Discussion Intel current and future Lakes & Rapids thread

[eek2121]

Those two points are entirely unrelated. PL2 being set to 228W has nothing to do with 5.3GHz ST boosts.

Did anyone say ST boosts for Alder Lake? Intel needs the extra headroom to push clocks up. ST boosts don't matter because the chip isn't going to use that much power. Multicore boosts, however...

 

[IntelUser2000]

5.3GHz isn't good news. "Up to" 20% faster and at same clocks? That's what you get from a huge ass core? I hoped for a 5GHz one and being "up to 20% faster".

He's also guessing based on ES2, but unless you know the behavior to a tee, how would you know how the QS performs?

Also I don't know how they expect scores to pretty much double unless Gracemont performs better than Sunny Cove. It lacks Hyperthreading remember? What about the lower clocks? I can believe it for mobile but no way for desktop.

Verdict: Fan fiction

 

[jpiniero]

Also I don't know how they expect scores to pretty much double unless Gracemont performs better than Sunny Cove. It lacks Hyperthreading remember? What about the lower clocks? I can believe it for mobile but no way for desktop.

Maybe not Sunny Cove but it would need to be better than Skylake for sure.

 

[uzzi38]

Did anyone say ST boosts for Alder Lake? Intel needs the extra headroom to push clocks up. ST boosts don't matter because the chip isn't going to use that much power. Multicore boosts, however...

Are you insinuating 5.3GHz all-core boosts for ADL-S?

I feel like that claim is going to end just like your previous claim that the Gracemont cores are going to clock to 5GHz. Which by the way, this supposed "leak" disputes.

Not that I believe it's accurate anyway.

 

[lobz]

ehhhh, that Alderlake QS leak's source seems not that reliable according to some people responsed.....
(if you ask me, my impression about that forum is there were so many 'Zen2 OC@4.8Ghz' rumor floating around there when it was 2018 and me as a victim was also cheated by them)

To be fair, Intel 5+ GHz leaks almost always came true. They've really learned on 14nm how to feed the speed demon
One of the most moronic Intel leaks floating around here, as far I can remember was Skylake with its 20% IPC uplift and reverse multi-threading... still shaking my head with my face in my palms.

 

[Saylick]

To be fair, Intel 5+ GHz leaks almost always came true. They've really learned on 14nm how to feed the speed demon
One of the most moronic Intel leaks floating around here, as far I can remember was Skylake with its 20% IPC uplift and reverse multi-threading... still shaking my head with my face in my palms.

Inb4 rumors of Meteorlake utilizing new architecture that reverse multi-threads the big core and little core into a Big Chungus™ core.

 

[diediealldie]

Inb4 rumors of Meteorlake utilizing new architecture that reverse multi-threads the big core and little core into a Big Chungus™ core.

Hmm...any sources? Actually, modern superscalar processors are already working like a reverse-multithreaded cores. I don't see any benefits in having unified schedulers and decoders on top of current CPU cores which already use them.

 

[Thunder 57]

To be fair, Intel 5+ GHz leaks almost always came true. They've really learned on 14nm how to feed the speed demon
One of the most moronic Intel leaks floating around here, as far I can remember was Skylake with its 20% IPC uplift and reverse multi-threading... still shaking my head with my face in my palms.

Reverse hyperthreading, the rumor that never dies. I think we can add SMT4 to that list. Funny thing about reverse hyperthreading is that if you know even the basics of how a CPU works one would know that it is simply not possible.

 

[DrMrLordX]

I think we can add SMT4 to that list.

At least SMT4 exists/functions when attempted. Not that anyone in the x86 world is champing at the bit to implement SMT4.

 

[Hulk]

ADL rumors/scores are doubly confounding (or perhaps "quadruply") since it has two different cores and sets of frequencies. I'm waiting for a ST score for Golden Cove AND same for Gracemont with defined frequencies during the benchmark run before I can even begin to think about real world performance.

I don't understand why if someone has an ES part or whatever that they just don't release actual benchmark data instead of this cryptic BS? Yeah, yeah, I know, click bait.

 

[eek2121]

Are you insinuating 5.3GHz all-core boosts for ADL-S?

I feel like that claim is going to end just like your previous claim that the Gracemont cores are going to clock to 5GHz. Which by the way, this supposed "leak" disputes.

Not that I believe it's accurate anyway.

I'm not going down this rabbit hole again. We will wait for more leaks.

 

[Saylick]

At least SMT4 exists/functions when attempted. Not that anyone in the x86 world is champing at the bit to implement SMT4.

Not unless your name rhymes with Itchy Itch.

Also, kudos for using the OG spelling for "chomping" at the bit.

 

[DrMrLordX]

Not unless your name rhymes with Itchy Itch.

. . . yeah.

Also, kudos for using the OG spelling for "chomping" at the bit.

 

[Thunder 57]

Not unless your name rhymes with Itchy Itch.

Also, kudos for using the OG spelling for "chomping" at the bit.

I was totally drawing a blank on who that was. Wonder whatever happened to that special fellow.

 

[LightningZ71]

I haven't seen him active on the other forum he used to inhabit either for a long while. While I wasn't a fan, I do hope that they are OK

 

[Exist50]

SMT4 for x86 is going to happen eventually. It's just a question of when. I'm guessing both AMD and Intel will have it by the end of '25. Reverse-SMT, meanwhile, is most likely a pipe dream.

 

[Cardyak]

It's not that reverse hyperthreading is impossible, it's just not been pragmatic from an engineering standpoint.

It's not just about how much performance you can extract using a given solution, but how much perf/area, perf/xtors, and perf/W.

Reverse hyperthreading and Out of Order execution both achieve the same goal of extracting ILP to increase performance, but Out of Order execution has always been a lot more efficient in terms of the number of transistors used. There's no guarantee this will continue forever, as future blocks of the CPU are scaled to larger sizes we may get to a point where Reverse HT becomes cheaper than OoO, but that has not happened yet, hence why it has never been implemented in a product so far.

 

[moinmoin]

Interestingly enough IBM had a patent on reverse simultaneous multi-threading filed in 2009 and granted in 2013 that they let lapse in 2017 due to not paying the necessary fee. Kinda says a lot how feasible and how much worth that approach must be.

 

[mikk]

If Raichu can be believed the scores are real but it depends on the clock speed, he posted a formula. ADL-S 8+8 is able to achieve ~11500 points with 5.0 Ghz+3.9 Ghz according to this formula but likely needs to run at 228W PL2. But even with a more appropriate MT clock speed at 4.5 Ghz+3.5 Ghz it's close to Ryzen 9 5950X performance in Cinebench MT.

 

[eek2121]

If Raichu can be believed the scores are real but it depends on the clock speed, he posted a formula. ADL-S 8+8 is able to achieve ~11500 points with 5.0 Ghz+3.9 Ghz according to this formula but likely needs to run at 228W PL2. But even with a more appropriate MT clock speed at 4.5 Ghz+3.5 Ghz it's close to Ryzen 9 5950X performance in Cinebench MT.

It won’t do any good to point this out. Too many AMD redacted.

It doesn’t matter that gracemont has a much faster AVX implementation than skylake (CBr20 uses AVX). It doesn’t matter that the golden cove cores will be 15%-20% faster than rocket lake or tiger lake. It doesn’t even matter that on paper, you can do basic math to see that Intel’s parts will be faster. Somehow AMD’s 8 month old Zen 3 architecture will be faster, even if it isn’t faster in benchmarks. Many of these guys will overlook the fact that Intel will be the first to launch a laptop chip with more than 8 cores, the first to launch a DDR5 platform, and the first to launch a true performance hybrid part.

The rest of us will continue to discuss the strengths and weaknesses of both companies and their products, speculate on future products, and have a good time in general. Some of us feel the need to defend our purchases. The rest of us (including me) just love technology, regardless of the company.





You cannot use the word ***boy on the tech forum.


esquared
Anandtech Forum Director

 

[Exist50]

If Raichu can be believed the scores are real but it depends on the clock speed, he posted a formula. ADL-S 8+8 is able to achieve ~11500 points with 5.0 Ghz+3.9 Ghz according to this formula but likely needs to run at 228W PL2. But even with a more appropriate MT clock speed at 4.5 Ghz+3.5 Ghz it's close to Ryzen 9 5950X performance in Cinebench MT.

I doubt a 228W PL2 would be sufficient for all core 5.0/3.9GHz.

 

[eek2121]

I doubt a 228W PL2 would be sufficient for all core 5.0/3.9GHz.

Remember this is 10nm, and without AVX512 features.

EDIT: The 11900k uses less than 250W for 8 cores loaded at 4.7Ghz with an AVX2 workload, for reference. Also see the chart below for the 11980hk. I suspect Tiger lake could see an all core boost of 4.7-5.0 ghz with sufficient cooling and a PL2 of just 125W. This will depend on workload of course, and everyone expects golden cove to run hotter…

 

[Gideon]

Here is the post itself, if people missed it (didn't see it directly linked):

https://twitter.com/x/status/1415332985772023811

What stood out the most to me, is how strong the small cores have to be for these scores.

I was sceptical about Alder lake beating 5950X for a very specific reason:

• Generally SMT gives you ~30% extra performance per core in cinebench no Zen 2
• Hyperthreading for Skylake is roughly in the same ballpark.Can't see apples-to-apples R20 numbers but it's 31% in R15 for instance

But according to Raichu's formula Gracemont cores get:

115 * 0.95= 109.25 points per GHz in R20

A 9700K from Anandtech bench gets 3633 points, as it's running at around it's all-core turbo of 4.6 ghz, it gets:

3633 / 8 / 4.6 = 98.7 points per GHz

But being so high clocked reduces the per ghz number somewhat let's pick a lower-clocked example:
A 6600K from bench (running at an "impressive" +100 Mhz all-core turbo) gets 1446 points

1446 / 4 / 3.6 = 100.41 points per GHz

So unless those 8 small cores are running really low frequencies (which according to Raichu it's not, it's actually slightly higher than 6600K) they at least 10% faster than Skylake per clock.

Similar comparisons can be done against Skylake and Rocket Lake:

Spoiler: Skylake with SMT vs Golden Cove with SMT

10900K is 6319 / 10 / 4.9 = 129 points per Ghz
11900K is 5900 / 8 / 4.7 = 155 points per Ghz
Golden Cove is (according to Raichu)
162*1.275 = 206.55 points per Ghz

TL;DR (if the info is correct)

Alder Lake Gracemont (small cores) have 10% better IPC than Skylake, running nearly top clocks Sunny Cove achieved.
Alder lake Golden Cove (big cores) have 60% better IPC than Skylake (and 32% better IPC than Rocket Lake) running at up to 5.3 Ghz.

At least in Cinebench.

P.S.
Bear in mind 5950X is still limited by it's rather tame TDP compared to intel, pushed to 250+W it can also score 11000 - 12000 points. A 6nm refresh should be in the same ballpark. If all this si accurate, this should bode very well for Raptor Lake

 

[lobz]

It won’t do any good to point this out. Too many AMD fanboys.

You've just rendered everything else you'd said in your post totally meaningless.

 

[vstar]

According to this, the 12900K has an all core MT boost of 5 GHz and a ST boost of 5.3 GHz for the big cores. For the little cores, ST tops out at 3.9 GHz and MT is at 3.7 GHz.

https://twitter.com/x/status/1415459621628571650