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.

 

[jpiniero]

Is It Worthwhile Running Intel Alder Lake With mitigations=off? - Phoronix

www.phoronix.com

Some Linux benchmarks comparing turning off the Smeltdown mitigations. Most of them it's only like 1% tops.

 

[IntelUser2000]

Yeah that was a mess. Kinda makes you wonder why they continued with heterogeneous core configs in a flagship product after that particular disaster.

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

??

Nicalandia are you saying Sunny Cove was crap? Because it wasn't bad. Remember Icelake? Tigerlake? They are pretty much using the same cores. Lakefield was bad because it underperformed not only Icelake(which was expected) but Tremont. The Jasper Lake 6W Pentium Silver N6000 outperforms the Core i5 L16G7 in both ST and MT workloads, and that's after they updated the Cinebench version so it can run on Sunny Cove. That's a failure. Not to mention we didn't get better battery life and devices were in the $1000 range.

Intel made few very different approaches that ended up being a failure. Lakefield, and Kabylake-G for example. Both failed not due to competitive pressure(some speculating the latter died due to Nvidia) but because they both sucked. It brought zero advantages that they were supposed to bring. The market is actually logical that way.

And I don't get DrMrLordX's assertion that hetero core configs are crap when its largely responsible for the high performance of Alderlake. You can see from Computerbase.de test that it allows it to perform it the same as if it had twice the TDP. The 125W 8P performs like 65W 8P+8E for example.

And this is going to get good in mobile.

 

[JoeRambo]

Popping the Hood on Golden Cove

Alder Lake (ADL) is the most exciting Intel launch in more than half a decade. For the first time since Skylake, Intel has launched a competitive desktop microarchitecture. But I’m sure you a…

chipsandcheese.com

Some incredible info and comparisons inside. These guys (girls?) really picked up a torch of proper architectural investigation. A highly recommended read!

 

[itsmydamnation]

Popping the Hood on Golden Cove

Alder Lake (ADL) is the most exciting Intel launch in more than half a decade. For the first time since Skylake, Intel has launched a competitive desktop microarchitecture. But I’m sure you a…

chipsandcheese.com

Some incredible info and comparisons inside. These guys (girls?) really picked up a torch of proper architectural investigation. A highly recommended read!

Thax for posting , kind of matches what we all know/have seen in other reviews , its not a bad uarch, its just not amazing and get's the performance it has by being bigger not smarter.

 

[Carfax83]

Thax for posting , kind of matches what we all know/have seen in other reviews , its not a bad uarch, its just not amazing and get's the performance it has by being bigger not smarter.

I think Raptor Cove will expose the full potential of the Golden Cove core and shore up its weaknesses and flaws. That's usually been the case in the past with Intel cores. The revision is always better, which is why I've decided to wait for Raptor Cove to do a full platform upgrade. I wonder how much performance Intel can further squeeze out of Golden Cove? They only need about 10% IPC to compete with what Zen 4 is rumored to bring which is a 20-25% IPC gain, which Intel should easily be able to achieve with microarchitectural tweaks, cache increase and faster DDR5.

 

[DrMrLordX]

And I don't get DrMrLordX's assertion that hetero core configs are crap

Lakefield was crap.

 

[TheELF]

Thax for posting , kind of matches what we all know/have seen in other reviews , its not a bad uarch, its just not amazing and get's the performance it has by being bigger not smarter.

Well all the benchmarks are completely dumb compute jobs so there is zero wiggle room for smarter doing anything.
Nobody benches the new scheduler because that would be extremely hard to do, nobody benches avx 512 instead only uses it to highlight power peaks, nobody gives quicksync or iGPU AI the time of day.

At the end of the day if all that reviews focus on is dumb compute then that's the type of CPUs we will get.

 

[nicalandia]

Nicalandia are you saying Sunny Cove was crap?

Just saying that a Desktop Sunny Cove/Tremont hybrid would not cut it for for desktop against Amd Current and up coming Desktop CPUs, they had to push Alderlake(the 12900K) to match a stock 5950X(When the 5950X is without limits the story changes). Do you thing a 8C/16T Sonny cove + 8C/8T Tremont Hybrid would keep up?

 

[Hulk]

Well all the benchmarks are completely dumb compute jobs so there is zero wiggle room for smarter doing anything.
Nobody benches the new scheduler because that would be extremely hard to do, nobody benches avx 512 instead only uses it to highlight power peaks, nobody gives quicksync or iGPU AI the time of day.

At the end of the day if all that reviews focus on is dumb compute then that's the type of CPUs we will get.

You are right about AVX512 and the iGPU being overlooked or always regarded negatively. The deep dive on Golden Cove Joe Rambo posted a link to does a good job benching the scheduler. In my experience Quicksync quality is so bad compared to (CPU only) Handbrake at equal bitrates I don't care about it.

 

[tomatosummit]

Didn't see it posted yet but igor's have done an av512 review thing.
A variety of results but it looks like on software that can use it 8gc with 512 perform similarly to 8gc+8gr with avx2
They seem to pay a lot of attention to the efficiency of the 512 implementation, I wonder if there's more clock headroom available when running avx512 code which could make it leap ahead somewhat.

Efficiency secret AVX-512 on Alder Lake - The resurrected instruction set in a practical test | Page 2 | igor´sLAB

AVX-512 was a hotly discussed topic around the launch of the new Intel Alder Lake CPUs. At first it was said that the P cores supported it in principle, but in order to be able to offer a common…

www.igorslab.de

 

[Heartbreaker]

AVX512 seems like a real flaw of the Alder Lake design. A lot of silicon that is just being turned off.

I wonder if they are going to squeeze AVX512 into to future E-Cores.

 

[nicalandia]

I wonder if they are going to squeeze AVX512 into to future E-Cores.

That would defeat the purpose of e-cores, to save on space and power

 

[igor_kavinski]

Intel Core i9-12900K E-Cores Only Performance Review

With Alder Lake, Intel is betting big on hybrid CPU core configurations. The Core i9-12900K has eight P(erformance) cores and eight E(fficient) cores. We were curious and tested the processor running the E-Cores only to see how well they perform against architectures like Zen 2, Zen 3, Skylake...

www.techpowerup.com

MySQL hates the E-cores.



Doom Eternal LOVES E-cores



4K gaming doesn't need P-cores. A new BIOS option needed to disable P-cores with a keypress to save power in such a scenario.

 

[Heartbreaker]

4K gaming doesn't need P-cores. A new BIOS option needed to disable P-cores with a keypress to save power in such a scenario.

You aren't going to use much CPU power in that scenario anyway, because the GPU bottle neck will give the CPU lots of idle cores.

 

[JoeRambo]

Doom Eternal LOVES E-cores

Looks more like a GPU limit @310FPS that is easy to hit esp with 30MB of L3 equipped CPU. P cores might scale to 500FPS and E cores could stop at 311FPS for all we know.

 

[coercitiv]

View attachment 53762

4K gaming doesn't need P-cores. A new BIOS option needed to disable P-cores with a keypress to save power in such a scenario.

To save what, 10 Watts in a gaming system that will already be using 400W+?



If you want to save power in a 4K gaming system you simply downclock your CPU. No need to reinvent the wheel.

 

[Ed1]

To save what, 10 Watts in a gaming system that will already be using 400W+?

View attachment 53765

If you want to save power in a 4K gaming system you simply downclock your CPU. No need to reinvent the wheel.

Those values are not gaming ones, they would be even lower there. He uses CB R23.

 

[Hulk]

Seeing how the primary purpose of Gracemont is MT performance in as little area as possible I still think adding HT w

Intel Core i9-12900K E-Cores Only Performance Review

With Alder Lake, Intel is betting big on hybrid CPU core configurations. The Core i9-12900K has eight P(erformance) cores and eight E(fficient) cores. We were curious and tested the processor running the E-Cores only to see how well they perform against architectures like Zen 2, Zen 3, Skylake...

www.techpowerup.com

View attachment 53760

MySQL hates the E-cores.

View attachment 53761

Doom Eternal LOVES E-cores

View attachment 53762

4K gaming doesn't need P-cores. A new BIOS option needed to disable P-cores with a keypress to save power in such a scenario.

Having examined many of these "E cores" only tests and tested my 12700K I won't have much trust in their validity. The fact that the P's can't be completely disabled means Process Lasso or some other software means must be used to try and isolate the E's and they are slippery as the P's always seem to sneak into the action.

Some examples. The E only Handbrake score in the Anandtech bench doesn't make sense. 87fps for the HVEC Handbrake test is clearly not correct.

Cinebench R23 E tests show results that when you add P and E separate results the combined total is much higher than when both are run concurrently. Either the P's are taking care of all background tasks, inflating the E score, or helping with the test, also helping the score.

At this point in time I like E core isolated scores that are obtained by subtracting P scores from the P+E scores. The E cores can be shut down in the BIOS so I have some trust in P only scores. Also an imperfect way to evaluate E cores but I think better than the affinity solution.

Also the elephant in the room is why Intel won't allow the E to "run the show" by themselves? Is there some tasks the P's must perform?

Anyway from testing on my 12700K I have some faith in the following.

P core at 4.7GHz with HT scores 2355. E cores score 973 in ST test at 3.8GHz and 909 subtracting P from P+E result, E's at 3.6 of course.

So, in CB R23 P's do 501 point per GHz while E's do 254 points per GHz.
P cores are 98.6% faster than E's in CB R23, or almost twice as fast. This means that at 4.7/3.6GHz and assuming P and 4E cluster requires the same die space the E's are 1.54 times more productive for the die area they consume.

And that of course is the reason for their existence.

Also I believe that the superwide design of the Gracemont cores is directly related to their "tuning" for MT performance. We have 6 decoders in 2x3 clusters and if I remember correctly 17 decode. But, the OoS is much more primitive than Golden Cove, meaning MT throughput will only be high if the incoming code is already highly parallel as Gracemont doesn't the smarts to work it out on the fly.

 

[coercitiv]

Those values are not gaming ones, they would be even lower there. He uses CB R23.

I know, that was the point: if 8 P cores @ 3.9Ghz with HT disabled use just 22W more than 8 E-cores under Cinebench (while offering 20% higher score), then gaming power delta is bound to be significantly less. In fact, if we were to aim for ISO performance, there's a very good chance the P-cores would end up using less power than the E-cores.

According to the same review the P-cores @ 3.9Ghzstill offer up to 50% more gaming performance over E-cores, which means we could easily drop the P-core clocks to 3Ghz and still match or beat the E-cores while massively dropping power usage.

 

[nicalandia]

Seeing how the primary purpose of Gracemont is MT performance in as little area as possible I still think adding HT w


Having examined many of these "E cores" only tests and tested my 12700K I won't have much trust in their validity. The fact that the P's can't be completely disabled means Process Lasso or some other software means must be used to try and isolate the E's and they are slippery as the P's always seem to sneak into the action.

Some examples. The E only Handbrake score in the Anandtech bench doesn't make sense. 87fps for the HVEC Handbrake test is clearly not correct.

Cinebench R23 E tests show results that when you add P and E separate results the combined total is much higher than when both are run concurrently. Either the P's are taking care of all background tasks, inflating the E score, or helping with the test, also helping the score.

At this point in time I like E core isolated scores that are obtained by subtracting P scores from the P+E scores. The E cores can be shut down in the BIOS so I have some trust in P only scores. Also an imperfect way to evaluate E cores but I think better than the affinity solution.

Also the elephant in the room is why Intel won't allow the E to "run the show" by themselves? Is there some tasks the P's must perform?

Anyway from testing on my 12700K I have some faith in the following.

P core at 4.7GHz with HT scores 2355. E cores score 973 in ST test at 3.8GHz and 909 subtracting P from P+E result, E's at 3.6 of course.

So, in CB R23 P's do 501 point per GHz while E's do 254 points per GHz.
P cores are 98.6% faster than E's in CB R23, or almost twice as fast. This means that at 4.7/3.6GHz and assuming P and 4E cluster requires the same die space the E's are 1.54 times more productive for the die area they consume.

And that of course is the reason for their existence.

Also I believe that the superwide design of the Gracemont cores is directly related to their "tuning" for MT performance. We have 6 decoders in 2x3 clusters and if I remember correctly 17 decode. But, the OoS is much more primitive than Golden Cove, meaning MT throughput will only be high if the incoming code is already highly parallel as Gracemont doesn't the smarts to work it out on the fly.

You could try a Virtual Machine and have it used only the P Cores, do the same for E cores and compare it, Process Lasso is obviously not the best since the P cores are taking care of other task the OS is doing while the e cores are being tested


Edit.
I would like to add that Virtual Machines can be use to test the strength or weakness of a particular CPU design, for example on Zen/Zen+ there was big penalties from apps jumping thru cpu chiplets. to test this you could design a Virtual Computer and assign core in the worst possible way. One could also do the same with Alderlake.

 

[Ed1]

You could try a Virtual Machine and have it used only the P Cores, do the same for E cores and compare it, Process Lasso is obviously not the best since the P cores are taking care of other task the OS is doing while the e cores are being tested


Edit.
I would like to add that Virtual Machines can be use to test the strength or weakness of a particular CPU design, for example on Zen/Zen+ there was big penalties from apps jumping thru cpu chiplets. to test this you could design a Virtual Computer and assign core in the worst possible way. One could also do the same with Alderlake.

Those using Processlasso make sure you are on the latest version as it adds support for Alder-lake and also has CPUsets, though I don't know if that would change outcome of your tests.

On my end with 12600k I get with E cores only CM-r15 602 and 1834 with just P cores. though it takes a few seconds for all P cores to load to 100% so should be higher..
All 10 cores it gets 2589

 

[nicalandia]

Those using Processlasso make sure you are on the latest version as it adds support for Alder-lake and also has CPUsets, though I don't know if that would change outcome of your tests.

On my end with 12600k I get with E cores only CM-r15 602 and 1834 with just P cores. though it takes a few seconds for all P cores to load to 100% so should be higher..
All 10 cores it gets 2589

A virtual Machine is a better way to test since the P cores will not allow the OS to crash if the E CPUs are 100% stressed. in a Virtual Machine the OS on the virtual hardware can't have access to the P cores

 

[Ed1]

A virtual Machine is a better way to test since the P cores will not allow the OS to crash if the E CPUs are 100% stressed. in a Virtual Machine the OS on the virtual hardware can't have access to the P cores

I forgot to mention, I have VMware player but it isn't working right as far as the core count setting that it uses. I think it needs a update for alder lake.
I can not get higher than 3 cores for it to use no matter my setting.

 

[DrMrLordX]

nobody benches avx 512

That's not entirely true. The problem is that the same general benchmarks you saw ran before Ice Lake/Tiger Lake/Rocket Lake came out are the ones being run after the launch of those CPUs. So now that we have a lot of consumer-level implementations of AVX-512, we still don't have a lot of software vendors taking advantage of it. That is reflected in the benchmark choices made by reviewers.

FYI there are a few benchmarks where AVX-512 does make a difference, like y-cruncher and 3DPM v2

 

[Heartbreaker]

The source of that infamous graph. Only Shadows of the Tomb Raider has that oddball performance with E-cores: