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

[bit_user]

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

If people aren't familiar, the M1/M2/M3 is the base model. It has a 4+4 configuration, like Lunar Lake. The idea of comparing it to an Alder Lake i9 is a bit silly. It should be compared to something more like an i3. The Pro is roughly equivalent to the i5. The Max is their i7 tier. The Ultra is their i9 tier.

Anyway, I don't really trust comparison sites like that. Some of those sites have approximate data the produced using a numerical performance model they fit to a handful of benchmark results, rather than using real test data for each CPU they list.

 

[-Fran-]

If people aren't familiar, the M1/M2/M3 is the base model. It has a 4+4 configuration, like Lunar Lake. The idea of comparing it to an Alder Lake i9 is a bit silly. It should be compared to something more like an i3. The Pro is roughly equivalent to the i5. The Max is their i7 tier. The Ultra is their i9 tier.

Anyway, I don't really trust comparison sites like that. Some of those sites have approximate data the produced using a numerical performance model they fit to a handful of benchmark results, rather than using real test data for each CPU they list.

Not to go against, but perhaps a better parallel would be the CCDs in AMD's design. That's how I've always seen the building blocks of the M line, but on steroids.

Regards.

 

[TheHerald]

If people aren't familiar, the M1/M2/M3 is the base model. It has a 4+4 configuration, like Lunar Lake. The idea of comparing it to an Alder Lake i9 is a bit silly. It should be compared to something more like an i3. The Pro is roughly equivalent to the i5. The Max is their i7 tier. The Ultra is their i9 tier.

I can't see the logic in that. Just because the M3 is the smaller chip apple makes doesn't make it comparable to the smallest chip someone else makes, since the transistor difference between an i3 12100 is like what, 10x?

 

[bit_user]

I can't see the logic in that. Just because the M3 is the smaller chip apple makes doesn't make it comparable to the smallest chip someone else makes, since the transistor difference between an i3 12100 is like what, 10x?

I never said to compare specifically the M3 to the i3-12100. I just said it's their lower tier.

The M-series are SoCs. They contain NPU, ISP, iGPU, etc. So, you can't go by gross transistor counts between them. I know you don't care about being fair, but I'm just saying... if you did.

Actually, rather than compare to desktop CPUs, the Gen 12 product it would make more sense to look at is something like the i3-1220P, although it's 28W. The iGPU is a little bit closer and they actually have an ISP and a tiny NPU.

 

[TheHerald]

I never said to compare specifically the M3 to the i3-12100. I just said it's their lower tier.

The M-series are SoCs. They contain NPU, ISP, iGPU, etc. So, you can't go by gross transistor counts between them. I know you don't care about being fair, but I'm just saying... if you did.

Actually, rather than compare to desktop CPUs, the Gen 12 product it would make more sense to look at is something like the i3-1220P, although it's 28W. The iGPU is a little bit closer and they actually have an ISP and a tiny NPU.

The only thing that would throw off my comparison (m3 ultra vs 13900k + 4090) is if the 4090 was insanely better in transistor / performance compraed to the m3 ultra. Is that the case?

 

[Kamen Rider Blade]

That's not what Keller said. He said he asked ARM to add a couple of custom instructions to convert between the data formats used by their Tensix DSP cores, but ARM said "no". Then, he turned to RISC-V, because it lets you do whatever the heck you want.

That was also one of the sticking points as well.

AMD has a custom IP business, which is how console makers gain use of their IP.

It isn't cheap either to hire them for their custom IP.

Intel has also talked about licensing their cores to fab customers, in which case I'd bet you might be able to get a few custom tweaks.

And what will it cost to do it? Probably more than what a StartUp like TensTorrent can afford to pay.

The first few gens of their chips with RISC-V used IP licensed from SiFive. It was only later that he decided to start designing their own, possibly as an alternate business model, in case the whole AI thing didn't work out for them.

But they plan on having a CPU & NPU as well, so either way, they decided to implement their own CPU.

 

[palladin9479]

The ISA doesn't really matter once you get past ultra low power devices. All uArchs run into the exact same scaling issues and use the same solutions.

Also you absolutely can not compare an Intel 12-14th gen CPU vs Apple Mx as they are on very different process nodes. Apple paid TSMC a ridiculous amount of money to have first access to their newest process nodes. This gives them an amazing performance efficiency advantage for a few years until Intel catches up with TSMC. Only for TSMC to have an even newer process.

 

[bit_user]

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 Alder Lake-S die is 215.25 mm^2, of which a Golden Cove P-core occupies 7.53 mm^2 or 3.50%. If we make a simplistic assumption that all transistors are equal size (which they're not), then that works out to approximately 147 Million Transistors.

The M3's P-core is only 2.49 mm^2 and just 1.71% of its 146 mm^2 die. That works out to approximately 426 Million Transistors, which is only 2.90x as many as Golden Cove.

The M3 gets a CB R23 (ST) score of 1904 points at 10 W, for an efficiency of 190.4 points/W. (source: https://www.notebookcheck.net/Apple...ormance-and-improved-efficiency.766789.0.html )

The i9-12900K gets a CB R23 (ST) score of 2030 (source: https://www.techpowerup.com/review/intel-core-i9-13900k/6.html ). I didn't find CPU-only power figures for that, but TechPowerUp's MP3 encode measured 32 W. That would work out to 63.4 points/W, which is almost exactly 1/3rd as efficient as the M3.

So, the M3 is 3x as efficient (ST) with a P-core having approx 2.9x as many transistors. That's a net improvement of 3.4% per Transistor.

So the m3 cores are 2.5 times larger than the 12900k.

Nope. To see why, consider that the i9 has twice as many P and E cores. So, if we assume the E-cores compare similarly as the P-cores did, then the M3 should actually spend just 1.45x as many transistors on its cores as the i9-12900K. So, your estimate was over by 72.4%.

 

[bit_user]

Apple paid TSMC a ridiculous amount of money to have first access to their newest process nodes. This gives them an amazing performance efficiency advantage for a few years until Intel catches up with TSMC. Only for TSMC to have an even newer process.

Apple's M3 is on the exact same TSMC N3B node as Lunar Lake. And they both have 4P + 4E cores, with no SMT. It is a very direct comparison.

To your point about new nodes, the M3 is old and about to be replaced by the M4, whereas Lunar Lake just launched 1 month ago.

Also you absolutely can not compare an Intel 12-14th gen CPU vs Apple Mx as they are on very different process nodes.

100% agree with this, but TheHerald was insistent on that comparison and I at least wanted to check his math.

 

[TheHerald]

The Alder Lake-S die is 215.25 mm^2, of which a Golden Cove P-core occupies 7.53 mm^2 or 3.50%. If we make a simplistic assumption that all transistors are equal size (which they're not), then that works out to approximately 147 Million Transistors.

The M3's P-core is only 2.49 mm^2 and just 1.71% of its 146 mm^2 die. That works out to approximately 426 Million Transistors, which is only 2.90x as many as Golden Cove.

The M3 gets a CB R23 (ST) score of 1904 points at 10 W, for an efficiency of 190.4 points/W. (source: https://www.notebookcheck.net/Apple...ormance-and-improved-efficiency.766789.0.html )

The i9-12900K gets a CB R23 (ST) score of 2030 (source: https://www.techpowerup.com/review/intel-core-i9-13900k/6.html ). I didn't find CPU-only power figures for that, but TechPowerUp's MP3 encode measured 32 W. That would work out to 63.4 points/W, which is almost exactly 1/3rd as efficient as the M3.

So, the M3 is 3x as efficient (ST) with a P-core having approx 2.9x as many transistors. That's a net improvement of 3.4% per Transistor.


Nope. To see why, consider that the i9 has twice as many P and E cores. So, if we assume the E-cores compare similarly as the P-cores did, then the M3 should actually spend just 1.45x as many transistors on its cores as the i9-12900K. So, your estimate was over by 72.4%.

But you literally just compared core per core, (my numbers were even higher than yours btw, lol), According to the numbers you just posted - m3 sucks in both efficiency and performance. You obviously know performance and power don't scale linearly but you literally skipped past it, lol. What's going to happen if you drop the Pcore to 10w as well? It's gonna lose what, 10-15-20% performance? So it's going to end up looking like the m3 is 20% faster while using 3 times the transistors. And that''s without taking into account the ecores which would tip the scales even further.

You are spinning the facts man...Even with the my calculations ending up at twice the transistor count for the 12900k it still ended up above and beyond the m3. Like seriously, don't you think having 3x as many transistors and ending up 20% faster at iso power is catastrophically bad, especially when compared to a 3 year old CPU - only the pcores of that cpu - and adding on top of that the Pcores are quite well known to being HUGE in the first place?

Now calculate the whole CPU like I did and youll see that the m3 is laughably bad, it uses way more transistors for the cores while being slower. Even at ISO watt it will end up being slower. That's terrible man.

EG1. We don't even have to estimate actually. 13400f has the same cores as aldelrake. It's scoring 1800 points @ 11 watts.
https://www.techpowerup.com/review/amd-ryzen-7-7800x3d/7.html

Man does it look bad for the m3.

 

[bit_user]

What's going to happen if you drop the Pcore to 10w as well?

But that's not the PL of any i9-12000 CPU. They all will boost to nearly as much, on a ST workload. Anyway, you're the one who picked the i9-12900K, not me. Blame yourself.

And that''s without taking into account the ecores which would tip the scales even further.

Okay, let's go all-core then.

The M3 got a CB23 (MT) score of 10454 @ ~27.9 W for an efficiency of 375 point per Watt.
Stock i9-12900K got a score of 27216 @ 257 W (using TechPowerUp's Blender numbers from above link) for 105.9 points per Watt.
So, MT performance is an even bigger win for the M3, at 3.54x of the i9-12900K. And that's in spite of lacking SMT.

Also, by image analysis, I found the Apple M3's E-core to be 31.7% as big as its P-core (source: https://www.tomshardware.com/news/annotated-apple-m3-processor-die-shots-bring-chip-designs-to-life ). This is roughly on-par with the ratio of Alder Lake's core sizes, which are about 30.4%, if we exclude cache in both cases. So, Apple isn't packing much larger E-cores, proportionally speaking.

Now, if you'd care to cite some CB23 (MT) numbers for i9-12900K @ 28W, then we could see just how they compare at iso-power.

You are spinning the facts man...

No, you've got us confused.

 

[TheHerald]

But that's not the PL of any i9-12000 CPU. They all will boost to nearly as much, on a ST workload. Anyway, you're the one who picked the i9-12900K, not me. Blame yourself.


Okay, let's go all-core then.

The M3 got a CB23 (MT) score of 10454 @ ~27.9 W for an efficiency of 375 point per Watt.
Stock i9-12900K got a score of 27216 @ 257 W (using TechPowerUp's Blender numbers from above link) for 105.9 points per Watt.
So, MT performance is an even bigger win for the M3, at 3.54x of the i9-12900K. And that's in spite of lacking SMT.

Also, by image analysis, I found the Apple M3's E-core to be 31.7% as big as its P-core (source: https://www.tomshardware.com/news/annotated-apple-m3-processor-die-shots-bring-chip-designs-to-life ). This is roughly on-par with the ratio of Alder Lake's core sizes, which are about 30.4%, if we exclude cache in both cases. So, Apple isn't packing much larger E-cores, proportionally speaking.

Now, if you'd care to cite some CB23 (MT) numbers for i9-12900K @ 28W, then we could see just how they compare at iso-power.


No, you've got us confused.

Im not comparing with the 12900k. Im comparing with x86. So using the 12900k power limits is irrelevant.

At 28w the 12900k doesnt even boot. At 35w (that's the lowest wattage it boots with) it scores around 12k.

But you skipped over those ST comparisons man. Did you see the 13400 numbers? 2.9x the transistor on the m3 for what, 15% more performance at iso power. That's horrible isn't it?

 

[bit_user]

Im not comparing with the 12900k. Im comparing with x86. So using the 12900k power limits is irrelevant.

The choice of product dictates the power limits.

At 28w the 12900k doesnt even boot.

What about i9-12950HX?

At 35w (that's the lowest wattage it boots with) it scores around 12k.

Source?

But you skipped over those ST comparisons man. Did you see the 13400 numbers?

F models don't count. It needs to be something with an iGPU. Your next best bet is the i5-12600. That gets 1866 and uses 15 W on MP3-encode, for 124 points/W.

Now, answer me this: if Alder Lake and Raptor Lake are so great, why did Intel bother to make Lunar Lake?

Intel seems quite clear about this:

​

So, either Intel is lying or you've just made another silly comparison. Which is it?

 

[bit_user]

You obviously know performance and power don't scale linearly but you literally skipped past it, lol. What's going to happen if you drop the Pcore to 10w as well? It's gonna lose what, 10-15-20% performance?

To be fair, if you're measuring efficiency and you cut the i9-12900K to 31.3% of its power, then you should also measure the M3 (on ST) using 3.1 W.

As you say, performance vs. power is nonlinear, so it's not fair to measure one product at the top of its power envelope and the other down at 1/3rd. But, that's just one of your usual tricks.

 

[TheHerald]

To be fair, if you're measuring efficiency and you cut the i9-12900K to 31.3% of its power, then you should also measure the M3 (on ST) using 3.1 W.

No it's not. It's called ISO power.

As you say, performance vs. power is nonlinear, so it's not fair to measure one product at the top of its power envelope and the other down at 1/3rd. But, that's just one of your usual tricks.

Of course it's fair to measure them at the same power. There is no such thing as a power envelope. CPUs don't have an intrinsic power draw. What wattage they use is a decision made by the marketing team.

It's literally you who is pulling tricks to not admit the obvious. This is really pathetic man.

The choice of product dictates the power limits.

No it doesn't. The exact same die as the 12900k exists in 4 or 5 (maybe even more) different power envelopes, from mobile to locked to T skus.

F models don't count. It needs to be something with an iGPU. Your next best bet is the i5-12600. That gets 1866 and uses 15 W on MP3-encode, for 124 points/W.

Why do F models not count? LOL. You literally insisted 5 titmes to leave the GPU out of the conversation. I mean still, even the 12600 absolutely nails the m3. 3x the transistors man, that's pathetic.

Now, answer me this: if Alder Lake and Raptor Lake are so great, why did Intel bother to make Lunar Lake?

Intel seems quite clear about this:

​

​

So, either Intel is lying or you've just made another silly comparison. Which is it?

I didn't say ald or raptor lake are great. Im saying the m3 is absolutely pathetic in transistors / performance and efficiency, and if that's the best arm can do, im not impressed. Looks horrible. Your own numbers prove it's horrible. 1 m3 core = 3P cores in terms of transistors. Obviously 3P cores are going to be a lot faster while using less power than the m3.

Source?

You have the CPU man, go test it, lol.


Just to sum up, in order for the m3 to be at least equal to an alderlake P core - the alderlake P core has to score ~650 points @ 10 watts in CBR23. Instead it scores 1800. Or,to go in reverse, the M3 P core needs to score 5700 points. Instead it score ~1950.

M3 needs 3 times the transistors for 10% more performance at the same power. That's embarrassing.

 

[bit_user]

No it's not. It's called ISO power.

Just because it has a name doesn't mean it makes sense to use in any context where you feel like trotting it out. If you're going to operate Alder Lake in some non-stock configuration that benefits its efficiency, then you need to do the same for the other product. Otherwise, the results are misleading. You don't want to mislead people, do you?

There is no such thing as a power envelope.

Yes, there is. If we're talking about the chip architecture, they all have a minimum and maximum power at which it will operate.

What wattage they use is a decision made by the marketing team.

That's a product-level designation and it no longer applies once you propose to run a CPU at other than stock power.

It's literally you who is pulling tricks to not admit the obvious.

The obvious fact that newer CPUs on smaller nodes use disproportionately more transistors for the performance they get? I've said that repeatedly - even including that nice Moore's Law graph. It's true of the M3, but it's even more true of Lunar Lake.

BTW, I was doing a little comparison with Skylake, when I noticed something fishy about your "4.8B Transistors" number. I had tried to find this figure for myself and only got one hit on google, where some rando posted a claim of 4.8B. It seems there's no official number, but what's known about Intel's 10 nm process is that it should be 100.76 MTr/mm^2. The 4.8 BTr figure you quoted works out to just 22.3 MTr/mm^2, which is only about 56% more dense than the original Skylake. Yet, according to the same graph, Intel 7 should be 2.26x as dense as 14 nm (and 2.71x as dense as 14 nm++, which chimes with what I recall).

TSMC 7nm HD and HP Cells, 2nd Gen 7nm, And The Snapdragon 855 DTCO

Update and analysis of TSMC 7-nanometer node low-power and high-performance cells, 2nd generation 7nm, and the design technology co-optimization (DTCO) effort that went into the Snapdragon 855 SoC.

fuse.wikichip.org

So, I'm pretty sure your entire house of cards comes tumbling down, if we had a real figure for Alder Lake. Even if we allow for the lower density of IOs and SRAMs and use a more conservative figure of half the measured density for Intel 7, that would still put Alder Lake (C0)'s transistor count at 10.8B, which is 2.26x as much as you claimed. That changes everything.

So, you really need to provide a good source on that 4.8B figure, or this entire tangent is completely out of gas.

No it doesn't. The exact same die as the 12900k exists in 4 or 5 (maybe even more) different power envelopes, from mobile to locked to T skus.

You picked the i9-12900K. Maybe you should've picked the i9-12900T, then. If we're comparing stock M3, then we should compare the other product at stock. It's only fair.

Why do F models not count? LOL.

Because the M3 was tested using an external monitor, meaning its display driver was active and its GPU was burning a nonzero amount of power.

You literally insisted 5 titmes to leave the GPU out of the conversation.

That was concerning GPU performance. However, if one set of CPU efficiency measurements is taken with an iGPU driving a monitor then the other should be as well.

I didn't say ald or raptor lake are great. Im saying the m3 is absolutely pathetic in transistors / performance and efficiency, and if that's the best arm can do, im not impressed. Looks horrible.

Lunar Lake is worse. That's the point, though. As bad as you claim the M3 is, it performs the same or better than Lunar Lake, while using less power.

You really cannot trash the M3 without doing even worse damage to Lunar Lake, because the M3 pwns it.

You have the CPU man, go test it, lol.

I do not have a i9-12900K. At work, I use a i9-12900, but it's running Linux and I can't just install random software on it.

Just to sum up,

1. You invented a new metric, because your old tricks no longer worked.
2. Nobody is judging products by such a metric - this is all you.
3. The new metric denies the impact of Moore's Law on CPU design.
4. The new metric is based on an unsourced and highly suspect claim of 4.8 B transistors for the C0 die.
5. You both acknowledged that perf/W is nonlinear, yet insist on measuring efficiency using a single point measurement that grossly favors one product.
6. Your first post in this thread wasn't even on topic. You were spoiling for a fight.
7. You got the answer you asked for, but it's clear your question wasn't in good faith, because you refused to accept it.

 

[bit_user]

Just to follow up on the question of transistor count, I found the Skylake i7 reportedly has 1.75B transistors on a 122.3 mm^2 die, which works out to 14.3 MTr/mm^2 or 32.0% of the graph's claimed 44.67 MTr/mm^2.

If we apply that same ratio to Alder Lake, then we get 6.95 B transistors, for the C0 die. That's 1.45x as much as claimed.

This lists the Skylake die sizes. It incorrectly implies the 1.75B figure applies to the i3 die, but I've elsewhere found a figure of about 1.4B transistors for that and confirmation that the 1.75B figure should've referred to the i7 die.

https://en.wikichip.org/wiki/intel/microarchitectures/skylake_(client)

Of course, just looking at the Skylake i7 and Alder Lake C0 dies, you can see that Alder Lake devotes much more of the the die to its CPU cores. So, I think the estimate based on Skylake's transistor mix is probably rather low.

 

[TheHerald]

Just because it has a name doesn't mean it makes sense to use in any context where you feel like trotting it out. If you're going to operate Alder Lake in some non-stock configuration that benefits its efficiency, then you need to do the same for the other product. Otherwise, the results are misleading. You don't want to mislead people, do you?

That's just silly. This isn't about alderlake or it's configuration. Since you acknowledge that power doesn't scale with power then the only fair way to test 2 products is at ISO power. Period. You are just making stuff up man. There is absolutely nothing unfair - quite the contrary - in comparing 2 products ceteris paribus. It's literally the one and only scientific method, we are trying to isolate for transistor count, everything else has to be equal. I get why you don't like the results but at least man up and admit that you don't like the results, don't try acting silly about it.

Yes, there is. If we're talking about the chip architecture, they all have a minimum and maximum power at which it will operate.

No they don't. There is no such thing as "maximum" power.

The obvious fact that newer CPUs on smaller nodes use disproportionately more transistors for the performance they get? I've said that repeatedly - even including that nice Moore's Law graph. It's true of the M3, but it's even more true of Lunar Lake.

Then say that instead of trying to act like alderlake doesn't completely obliterate the M3. The problem here is you are trying to skew facts (your own freaking numbers) to act like m3 is better than alderlake when it's clearly not. You could instead have said "YES, alderlake obliterates the crap out of m3, but that's because newer nodes use way more transistors". Then we could focus on that. But no, you are refusing to admit that in fact m3 needs 3 times as many transistors for 15% better ISO performance, which is HORRIBLE.

BTW, I was doing a little comparison with Skylake, when I noticed something fishy about your "4.8B Transistors" number. I had tried to find this figure for myself and only got one hit on google, where some rando posted a claim of 4.8B. It seems there's no official number, but what's known about Intel's 10 nm process is that it should be 100.76 MTr/mm^2. The 4.8 BTr figure you quoted works out to just 22.3 MTr/mm^2, which is only about 50% more dense than Skylake (if that). Yet, according to the same graph, Intel 7 should be 2.71x as dense as 14 nm++ (which chimes with what I recall).

​

TSMC 7nm HD and HP Cells, 2nd Gen 7nm, And The Snapdragon 855 DTCO

Update and analysis of TSMC 7-nanometer node low-power and high-performance cells, 2nd generation 7nm, and the design technology co-optimization (DTCO) effort that went into the Snapdragon 855 SoC.

fuse.wikichip.org

​

So, I'm pretty sure your entire house of cards comes tumbling down, if we had a real figure for Alder Lake. Even if we allow for the lower density of IOs and SRAMs and use a more conservative figure of half the measured density for Intel 7, that would still put Alder Lake (C0)'s transistor count at 10.8B, which is 2.26x as much as you claimed. That changes everything.

So, you really need to provide a good source on that 4.8B figure, or this entire tangent is completely out of gas.

GREAT, that's a lot more productive than the nonsense you came up with. If that number is wrong that changes everything, OBVIOUSLY. Problem is even when you believed those numbers are correct you still refused to accept that the m3 needing 3 times as many transistors is freaking BAD. Absolutely horrendous. Still my estimations on the cores are correct, 54% of the die size are the cores. If we can find the actual transistor count of the 12900k....

You picked the i9-12900K. Maybe you should've picked the i9-12900T, then. If we're comparing stock M3, then we should compare the other product at stock. It's only fair.

Why? They are the same chip. Power limits are completely utterly irrelevant.

Because the M3 was tested using an external monitor, meaning its display driver was active and its GPU was burning a nonzero amount of power.


That was concerning GPU performance. However, if one set of CPU efficiency measurements is taken with an iGPU driving a monitor then the other should be as well.

And that's completely irrelevant as well. TPU is using an external GPU, whether it's an F or a non F model doesn't matter, they are not using the iGPU. You are just making up stuff as you go man.

Lunar Lake is worse. That's the point, though. As bad as you claim the M3 is, it performs the same or better than Lunar Lake, while using less power.

You really cannot trash the M3 without doing even worse damage to Lunar Lake, because the M3 pwns it.

Sure, and I have no problem admitting it. That's the difference between you and me. 😆

Your own numbers (even if now it turns out they are wrong) showed how freaking terrible the m3 was, yet you refused to admit it, lol

1. You invented a new metric, because your old tricks no longer worked.
2. Nobody is judging products by such a metric - this is all you.
3. The new metric denies the impact of Moore's Law on CPU design.
4. The new metric is based on an unsourced and suspect claim of 4.8 B transistors for the C0 die.
5. You both acknowledged that perf/W is nonlinear, yet insist on measuring efficiency using a single point measurement that grossly favors one product.
6. Your first post in this thread wasn't even on topic. You were spoiling for a fight.
7. You got the answer you asked for, but it's clear your question wasn't in good faith, because you refused to accept it.

1) That's literally the only metric that matters when you cross comparing x86 with arm. What else would you use?
2) What metric are others using to compare x86 with arm and how are those other metrics more fair?
3) No it does not
4) Maybe that number is wrong. Still with your new 10b metric the end result does not change. Sure m3 gets close, but its still lagging behind alderlake.
5) Exactly, because it's not linear it makes 0 sense to compare 2 products at different power levels which is what you are trying to do. It's obvious to me that m3 is incredibly inefficient, that's exactly why you are trying to put it at a severe voltage curve advantage. Imagine how bad it is that you have to resort to that tactic. You are pulling the same with AMD chips. That doesn't inspire confidence about their efficiency either 😆

You are basically saying these cpus are so incredibly efficient that unless I put them at 3 times better voltage curve, they are going to get roflstomped. Oh well...
6) Irrelevant even if true.
7) No, you are the one refusing to accept even your own numbers man.

 

[TheHerald]

Just to follow up on the question of transistor count, I found the Skylake i7 reportedly has 1.75B transistors on a 122.3 mm^2 die, which works out to 14.3 MTr/mm^2 or 32.0% of the graph's claimed 44.67 MTr/mm^2.

If we apply that same ratio to Alder Lake, then we get 6.95 B transistors, for the C0 die. That's 1.45x as much as claimed.

This lists the Skylake die sizes. It incorrectly implies the 1.75B figure applies to the i3 die, but I've elsewhere found a figure of about 1.4B transistors for that and confirmation that the 1.75B figure should've referred to the i7 die.

​

https://en.wikichip.org/wiki/intel/microarchitectures/skylake_(client)

​

Of course, just looking at the Skylake i7 and Alder Lake C0 dies, you can see that Alder Lake devotes much more of the the die to its CPU cores. So, I think the estimate based on Skylake's transistor mix is probably rather low.

If we go by 7b - my 54% is correct (literally measured them on a die shot) that's still 3,78b for all 16 cores. M3 has no leg to stand on my man.

But as i've said, numbers here aren't the issue. The issue is that even when you thought those numbers are correct and that the m3 needed 290% the transistors for 15% better ISO performance, you still refused to acknowledge how terrible that is. So even if we get to the correct numbers, youll still come up with excuses if they don't match your beliefs. That's incredible...

 

[bit_user]

the only fair way to test 2 products is at ISO power. Period.

This is a canard. The way to test products is how people will actually use them. Intel does sell a laptop CPU with that same die. It's called the i9-12950HX and it doesn't run in the M3's power envelope.

Also, as we've discussed, it's ridiculous to use a product in a completely different class. You should be comparing the i9-12900-class CPU to something like the M3 Max or Ultra, where there's better price, power, and core-count parity. That's the proper way to compensate for the power limits being wildly different.

As I had said, if you wanted to compare older-generation Intel products against something like the M3, then you should use something based on the 6+8 core P-series.

There is absolutely nothing unfair - quite the contrary - in comparing 2 products ceteris paribus.

Using Latin doesn't change the fact that if your testing conditions are abnormal, then your test is irrelevant.

It's literally the one and only scientific method,

The proper design of an experiment is dependent on what you're trying to measure, which depends on why. You cannot separate them.

we are trying to isolate for transistor count,

You're trying to distract from the fact that you want to measure efficiency vs. transistor count. Efficiency is perf/W, which is a curve. Therefore, the proper experiment design should be to measure the perf/W curve of each product and compare those.

By picking a point that favors one product, you just end up with a misrepresentation. I know that's not what you want, so I don't understand why you disagree.

I get why you don't like the results

I don't like the proposition. There's no precedent for this, it has no practical consequence, and nobody seems to care about it except you.

You're clearly just trying to gin up a metric that makes Intel not look terrible, which is sadly the case when Lunar Lake is directly compared to the M3. I mean, we finally have the perfect apples-to-apples comparison and so I guess it shouldn't be a surprise that you throw a wild redirection.

And that's completely irrelevant as well. TPU is using an external GPU, whether it's an F or a non F model doesn't matter,

Okay, then we need data from somewhere that's measuring package power while using the iGPU.

1) That's literally the only metric that matters when you cross comparing x86 with arm. What else would you use?

No. Many people have benchmarked x86 vs. ARM and nobody is using performance per transistor.

2) What metric are others using to compare x86 with arm and how are those other metrics more fair?

Absolute performance and perf/W.

it makes 0 sense to compare 2 products at different power levels

It makes even less sense to compare two products in completely different market segments. What you're trying to do makes about as much sense as comparing an i3 to a ThreadRipper.

 

[TheHerald]

This is a canard. The way to test products is how people will actually use them.

I disagree with that as well but that's completely irrelevant anyways. We are not comparing products. We are comparing arm with x86. You are just playing games.

If you feel like transistor count comparison is meaningless then how do you determine what models to compare with each other? Who's to say that the m3 max competes with X cpu instead of Y?

I mean let me make it even easier. Why is amd selling the 7600x for 170$ and the 7995wx for 9.999$? Why aren't they selling the 7995 for 170$ as well?

As a customer i care about the performance I'm getting for the money I'm spending. If an arm cpu needs 25 times the transistors to match an x86 then it's going to cost 25 times as much money. In which case what's the point of even comparing the two, cause then I might as well buy the x86 chip that also costs 25 times as much and compare the arm chip with that one.

If Intel moves to arm and ends up needing 25b transistors to match the performance of the 12900k which is 7b transistors then I'm telling you it ain't gonna launch at 589$

 

[bit_user]

So, here's how ridiculous it is to compare performance per transistor and why I utterly reject that metric. According to the graph I posted above, Intel 14 nm++ has a density of 37.22 MTr/mm^2. Coffee Lake was made on that node. The i9-9900K die is 174 mm^2 and a single core is about 4.60% of it, which works out to about 8.01 mm^2 (source: https://en.wikichip.org/wiki/intel/microarchitectures/coffee_lake).

Now, since we have figures for density on both that node and Intel 7, I'm just going to apply the scale difference between those density figures, without trying to account for the mix of high/low density transistors, again under the assumption that it's going to be similar for both cores.

So, what we end up with is a Coffee Lake core having a nominal transistor count of 298.0 M and an Alder Lake P-core with a nominal transistor count of 758.7 M. In order to achieve equal perf/Tr, the i9-12900K would need to be 2.55x as fast as the i9-9900K, on single-threaded tasks. Let's see how that worked out, shall we?

Source: https://www.anandtech.com/show/1704...hybrid-performance-brings-hybrid-complexity/7
​

Whoops! Looks like Golden Cove only managed a 1.368x speedup on integer and 1.469x speedup on floating point. By your definition, Golden Cove / Alder Lake was an abject failure!

So, do you now see why I find this metric of perf/transistor so absurd? It tells us something that's both irrelevant and that we (should) already know!

 

[TheHerald]

So, here's how ridiculous it is to compare performance per transistor and why I utterly reject that metric. According to the graph I posted above, Intel 14 nm++ has a density of 37.22 MTr/mm^2. Coffee Lake was made on that node. The i9-9900K die is 174 mm^2 and a single core is about 4.60% of it, which works out to about 8.01 mm^2 (source: https://en.wikichip.org/wiki/intel/microarchitectures/coffee_lake).

Now, since we have figures for density on both that node and Intel 7, I'm just going to apply the scale difference between those density figures, without trying to account for the mix of high/low density transistors, again under the assumption that it's going to be similar for both cores.

So, what we end up with is a Coffee Lake core having a nominal transistor count of 298.0 M and an Alder Lake P-core with a nominal transistor count of 758.7 M. In order to achieve equal perf/Tr, the i9-12900K would need to be 2.55x as fast as the i9-9900K, on single-threaded tasks. Let's see how that worked out, shall we?

​

​

Source: https://www.anandtech.com/show/1704...hybrid-performance-brings-hybrid-complexity/7​

​

Whoops! Looks like Golden Cove only managed a 1.368x speedup on integer and 1.469x speedup on floating point. By your definition, Golden Cove / Alder Lake was an abject failure!

So, do you now see why I find this metric of perf/transistor so absurd? It tells us something that's both irrelevant and that we (should) already know!

Of course it is. I've already said that a couple of times. Intel's pcores are considered HUGE for the given performance. That's why I used those for the comparison, I wanted to give the m3 every opportunity to shine. Yet it failed.

Saying you reject transistor comparisons is crazy. The whole reason ecores and zen5c cores exist is because of transistor to performance metrics. Literally the only reason the 7995wx is faster than the 7600x are the transistors. It has more transistors and therefore more cores = more performance.

 

[bit_user]

Of course it is. I've already said that a couple of times. Intel's pcores are considered HUGE for the given performance. That's why I used those for the comparison, I wanted to give the m3 every opportunity to shine. Yet it failed.

They really didn't. They achieved ST performance parity at much lower clocks (peak: 4.06 GHz) than Alder Lake, where the i9-12900K boosts up to 5.2 GHz. So, roughly speaking, that's about 28% better IPC, which is absolutely huge! Lion Cove is a whopping 81.9% bigger than the M3's P-core, and it only delivered gen-on-gen IPC gains of 14% over Redwood Cove, which is within a couple % of Raptor Cove.

That's how you deliver strong performance at low power: through high IPC. That takes a lot of transistors. So, it's really not surprising the Apple P-cores use a lot. What's surprising is how much smaller and faster they are than Lion Cove! Right? Just 55% as big, and yet they're still 16.5% faster on CB (ST), while using less power!

On the other hand, if you're someone like Intel or AMD, you'd like to make smaller cores and clock them higher, because that's how you deliver better perf/$. Because Apple is vertically-integrated and sells only relatively low core-count CPUs, it can actually afford to make bigger cores, although I think it's telling that they're still much smaller than Lion Cove!

Saying you reject transistor comparisons is crazy.

What's crazy is to focus on perf/Tr across very different nodes and design targets. That's silly. If you disagree, explain to me why an end-user would care about that!

The whole reason ecores and zen5c cores exist is because of transistor to performance metrics.

Now, if you want to talk about AMD's C cores (and probably this applies to Intel's E-cores, to some extent), they're smaller in part because they could use finer pitch structures, if they don't have to clock as high. You can turn around and use that additional density and lower frequency target to deliver more IPC. So, it's a different design philosophy than trying to build the highest-clocking cores.

 

[TheHerald]

They really didn't. They achieved ST performance parity at much lower clocks (peak: 4.06 GHz) than Alder Lake, where the i9-12900K boosts up to 5.2 GHz. So, roughly speaking, that's about 28% better IPC, which is absolutely huge! Lion Cove is a whopping 81.9% bigger than the M3's P-core, and it only delivered gen-on-gen IPC gains of 14% over Redwood Cove, which is within a couple % of Raptor Cove.

That's how you deliver strong performance at low power: through high IPC. That takes a lot of transistors. So, it's really not surprising the Apple P-cores use a lot. What's surprising is how much smaller and faster they are than Lion Cove! Right? Just 55% as big, and yet they're still 16.5% faster on CB (ST), while using less power!

Man, that's exactly (exactly exactly) what I said in my first post and you said I'm coping. I literally said M3 delivers good efficiency because of low power and a big core.

Now you are saying the same thing, that it takes a lot of transistors. Yes, it does, but it's absolutely useless for desktops and only useful for mobile. MT is much more important in the desktop desgment, that's why they are trying to pack more cores instead of just enlarging the core to achieve better ST efficiency. The whole reason intel decided to add ecores is to not sacrifice MT performance due to their design needing more transistors for increased IPC.

The bottomlone is M3 isn't particularly impressive. It sacrifices a ton of MT performance and efficiency to achieve good ST performance and efficiency. I don't see any obvious benefit to arm over x86.