add the fact that it's been many, many years since Intel or AMD CPUs were actually x86 CPUs. In current "x86" CPUs today, actual x86 ISA code is a tiny portion of the CPUs, and is only there for backwards compatibility with much older software that required actual x86 code in the CPUs. The only people who still think that Intel and AMD are still selling real x86 ISA CPUS are Mac folks, apparently...Apple is always 20 years behind when they talk about Intel and x86... It's been the rule for Apple for the last 30 years, at least. (The exception being the brief period when Jobs put Intel CPUs into Macs for a while...then, even Apple was "x86.")

So why keep calling it "x86"? Because even though both AMD and Intel CPUs have tiny portions of their transistor densities for the purpose of compatibility with x86 software, and "x86" is simply a convenient term for older software & hardware compatibilities. I've got games 30-35 years old that run like a top on my AMD 3900X. That's the value of backwards compatibility.

 

[bit_user]

absolutely beats the m3 to a pulp.

Source? What independent data supports this assertion?

I posted my data. ST & MT performance (on CB24 + GB6) and efficiency on same + core sizes.

The funny part is that your excuse makes no sense. Somehow a node shrink from Intel 7 to TSMC 4 causes a CPU to require 7 times as many transistors for the same performance?

Hey, just look at how many more transistors Lion Cove is using!

The point of increasing transistor count is that it lets you boost IPC, which helps increase performance without destroying efficiency. This is standard stuff. It's why transistor count has increased at an exponential rate, since the dawn of CPUs.

​

Then why do companies want to shrink their nodes? They would be looking to enlarge them if that was the case.

Huh?

 

[TheHerald]

Source? What independent data supports this assertion?

I posted my data. ST & MT performance (on CB24 + GB6) and efficiency on same + core sizes.


Hey, just look at how many more transistors Lion Cove is using!

The point of increasing transistor count is that it lets you boost IPC, which helps increase performance without destroying efficiency. This is standard stuff. It's why transistor count has increased at an exponential rate, since the dawn of CPUs.

​

​

Huh?

There are no assertions. I measured the die shots of m3 and 12900k, did the math, and ended up at the conclusion that m3 needs 7 times as many transistors to achieve the same performance as a 3 year old x86 chip. Grannted, it's doing it with less power, but even if you normalize for power youll still end up at 3-4 times more transistors. Unless m3 specifically is terrible and there are other arm chips that can do better, arm looks like it's incredibly inefficient , underachieving in both efficiency and performance based on their transistor count.

You are now arguing with facts my man. M3 literally uses 2.7 times more transistors for 1/3 of the performance of a 3 year old x86 chip, it doesn't freaking get more obvious than that.

 

[stuff and nonesense]

Wanted to add the fact that it's been many, many years since Intel or AMD CPUs were actually x86 CPUs. In current "x86" CPUs today, actual x86 ISA code is a tiny portion of the CPUs, and is only there for backwards compatibility with much older software that required actual x86 code in the CPUs. The only people who still think that Intel and AMD are still selling real x86 ISA CPUS are Mac folks, apparently...Apple is always 20 years behind when they talk about Intel and x86... It's been the rule for Apple for the last 30 years, at least. (The exception being the brief period when Jobs put Intel CPUs into Macs for a while...then, even Apple was "x86.")

So why keep calling it "x86"? Because even though both AMD and Intel CPUs have tiny portions of their transistor densities for the purpose of compatibility with x86 software, and "x86" is simply a convenient term for older software & hardware compatibilities. I've got games 30-35 years old that run like a top on my AMD 3900X. That's the value of backwards compatibility.

Transmeta Crusoe was the key, dismantle the CISC instructions and pass them through a wrapper to a code agnostic RISC core, these days cores. The wrapper soon got copied and adopted by Intel/AMD.

Instruction decoders are baffling to me. So much complexity juggling what registers are pretending to be whatever type of register the programmer or the compiler decided they should be based on whether they are 8 bit, 16 bit, 32 bit, 64 bit values, while taking whatever language and translating it into some short simple processor commands. To my limited brain, simply amazing.

 

[bit_user]

There are no assertions. I measured the die shots of m3 and 12900k, did the math, and ended up at the conclusion that m3 needs 7 times as many transistors to achieve the same performance as a 3 year old x86 chip.

First, your estimates are a joke. A bad one. Second, how did you establish performance?

Third, however bad you're saying the M3 is just makes Lunar Lake look that much worse. Lions Cove is 82% bigger than the M3's P-core and yet the M3 has 15% better ST performance, so you're really throwing Lunar Lake in the gutter!

You are now arguing with facts my man.

They're not facts. They're sketchy estimates you made and wouldn't hold up well to scrutiny, except that the premise is so deeply flawed it's not even worth trying to check them.

 

[bit_user]

Transmeta Crusoe was the key, dismantle the CISC instructions and pass them through a wrapper to a code agnostic RISC core, these days cores. The wrapper soon got copied and adopted by Intel/AMD.

Intel released the Pentium Pro, about 5 years earlier. That was the first time I ever read about Intel using RISC-like micro-ops.

Note that microcode != micro ops. Microcode is a more general and squishy term, but if we're talking about what pops out of instruction decoders, those can be something implemented in ways other than RISC-like instructions.

 

[TheHerald]

First, your estimates are a joke. A bad one. Second, how did you establish performance?

Third, however bad you're saying the M3 is just makes Lunar Lake look that much worse. Lions Cove is 82% bigger than the M3's P-core and yet the M3 has 15% better ST performance, so you're really throwing Lunar Lake in the gutter!


They're not facts. They're sketchy estimates you made and wouldn't hold up well to scrutiny, except that the premise is so deeply flawed it's not even worth trying to check them.

Sure, lunar lake is despicable. That doesn't change anything regarding x86 vs arm. That's why you are deflecting

 

[bit_user]

Sure, lunar lake is despicable. That doesn't change anything regarding x86 vs arm. That's why you are deflecting

We already established this wasn't a conclusive referendum on x86 vs. ARM. It was just about looking at one very current and relevant matchup, that's about as apples-to-apples as you can possibly get!

 

[stuff and nonesense]

Intel released the Pentium Pro, about 5 years earlier. That was the first time I ever read about Intel using RISC-like micro-ops.

Note that microcode != micro ops. Microcode is a more general and squishy term, but if we're talking about what pops out of instruction decoders, those can be something implemented in ways other than RISC-like instructions.

Fair enough… btw not talking about microcode.

The idea Crusoe targeted was CISC -> abstracted “native” through a translation layer.

 

[TheHerald]

We already established this wasn't a conclusive referendum on x86 vs. ARM. It was just about looking at one very current and relevant matchup, that's about as apples-to-apples as you can possibly get!

If it's not, and you admit it's not, then what's the point of even bringing it up? Repeatedly?

 

[stuff and nonesense]

We already established this wasn't a conclusive referendum on x86 vs. ARM. It was just about looking at one very current and relevant matchup, that's about as apples-to-apples as you can possibly get!

FYI
Don’t know how reputable the site is but for what it’s worth

https://nanoreview.net/en/cpu-compare/intel-core-i9-12900k-vs-apple-m3

https://nanoreview.net/en/cpu-compare/intel-core-i9-12900k-vs-apple-m3-max

I’ll leave you to draw your own conclusions

 

[TheHerald]

FYI
Don’t know how reputable the site is but for what it’s worth

https://nanoreview.net/en/cpu-compare/intel-core-i9-12900k-vs-apple-m3

https://nanoreview.net/en/cpu-compare/intel-core-i9-12900k-vs-apple-m3-max

I’ll leave you to draw your own conclusions

I found this review that is a bit more comprehensive.

View: https://www.youtube.com/watch?v=buLyy7x2dcQ


The m2 ultra (2023 chip) against a 13900k (2022) + 4090 (2022). The m2 ultra is 134b transistors, the 13900k + 4090 are below 100.

Obviously the m2 ultra was slower on pretty much everything, sometimes by a huge margin (like 3-4 times!!!). Of course the m2 ultra uses less power and we have to normalize for that as well but I can't see how restricting the 13900k to 65w will drop it's performance by 3-4 times. It won't.

Linus has another review done with 3090 + 12900k vs m1 ultra with similar results

View: https://www.youtube.com/watch?v=8YjMIjLLIwA

 

[-Fran-]

Dang. Good thing I wasn't drinking my tea. The hallucinations are wild from the rogue* AI.

Regards.

 

[bit_user]

I found this review that is a bit more comprehensive.
...
The m2 ultra (2023 chip) against a 13900k (2022) + 4090 (2022). The m2 ultra is 134b transistors, the 13900k + 4090 are below 100.

I don't make any claims about their GPU performance. Like I said, if you want to compare GPU performance, that's a different thread, because now you're not talking about x86 or ARM, whatsoever.