Question Alder Lake - Official Thread

[Hulk]

With the release of Alder Lake less than a week away and the "Lakes" thread having turned into a nightmare to navigate I thought it might be a good time to start a discussion thread solely for Alder Lake.

 

[coercitiv]

I'm sure there will be plenty of 4+8 laptops that will at least come with a dGPU.

I would expect 2+8 with dGPU. The scheduler will either be updated to scale properly or branched out to behave differently on mobile vs. desktop.

 

[Heartbreaker]

I would expect 2+8 with dGPU. The scheduler will either be updated or branched out to behave differently on mobile vs. desktop.

I expect not. 2+8 will be aimed only at ULTRA LOW power, for tablets and convertibles like the Surface Pro. Those never have dGPU.

 

[coercitiv]

There are no 4+8 Alder Lake CPUs at all. None even rumored.

I expect not. 2+8 will be aimed only at ULTRA LOW power, for tablets and convertibles like the Surface Pro. Those never have dGPU.

What exactly do you think Intel will build their Alder Lake mobile lineup on?! 2+8 for 9W and bellow and then 6+8 SKUs for anything above 10W?! Before writing your posts did you at least stop to consider that a 2+8 chip can easily scale to 20W+ or that 4+8 chips are clearly on the table when segmenting your lineup based on 6+8 silicon?

Here's a rumor for you. Not necessarily true, but at least makes a common sense effort to envision a proper line-up. It dates back to March this year.

 

[tomatosummit]

There are no 4+8 Alder Lake CPUs at all. None even rumored.

Not counting cutdowns?
If this image is still correct then mobile cpus keep 8e cores until there are only 2p cores left.
I'd not noticed the H55 segment has the full 8+8 cpu until now.

 

[jpiniero]

Not counting cutdowns?
If this image is still correct then mobile cpus keep 8e cores until there are only 2p cores left.
I'd not noticed the H55 segment has the full 8+8 cpu until now.

Yep. It is true that the SKU list could change before release but I would expect plenty of 4+8 models out there.

 

[dullard]

Edit: Misread the post that I was commenting on.

 

[Markfw]

You recently chastised someone for making a hyperbolic statement regarding 12900K performance vs. 5950X. In all fairness, do you think your statement is a little over the top? And to quote you, "do you own one?"

Games need certain threads to stay on the P cores and the scheduler in rare cases isn't doing this correctly. It's a pretty simple press of the scroll lock to disable the E's in those rare cases.

I have a 12700K and don't game. I have ZERO issues. In fact, this is the most stable system I've ever built.

Besides the game issue I mentioned, which I don't think qualifies as "serious issues," what other issues are you referring to?

I was commenting on what the others have observed. A lot of this "blame" I attribute to MS, as its their OS that is mostly the problem. I have actually thought of getting a 12700k, but I would rather spend the money on another EPYC due to the efficiency.

 

[Heartbreaker]

Here's a rumor for you. Not necessarily true, but at least makes a common sense effort to envision a proper line-up. It dates back to March this year.

Ok, you got me. Such rumors exist.

What exactly do you think Intel will build their Alder Lake mobile lineup on?! 2+8 for 9W and bellow and then 6+8 SKUs for anything above 10W?! Before writing your posts did you at least stop to consider that a 2+8 chip can easily scale to 20W+ or that 4+8 chips are clearly on the table when segmenting your lineup based on 6+8 silicon?

Even at 20W it's extremely unlikely anyone will be pairing a dGPU with the 2+8 part.

Even if there are 4+8 parts, that get paired with dGPU, they won't be top end dGPU parts.

Certainly nothing like the RX 6900 XT from that one cherry picked slide you pulled out of context, to fuel your FUD.

 

[Zucker2k]

These CPUs have serious issues IMO.

imagine them having such "serious issues" and still performing this well against the competition. ADL has the potential to perform even better! Best chips on desktop at the moment, yet people can't stop hating on them.

 

[nicalandia]

Anybody know where this chart came from?
Post #334

Guru3D, they have determined that for IPC is best to test CineBench R15 instead

Intel Core i9-11900K processor review (Page 7)

We review Intel's new Desktop flagship processor, the 8-core Core i9-11900K. The 11th generation Intel Core desktop processors previously known as RKL aka Rocket Lake-S has been released in an effort to take on the battle with AMD.

www.guru3d.com

 

[Hulk]

Guru3D, they have determined that for IPC is best to test CineBench R15 instead

Intel Core i9-11900K processor review (Page 7)

We review Intel's new Desktop flagship processor, the 8-core Core i9-11900K. The 11th generation Intel Core desktop processors previously known as RKL aka Rocket Lake-S has been released in an effort to take on the battle with AMD.

www.guru3d.com

I'm not so sure that is what they are implying.

"According to the developers, the software has been "extensively developed to exploit the performance of new hardware as possible." The results are unsurprisingly not comparable with those from earlier versions."

I believe they are speaking more generally to Cinebench and not the specific version. Since the differences in IPC are magnified as you move toward later versions I think it reasonable to believe that the instruction parallelism has increased. Wider, smarter (OoS better) are better able to keep the pipes running and produce better scores and are a more accurate representation of CPU performance in other applications.

I think the reality is that they are using CB R15 because those are the test results they have on hand and don't want to retest all of the CPU's on CB R23 @3500MHz. It's a lot of work.

Golden Cove is 19% better than Zen 3 in Cinebench R23. That's a significant difference they skip over with this older test.

Cinebench R23		ST	ST	Points		MT	MT	Point per	Points
	# Cores	Frequency	Score	per GHz		Frequency	Score	Core	per GHz
12700K	12	4.700					22477
12700K - P Cores Only	8	4.700	1854	394		4.700	18843	2355	501	27.0%
12700K - E Cores Only	4	3.800	1027	270		3.600	3892	973	270	0.0%
11900K	8	5.100	1678	329		5.000	15978	1997	399
11700K	8	4.600	1517	330		4.600	14320	1790	389			19%	Zen 3 to Golden Cove
11600K	6	4.600	1537	334		4.600	10822	1804	392
				331					394	18.9%
												27%	Rocket to Alder Lake
5600X	6	4.600	1537	334		4.300	10812	1802	419			23%	Skylake to Rocket
5800X	8	4.700	1594	339		4.550	15310	1914	421			8%	Haswell to Skylake
5900X	12	4.800	1590	331		4.250	21367	1781	419
5950X	16	4.900	1669	341		3.900	26291	1643	421			57%	Skylake to Alder
				336					420	24.9%		69%	Haswell to Alder Lake
10900KF	10	4.900	1317	269		5.000	15653	1565	313
10700KF	8	4.800	1286	268		5.000	12818	1602	320
10600K	6	4.700	1267	270		4.800	9036	1506	314
	4	3.400	891	262		2.528	3314	829	328
				267					319	19.3%

 

[jpiniero]

Even if there are 4+8 parts, that get paired with dGPU, they won't be top end dGPU parts.

If the H45 i5 ends up being 4+8 it won't be that niche. You wouldn't get a top end dGPU part, more like midrange.

 

[Heartbreaker]

If the H45 i5 ends up being 4+8 it won't be that niche. You wouldn't get a top end dGPU part, more like midrange.

And that will be mid-range mobile, so you will end up GPU limited 99%+ of the time.

It's just manufactured FUD to look at one, out of context case , combining top end desktop GPU, with disabling some performance cores, to proclaim the design is a mess.

 

[Markfw]

imagine them having such "serious issues" and still performing this well against the competition. ADL has the potential to perform even better! Best chips on desktop at the moment, yet people can't stop hating on them.

While I admit they are doing pretty good, the 12900k uses WAY too much power, and you have to tune the others down to get close to Ryzen in efficiency.

And then there is windows.....

 

[IntelUser2000]

While I still anticipate the 2p+8e core cpus I'm not sold on desktop e-cores. They feel like marketing gimmick today.
Give me 2-4 for low power draw in idle and tertiary tasks or a huge 32+ array for specialised workstation tasks.
I think it's telling how apple's m1++ is laid out. 8cores for performance and 2 efficient for extended battery life that's letting them scale across performance laptops, DTR and mid range workstations (big imacs).

Yes "gimmicks" are what allowed even the 12600K to kick the crap out of the previous gen 11900K flagship, and even in MT allow Intel to be competitive not just in performance but performance per price with Zen 3.

Remember Lakefield? Where the whole Sunny Cove + Tremont setup at 7W ended up being slower than the 6W Tremont-only Jasper Lake setups in MT while being similar in ST?

We went from that to coming close to 5950X. No, I don't see that as a gimmick at all. People are just grasping at straws. This is the future, where even in general purpose CPUs cores are optimized for tasks.

 

[dullard]

And that will be mid-range mobile, so you will end up GPU limited 99%+ of the time.

Just to put context into your discussion:

Midrange 45W i5 gaming laptops are usually configurable with NVidia RTX 3050, 3050 Ti, or 3060 GPUs.

Dell XPS 15 Laptop | Dell USA

www.dell.com

| Dell USA

www.dell.com

Computers, Monitors & Technology Solutions | Dell USA

Dell provides technology solutions, services & support. Buy Laptops, Touch Screen PCs, Desktops, Servers, Storage, Monitors, Gaming & Accessories

www.dell.com

Higher end 45W gaming laptops, usually have i7s, not i5s. They tend to be configurable with NVidia RTX 3070 or 3080 GPUs.

Computers, Monitors & Technology Solutions | Dell USA

Dell provides technology solutions, services & support. Buy Laptops, Touch Screen PCs, Desktops, Servers, Storage, Monitors, Gaming & Accessories

www.dell.com

Computers, Monitors & Technology Solutions | Dell USA

Dell provides technology solutions, services & support. Buy Laptops, Touch Screen PCs, Desktops, Servers, Storage, Monitors, Gaming & Accessories

www.dell.com

I agree, it does seem silly to put a 4 P core mobile chip with the highest end video card. But, you really don't need the highest end video card to do mobile gaming. You just need to turn settings down a couple notches.

Also, everyone should keep in mind, this discussion is about just 1 single game before patches have been released. I wouldn't consider this to be the truth going forwards once the games are designed with E cores in mind.

 

[nicalandia]

I think the reality is that they are using CB R15 because those are the test results they have on hand and don't want to retest all of the CPU's on CB R23 @3500MHz. It's a lot of work.

Golden Cove is 19% better than Zen 3 in Cinebench R23. That's a significant difference they skip over with this older test.

Cinebench R23 is not really a good IPC measurement for older CPUs that don't have newer instructions, CB R23 takes advantage of Intel AVX. Cinebench R15 does not use AVX so it's more of a true IPC test


You can check as to why CB R15 is better for IPC here:

Analyzing Zen 2’s Cinebench R15 Lead

Cinebench R15 (CBR15) is a popular benchmark based on Cinema4D’s 3D rendering engine. It can utilize all available CPU threads, but here we’ll be analyzing it in single thread mode. In …

chipsandcheese.com

"Correlation between IPC and branch prediction accuracy. There’s a clear positive correlation here, for both CPUs"

 

[dullard]

This is the future, where even in general purpose CPUs cores are optimized for tasks.

You are correct that it isn't a gimmick. But, even Intel isn't really banking on 8 E cores for very many of their desktop chips. Only the i9 line has 8 E cores. The large majority of Intel's desktop Alder Lake chip versions will have 4 E cores or less. The 12600K that you just praised for beating the 11900K doesn't even have 8 E cores.

At the state of software now (Windows 11 in early stages and almost no software has been optimized for E cores yet), I tend to agree with @tomatosummit. 4 E cores seems to be the sweet spot right now and 16+ E cores will be the sweet spot in the future. 8 E cores on the desktop just isn't quite optimum.

 

[IntelUser2000]

Remember the 6700K has 8MB L3 cache, while the 12600K has 20MB and 12900K has 30MB.

I don't know why they didn't have the 8E config test on Techspot. That's an oversight.

Better yet, 6700K should have been tested with HT.

You can see from 9100F vs 10100 in TPU review that in games like BFV, Doom Eternal, Cyberpunk 2077, RDR2, SoTR, the presence of Hyperthreading affects the results quite drastically.

The interconnect latencies are still high on the E cores but having thread count equalized will significantly reduce the gap between the two.

You are correct that it isn't a gimmick. But, even Intel isn't really banking on 8 E cores for very many of their desktop chips. Only the i9 line has 8 E cores. The large majority of Intel's desktop Alder Lake chip versions will have 4 E cores or less. The 12600K that you just praised for beating the 11900K doesn't even have 8 E cores.

Yes, it's pretty clear Alderlake was meant to be on mobile, seeing as how their efficiency is run out of the window on the 12900K.

It's just a matter of segmentation. Intel knows that power efficiency, performance are all knobs for a premium product. That's why the Core parts have all the power management settings while Celeron and the Pentium parts have the power states gimped. Do you think the Celerons and Pentiums had the C-states cut because they thought it wasn't "worth it"? No it's segmentation simple as that.

The chips with E cores are significantly faster in MT than those without anyways.

Like I said, the ideal hybrid is being able to absolutely maximize the size of the low thread count core and have MT tasks run by the smaller cores. The Big.Bigger is better because Big.Little has the issue where the gap is too big between the two and in real world and in Windows setups with countless applications that will suffer from substantial context switching performance issues. I bet the hybrid is going to be a huge contributor in catching up with ARM SoCs as well.

Big changes don't happen overnight. Hyperthreading was extremely poor on the Xeons, but improved with the Pentium 4 versions that came later and was perfected on Nehalem.

 

[Hulk]

Cinebench R23 is not really a good IPC measurement for older CPUs that don't have newer instructions, CB R23 takes advantage of Intel AVX. Cinebench R15 does not use AVX so it's more of a true IPC test


You can check as to why CB R15 is better for IPC here:

Analyzing Zen 2’s Cinebench R15 Lead

Cinebench R15 (CBR15) is a popular benchmark based on Cinema4D’s 3D rendering engine. It can utilize all available CPU threads, but here we’ll be analyzing it in single thread mode. In …

chipsandcheese.com

That is a valid argument.

But I will counter argue CB R23 is a more valid representation of ADL performance because ADL performance in CB R23 is more representative of how it performs in other applications than R15. Predicting performance in other applications is the primary purpose of benchmarks and that is why I would argue R23 is a better predictor than R15.

 

[LightningZ71]

Why would we want to exclude AVX2 from modern performance benchmarking? Using older programs with limited instruction sets is a purely academic test of what the CPU's performance would have been when that benchmark was initially released. We don't live in the world of 6 years ago. We live in 2021, using largely 2021 software. Yes, there are certainly some organizations that use very outdated software (I still support one system that was written in COBOL many, MANY years ago), but those organizations know exactly what they are looking for. Arguing that X processor is Y% better in IPC on what was a typical instruction set from 6+ years ago just isn't that relevant. How does it run what most of us are using TODAY?

It largely doesn't matter much in practice. Even most 2 core, 4 thread Ice Lake/ Tiger Lake laptops are plenty fast enough for 90% of population's system usage. About the only places where it seems to matter (outside of supercomputing, which uses a different class of CPU) is video encoding, video and graphic arts, and compiling.

Personally, I want benchmarks that are relevant to what is currently leading edge. If I'm dropping a lot of coin on a new computer, I want to know how it's going to perform on the stuff that's state of the art today, as that's what it's more likely to face going forward. That means AVX2, and, as it's the most modern thing out there at the moment, AVX-512 in it's various levels. I would want to know that Mobile Ice Lake, Tiger Lake, and Rocket Lake have crummy AVX-512 performance as compared to the server chips. It would indicate to me that I shouldn't waste my time with them, except to qualify code for operability (though, not suitability as the performance wouldn't be representative).

 

[nicalandia]

That is a valid argument.

But I will counter argue CB R23 is a more valid representation of ADL performance because ADL performance in CB R23 is more representative of how it performs in other applications than R15. Predicting performance in other applications is the primary purpose of benchmarks and that is why I would argue R23 is a better predictor than R15.

But ADL is NOT 19% ahead on IPC(Clock vs Clock) in real world performance.

 

[Hulk]

Yes but it's also not 1% ahead

 

[DrMrLordX]

Remember Lakefield?

Yeah that was a mess. Kinda makes you wonder why they continued with heterogeneous core configs in a flagship product after that particular disaster. The end result is certainly interesting, and gives the tech community much to ponder. It's really a shame that even after Lakefield that MS is struggling to get their scheduler right, though.

 

[nicalandia]

Yeah that was a mess. Kinda makes you wonder why they continued with heterogeneous core configs in a flagship product after that particular disaster. The end result is certainly interesting, and gives the tech community much to ponder. It's really a shame that even after Lakefield that MS is struggling to get their scheduler right, though.

Golden Cove>> Sunny Cove/Rocket Lake, they would never dare to release Lakefield type of performance on the desktop