Discussion Intel current and future Lakes & Rapids thread

[SiliconFly]

Just out of curiosity... comparing Apple M3 Max and Intel (previous) 13th gen mobile 13980HX geekbench scores.

M3 Max single thread is closer to Intel 13th gen mobile 13980HX. So, lets ignore it for a while.

But M3 Max MT is like 10% faster than 13980HX.

The crazy part is M3 Max is a 16C (12P+4E) 30W part I think & 13980HX is a 24C (8P+16E) 55W part with a PL2 of 157W!!!

What gives? Why so much discrepancy? I mean, a 30W apple part comfortably beating a intel part with PL2 of 157W. Is it even right to compare these two as they both belong to different price segments.

 

[Geddagod]

And you believe you can deduce ARL's Fmax based on it's TDP? Interesting!!!

Fmax doesn't matter, peak ST perf matters, and that's ~5%, from 2 different OEM slides

It's very interesting that you want to declare ARL a lost cause based on one quirky slide.

2, both look to be legit as well.

information properly vetted,

Both look pretty legit

Saying ARL is a lost cause is pretty extreme, but it's highly likely that ARL is a disappointment tbh

 

[Geddagod]

Considering all the hoopla around N3B, ARL-S might also be on 20A. After all, it's node agnostic and dual sourced.

Not the 8+16 die, maybe the 6+8

Thats what I said too

It's not too early for ARL perf leaks/speculation, since Intel themselves are speculating about ARL perf

Not even a single ES power usage benchmark this late in the cycle? Its definitely unlike intel.

There's that v/f curve Raichu leaked
But ye, mobile power draw is never leaked as early or as much as their desktop counterparts. MTL doesn't have desktop variants.

 

[igor_kavinski]

The crazy part is M3 Max is a 16C (12P+4E) 30W part I think & 13980HX is a 24C (8P+16E) 55W part with a PL2 of 157W!!!

I think this is part of the reason for Microsoft to make the effort for Windows on ARM. They got tired of the x86 players trying to make efficient enough SoCs.

 

[Viknet]

The crazy part is M3 Max is a 16C (12P+4E) 30W part I think

In reality it will be closer to 70-90W.
M2 Max with 8P+4E consumed 50-60W at max cpu-only load.

 

[SiliconFly]

In reality it will be closer to 70-90W.
M2 Max with 8P+4E consumed 50-60W at max cpu-only load.

Oh sorry. I was under the impression that M2 Max was capped at 30W or so and can't exceed the limit. My bad.

 

[SiliconFly]

I think this is part of the reason for Microsoft to make the effort for Windows on ARM. They got tired of the x86 players trying to make efficient enough SoCs.

MTL is gonna bring in some decent level of power efficiency. And ARL should hopefully improve it further. I don't want to speculate as there isn't good data yet about MTL & ARL power efficiency gains. But I think there's a decent chance ARL & Zen 5 might reach parity with Apple silicon in power efficiency next year.

 

[dullard]

MTL is gonna bring in some decent level of power efficiency. And ARL should hopefully improve it further. I don't want to speculate as there isn't good data yet about MTL & ARL power efficiency gains.

You could say there isn't good data on unreleased products. But Intel has provided their estimates several times for Meteor Lake (>20% power efficiency, not much performance increase):

https://www.techpowerup.com/review/intel-meteor-lake-technical-deep-dive/images/intel-4-03.jpg

https://www.techpowerup.com/review/intel-meteor-lake-technical-deep-dive/images/intel-4-13.jpg

 

[SiliconFly]

You could say there isn't good data on unreleased products. But Intel has provided their estimates several times for Meteor Lake (>20% power efficiency, not much performance increase):

https://www.techpowerup.com/review/intel-meteor-lake-technical-deep-dive/images/intel-4-03.jpg

https://www.techpowerup.com/review/intel-meteor-lake-technical-deep-dive/images/intel-4-13.jpg

If you see the two TPU links you posted, it talks about the power efficiency gains of Intel 4 over Intel 7 using Intel's version of industry standard core.

Even though MTL is on Intel 4, doesn't mean the power efficiency gains of MTL over RPL is the same as the power efficiency gains of Intel 4 over Intel 7. They're very different.

In fact, I'm expecting a lot more power efficiency gains due to the new MTL architecture.

Here's an excerpt from an older post for reference:
"...LP E-cores, DLVR, NOC, powering-down tCPU/tGPU tiles independently, an efficient SoC tile design that can run totally independent without powering up tCPU/tGPU (eg, while watching youtube), etc. Intel has claimed on multiple occasions that MTL SoC (not tCPU alone) can deliver up to 2X efficiency (probably under some normal workloads)..."

 

[dullard]

In fact, I'm expecting a lot more power efficiency gains due to the new MTL architecture.

Here's an excerpt from an older post for reference:
"...LP E-cores, DLVR, NOC, powering-down tCPU/tGPU tiles independently, an efficient SoC tile design that can run totally independent without powering up tCPU/tGPU (eg, while watching youtube), etc. Intel has claimed on multiple occasions that MTL SoC (not tCPU alone) can deliver up to 2X efficiency (probably under some normal workloads)..."

Running without powering up the CPU tile or without powering up the GPU tile is wildly different from the discussion at hand though. Which specifically was benchmarks running at full power. Yes, Meteor Lake will be drastically more efficient when it isn't fully utilized (at least for the specific tile that is powered off it is infinitely more efficient than an unused tile that is powered on). But in terms of the discussion of geekbench-like scores at PL2 of 157 W, you won't see the CPU tile powered off and will see much closer to the Intel 4 vs Intel 7 >20% power efficiency gains.

 

[mikk]

You think the separate voltage rail for E-cores won't improve efficiency? I think we should wait and see. The base clock speeds from the leaked models are closer to 30% higher. Afaik Raichu said the target was 1.5x efficiency improvement.

 

[Goop_reformed]

And you believe you can deduce ARL's Fmax based on it's TDP? Interesting!!!


Do you even believe in your own words? Cos' its exciting.

It's very interesting that you want to declare ARL a lost cause based on one quirky slide. But things work a bit differently here. Unless we have multiple credible sources and information properly vetted, no one here is gonna take it seriously.

Why are you so defensive. Take that slides at face value and move on. The only concrete infos right now are those slides. if you believe those aren't indicative of final performance then you must also accept the possibility of worse final performance for arrow lake.

I have also not declared arrow lake to be a lost cause nor infer that in anyway.

>But things work a bit differently here.

You're literally are a new member, just stop this. It's foolish.

 

[Joe NYC]

Wish some Robinhood hacker would penetrate into an OEM network and leak these

Or Gigabyte just being Gigabyte...

 

[DrMrLordX]

What gives? Why so much discrepancy? I mean, a 30W apple part comfortably beating a intel part with PL2 of 157W. Is it even right to compare these two as they both belong to different price segments.

Because it's Geekbench.

 

[S'renne]

Imo as long as efficiency improves its still...a bearable improvement on both mobile and desktop platforms, although this does make me want a Zen 5 CPU more instead but AMD's supplies aren't great for mobile high end

 

[uzzi38]

Igor thinks different.

Btw, the slide you keep referring to seems to have been cooked up by some Intel marketing guy who has no inkling of engineering. ARL-S is a 8+16+2 product, not a 8+16+1 as mentioned in that slide.

Of course the final product could perform differently, they're called projections for a reason. They're the targets Intel have set for themselves to hit effectively. They could exceed or fall behind those targets when it comes to the real deal, but thats something we'll only find out when retail silicon comes to market (or is close to doing so).

 

[uzzi38]

Imo as long as efficiency improves its still...a bearable improvement on both mobile and desktop platforms, although this does make me want a Zen 5 CPU more instead but AMD's supplies aren't great for mobile high end

Both MTL and ARL should be a decent efficiency step up thankfully, I'm not worried about that.

I really think the "problem" with ARL's relatively low perf gains is more due to Intel not being able to push clocks high on N3 rather than a problem with the core itself. Perhaps due to a lack of expertise on non-internal nodes, or perhaps due to N3 itself not being able to consistently hit higher clocks. Couldn't tell you. But best I can say is I wouldn't be surprised if we find ARL clocking a significant amount below RPL even on desktop (please let me clarify: this is an expectation, not a leak).

And that probably has knock on effects to MT perf too, because while they might be able to reign in power considerably, they might not be able to achieve higher MT clocks. So having a PL2 above (what was it, 177w?) might be a result of that as well.

 

[eek2121]

Just out of curiosity... comparing Apple M3 Max and Intel (previous) 13th gen mobile 13980HX geekbench scores.

M3 Max single thread is closer to Intel 13th gen mobile 13980HX. So, lets ignore it for a while.

But M3 Max MT is like 10% faster than 13980HX.

The crazy part is M3 Max is a 16C (12P+4E) 30W part I think & 13980HX is a 24C (8P+16E) 55W part with a PL2 of 157W!!!

What gives? Why so much discrepancy? I mean, a 30W apple part comfortably beating a intel part with PL2 of 157W. Is it even right to compare these two as they both belong to different price segments.

You are comparing N3 to something similar to N7. Compare an N7 chip from Apple and things won’t look so impressive. Or wait for Meteor Lake or Arrow Lake to drop and compare those.

 

[SiliconFly]

Running without powering up the CPU tile or without powering up the GPU tile is wildly different from the discussion at hand though. Which specifically was benchmarks running at full power. Yes, Meteor Lake will be drastically more efficient when it isn't fully utilized (at least for the specific tile that is powered off it is infinitely more efficient than an unused tile that is powered on). But in terms of the discussion of geekbench-like scores at PL2 of 157 W, you won't see the CPU tile powered off and will see much closer to the Intel 4 vs Intel 7 >20% power efficiency gains.

I'm talking about real world power efficiency & not "benchmarks running at full power".

 

[SiliconFly]

Why are you so defensive. Take that slides at face value and move on. The only concrete infos right now are those slides. if you believe those aren't indicative of final performance then you must also accept the possibility of worse final performance for arrow lake.

I have also not declared arrow lake to be a lost cause nor infer that in anyway.

>But things work a bit differently here.

You're literally are a new member, just stop this. It's foolish.

I'm not new to AT if thats what you mean. And it's not foolish. You keep obsessing over that slide. But I'm moving on.

 

[dullard]

Just out of curiosity... comparing Apple M3 Max and Intel (previous) 13th gen mobile 13980HX geekbench scores.

M3 Max single thread is closer to Intel 13th gen mobile 13980HX. So, lets ignore it for a while.

But M3 Max MT is like 10% faster than 13980HX.

The crazy part is M3 Max is a 16C (12P+4E) 30W part I think & 13980HX is a 24C (8P+16E) 55W part with a PL2 of 157W!!!

I'm talking about real world power efficiency & not "benchmarks running at full power".

Are there two of you using this account? First you mention Geekbench scores with max single thread, max multithread, and high PL2 values. Then you claim you aren't talking about benchmarks at full power. Which one of the two SiliconFly personalities am I responding to now?

 

[SiliconFly]

Are there two of you using this account? First you mention Geekbench scores with max single thread, max multithread, and high PL2 values. Then you claim you aren't talking about benchmarks at full power. Which one of the two SiliconFly personalities am I responding to now?

The first one is about 13980HX vs M3 Max. The second one is about MTL's real world power efficiency.

 

[trivik12]

I am skeptical M3 Max drawing just 30w for CPU intensive applications. Let us wait for independent benchmarks. m2 max avg power draw was 34w.

Apple MacBook Pro 16 Review | 2023 M2 Max Performance

Hands-on M2 Max MacBook Pro 16-inch review. Performance testing and benchmarks for Apple's 2023 laptop computer.

appleinsider.com

This time around there are 4 more p-cores and its clocked higher. It will draw more power for sure. Or they will throttle the cpu.

 

[msj10]

I really think the "problem" with ARL's relatively low perf gains is more due to Intel not being able to push clocks high on N3 rather than a problem with the core itself. Perhaps due to a lack of expertise on non-internal nodes, or perhaps due to N3 itself not being able to consistently hit higher clocks. Couldn't tell you. But best I can say is I wouldn't be surprised if we find ARL clocking a significant amount below RPL even on desktop (please let me clarify: this is an expectation, not a leak).

Apple just gained 500MHz on the M3 using N3B. I am getting more skeptical of these major clock regression rumors of ARL. the poor ST performance might just be because LNC IPC uplift is disappointing or because these projections aren't final.

 

[DrMrLordX]

Apple just gained 500MHz on the M3 using N3B. I am getting more skeptical of these major clock regression rumors of ARL. the poor ST performance might just be because LNC IPC uplift is disappointing or because these projections aren't final.

Sorry to repeat myself (for those that have read what I've written elsewhere), but it's been reported multiple times that Intel won't be taking any N3B wafers at all. Arrow Lake compute tiles are likely to be on 20a only.