Intel Core 12th Gen Alder Lake Preview

[Minus Infinity]

LOL 1080p gaming benchmarks. Yeah I buy 12900K for $600 to go with my 3070/3080 to play 1080p on my 4K monitor. Oh right e-sports.

Let's wait for actual independent meaningful reviews before wetting our pants. No doubt AL is huge improvement and AMD should be worried, but actual real world user experience won't be much different overall even against Zen 3. At 1440p let alone 4K there'll be bugger all difference for gamers unless you maybe have a 3090/6900

 

[RandallFlagg]

There is a reason to read multiple sites, Anand gives worst case, GN gives best, and TPU tends to give about avg.


If intel wants any credibility they have to fire Ryan Shrout... benchmarking on a build they know hurts AMD, giving them lower than recommended memory speeds.
Claiming a win when they are barely on par, and releasing the slide deck when AMD is showing off its Quarterly performance.

I'm not sure what all the indignation over that is, except maybe faux. The impact is being greatly exaggerated in the aggregate, the bug was not understood to be a problem with Win 11 at the time the benchmarks were done, and Intel is not responsible to handle MS/AMD bugs. To wit, Intel Lakefield was a bust largely due to the scheduler in Win 10 not effectively using its E-cores. That's been going on forever and I don't see people whining about it.

To Illustrate, this was the geomean difference at 1080P with the buggy Win 11 vs Win 10. Looks like about 1-2% to me, varying a bit by title :

On RAM - Given that he could have put DDR5-6000 in the Intel rig, that's not quite fair. As it is he used DDR4-3200 CL14-14-14-34 on AMD which is pretty decent RAM, the Gen 12 Intel box had DDR5-4400 which is among the slowest memory speeds for DDR5 and not even the maximum JEDEC standard. These memory speeds do not in any way look like an unfair comparison to me though. It's probably just selected to be JEDEC speed with XMP timings, which is a very middle-of-the-road approach.

I am hoping reviewers take a close look at DDR4 and DDR5 performance scaling on AL.

 

[Halo3Addict]

I'm not sure I understand the need for E cores on a desktop. Laptops I understand, it's all about efficiency. But desktops just want raw power.

Looking forward to reviews

 

[Chaitanya]

Looking at TDP values no wonder the size of VRM cooling assemblies have grown even on M-ITX offerings.

 

[Makaveli]

If intel wants any credibility they have to fire Ryan Shrout... benchmarking on a build they know hurts AMD, giving them lower than recommended memory speeds.

lol they won't fire him that is why he got hired.

 

[WhoDecidedThat]

If Intel is telling the truth (which is very doubtful) - Gracemont is only 20% behind Golden Cove?!! Damn, 8x Gracemont is looking like a very impressive processor for budget laptops. Hell, even budget gaming laptops with RTX 3050/3060s will be well suited with 8x Gracemont cores.

 

[Crackong]

If Intel is telling the truth (which is very doubtful) - Gracemont is only 20% behind Golden Cove?!! Damn, 8x Gracemont is looking like a very impressive processor for budget laptops. Hell, even budget gaming laptops with RTX 3050/3060s will be well suited with 8x Gracemont cores.

With the die shot we can see 4x Gracemont cores took 1 slot in the ringbus and roughly 1.5x the size of a goldencove core
If anyone believes one Gracemont is only 27% behind Golden Cove, then

The best config should be a 26 core 28 Thread CPU : 2P + 24E = 128*2+ 101*24 = 2680 %
For comparison, 8P+8E = 128*8 + 101*8 = 1832 %
( The numbers are based on % vs one Comet Lake-S core)
A crazy 46% more MT performance vs the 12900k, yet still have 2 P cores for single threaded tasks.
And the scheduler should help the rest, right?

This 26 core 28 thread CPU will smash anything on the market, right ?

Since Intel didn't do that, that Gracemont performance claim is 99.99% not the whole picture of it.

 

[Zubasa]

If Intel is telling the truth (which is very doubtful) - Gracemont is only 20% behind Golden Cove?!! Damn, 8x Gracemont is looking like a very impressive processor for budget laptops. Hell, even budget gaming laptops with RTX 3050/3060s will be well suited with 8x Gracemont cores.

Notice the wording, it is single thread not single core. Golden/Cypress/Sunny Cove and Skylake all have 2 threads.
That is another 20~30% performance per Core not accounted for in that slide.

 

[regs]

Notice the wording, it is single thread not single core. Golden/Cypress/Sunny Cove and Skylake all have 2 threads.
That is another 20~30% performance per Core not accounted for in that slide.

But they are small, so you can fit 2 Gracemonts into 1 Skylake core. Still 1 cove + 4 monts, 2 coves + 8 monts or even 1 cove + 12 monts would be better option, than just only monts, even in notebooks. Meteor Lake will expand number of monts cores.

 

[PanicLake]

Alder Lake
Tree Lake
Deadhead Lake

who comes up with these names?!?!?

edit and no more 14nm ++++++++ jokes……..sigh……..

They are all name of existing lakes in US
Alder lake

Alder Lake · Washington

★★★★★ · Lake

www.google.com

Tree Lake

Tree Lake · Alban, WI 54473

★★★★☆ · Lake

www.google.com

Deadhead Lake

Deadhead Lake · Washington 98068

★★★★★ · Lake

www.google.com

 

[First Strike]

This 26 core 28 thread CPU will smash anything on the market, right ?

Since Intel didn't do that, that Gracemont performance claim is 99.99% not the whole picture of it.

One reason I'm thinking is probably due to bandwidth. These four cores has to compete for the same bandwidth from the ring (presumably 102GB/s bidirectional), while also competing with 9 other ring nodes. For single-thread performance demonstration, the E-core should not have any competitor at all. I guess if you only pair a few crippled little cores with unaffected big cores then it's okay. The more the worse and eventually becomes like Epyc CCXs connected by a feeble ring bus.

 

[cst1992]

I used a Thermaltake Contact Silent 12 150W cooler. At stock it was perfectly fine. I find it hard to believe your 4690K if not power unlocked is stressing such a cooler. Sounds like a bad or very old paste job (it needs to be re-pasted every 3 years max).

It's Conductonaut on the die + Arctic MX-4 on the IHS, changed about a year ago.

4690k is running at 4.3GHz stable on stock voltage.

 

[Valantar]

One reason I'm thinking is probably due to bandwidth. These four cores has to compete for the same bandwidth from the ring (presumably 102GB/s bidirectional), while also competing with 9 other ring nodes. For single-thread performance demonstration, the E-core should not have any competitor at all. I guess if you only pair a few crippled little cores with unaffected big cores then it's okay. The more the worse and eventually becomes like Epyc CCXs connected by a feeble ring bus.

Not only that, but they're connected to the L3 by the same connection, meaning they're further disadvantaged compared to a Zen CCX. No doubt they'd be able to change this if there was a use case that warranted it (I could see a 48+ core server chip using these on a dual or triple ring bus, or a mesh), but for MSDT? Nah, that sounds too expensive.

 

[qubit]

I'd much rather have 16 full performance cores rather than this. That would be optimised and would wipe the floor with AMD.

 

[Crackong]

I'd much rather have 16 full performance cores rather than this. That would be optimised and would wipe the floor with AMD.

Intel's current ringbus design only supports up to 10 BIG cores.....so nope
And we all knew what happens with 16 cores in mesh.

MCM is the future, but intel have to test the waters with ponte vecchio first.

 

[Valantar]

Intel's current ringbus design only supports up to 10 BIG cores.....so nope
And we all knew what happens with 16 cores in mesh.

MCM is the future, but intel have to test the waters with ponte vecchio first.

They could always run a dual ring bus, which would put them more or less on par with AMD in terms of latencies across core groupings.

 

[qubit]

@W1zzard If you haven't already done so in the review, it would be nice to add a section where you disable the high efficiency cores in the bios and just benchmark it with the 8 high performance ones. You can then do the reverse and compare again. The difference in overall performance will be interesting to see.

 

[TheoneandonlyMrK]

@W1zzard If you haven't already done so in the review, it would be nice to add a section where you disable the high efficiency cores in the bios and just benchmark it with the 8 high performance ones. You can then do the reverse and compare again. The difference in overall performance will be interesting to see.

Apparently the E core's can be disabled, but one P core has to be enabled to boot.

 

[Turmania]

If they consume around same power as their rival counterparts, I would call it a win for Intel. So just have to wait for reviews before jumping to any conclusion. But Intel CEO was very positive about the product. So its his reputation on the line.

 

[qubit]

Apparently the E core's can be disabled, but one P core has to be enabled to boot.

In that case, it might be possible to mask it out in Windows. Boot performance couldn't be compared, though. Wouldn't surprise me if it requires W11 to properly utilise those cores, including masking them off.

Wouldn't it be ironic if one of these was installed to a mobo without a TPM and then didn't qualify for W11?!

 

[Vayra86]

You know, as promising as the first half of that slide deck looks,

that's about how the second half creates lot of doubt about all the performance claims. HOW MANY pages of disclaimer and explanation is there? Its more than the official slides themselves. They took an effort to include so much fine print you'll simply go TL DR on it.

That tells me a lot more than the first half of the slides.

That, and the fact PL2 is set at 241W while the spec sheet tells us Max turbo at 150W. The lies are still in the product and the improvements are mostly achieved by high frequency. So sure, IPC uplift, I'll believe it. In a best case scenario. Let's see where the locked CPUs end up... but my first look at this is the biggest change in the real world is the marketing, not the CPU's real world performance.

 

[WhoDecidedThat]

Notice the wording, it is single thread not single core. Golden/Cypress/Sunny Cove and Skylake all have 2 threads.
That is another 20~30% performance per Core not accounted for in that slide.

here's my perspective on things -

let's say a golden cove core running 1 thread gives 100 units/sec performance.
golden cove running 2 threads gives 125 units/sec performance.

1 gracemont core running 1 thread gives 78 units of performance
2 gracemont cores running 2 threads give 146 units/sec performance

now this is iso frequency.
we know that golden cove can clock higher than gracemont.
so those golden cove cores can easily run at 20% higher clock speeds and match 2 gracemont cores in terms of performance.

the questions is - what consumes more silicon area and more power - 2 gracemont cores at 3.9 GHz or 1 golden cove core running 2 threads at 4.7 GHz?

@W1zzard If you haven't already done so in the review, it would be nice to add a section where you disable the high efficiency cores in the bios and just benchmark it with the 8 high performance ones. You can then do the reverse and compare again. The difference in overall performance will be interesting to see.

now that would be telling wouldn't it.
8 P cores running an application v/s 8 E course running an application with information on the power consumption difference between the two.

Theoretically, 1.27x IPC * (5 GHz/3.9 GHz) * 1.25x Hyper Threading = 2.03 times the multithreaded performance of 8 P cores running 16 threads vs 8 E cores running 8 threads.

 

[Zubasa]

here's my perspective on things -

let's say a golden cove core running 1 thread gives 100 units/sec performance.
golden cove running 2 threads gives 125 units/sec performance.

1 gracemont core running 1 thread gives 78 units of performance
2 gracemont cores running 2 threads give 146 units/sec performance

now this is iso frequency.
we know that golden cove can clock higher than gracemont.
so those golden cove cores can easily run at 20% higher clock speeds and match 2 gracemont cores in terms of performance.

the questions is - what consumes more silicon area and more power - 2 gracemont cores at 3.9 GHz or 1 golden cove core running 2 threads at 4.7 GHz?

The last question is very difficult to answer, because each architecture and each process node has its own power vs frequency curve.
All that Intel slide shown is how the 2 designs performs at some given frequency.
We have no idea what that frequency is or how much power-draw for each design in that specific scenario.
Some designs scale well with higher power while others actually don't scale very well at very low power.

As for die space I would assume the e-cores are significantly more efficient or else Intel wouldn't have jump through hoops to make it happen.
Also there is the fact that Intel had problems with 10nm / Intel 7 yields for years. It makes sense for Intel to go as die-space efficient as possible to improve yield.

 

[Valantar]

It will definitely be interesting to see how performance on these pans out in the real world, especially with patched W11 and Zen3. Looking forward to the review!

A few thoughts:
- I like the change in power terminology, the death of TDP, the rise of PBP(PL1), MTP(PL2) and the officially sanctioned MCE mode for K SKUs (with PL1=PL2). Much clearer.

- The promise of ISO MT performance to a 241W 11900K at 65W is downright staggering. If true, that will be immensely beneficial for mobile workstations. Of course I'm expecting this to be - at best - true only in applications that scale nearly perfectly with more cores, and otherwise highly variable. Still, quite the claim!

- The biggest, most interesting thing to me though: XMP 3.0. User-configurable XMP profiles with human readable names? THANK YOU. No more CMOS resets losing your saved memory profiles, and you won't have to recall whether custom profile 1 or 2 was the one you set up to use two years ago. That live on/off switching for XMP is also really cool, though I see it being far more useful in mobile, yet its limited to overclockable SKUs, which for the most part will mean desktops. Kind of weird. But I guess we might see it on high end mobile parts as well? I also don't quite see why this couldn't switch between JEDEC profiles with equal ease, especially in mobile (where XMP is essentially nonexistent). More promising for the future than immediately useful, but I like it.

 

[ppn]

the questions is - what consumes more silicon area and more power - 2 gracemont cores at 3.9 GHz or 1 golden cove core running 2 threads at 4.7 GHz?

it's 4E core vs 1P core at same area, same clocks 5.2Ghz, if 1.27 IPC and 1,25 hyperthreading/. 4x vs 1.6x at same power.