Intel Core i9-12900K E-Cores Only Performance

[W1zzard]

at various TDP levels? 15,25,35,65, 105,125,150?

doing DDR5 scaling first, then maybe TDP testing if I can find the time

 

[napata]

I look at alder lake in general as a high consumption mess
These E-cores only? THESE impress me.

An E-core based laptop for example, would be quite interesting to see


(To be clear, this is sarcasm. Even the E-cores are power hungry monsters)

What are you talking about? AL is at the top in efficiency. Only the 5900x & 5950x are significantly better.

https://tpucdn.com/review/intel-core-i9-12900k-e-cores-only-performance/images/efficiency-multithread.png

 

[Xuper]

hmm , interesting.E-cores is 3.9% slower than Ryzen 3300x in 1080p and power draw almost on par with 3300x, It's just Ryzen 3300x.

 

[birdie]

Looking at the last chart, Cinebench energy usage, the "energy efficent" cores are in practice less efficient than the P-Cores, and less efficient than all the Zen 3 CPUs, when you actually need to get some work done.

And when you enable both the P-Cores and the E-Cores, the energy usage drops by just 3% compared to using just the P-Cores. Is this worth the hassle of having to deal with all the quirks of a hybrid architecture?

Even in the single threaded SuperPI energy test, the E-Cores were barely more efficient than the 5800X cores.

While the big/little architecture had some success in the mobile device market, I don't think Intel's implementation is worth the hassle for the time being.

That being said, if I was looking to build a new computer from scratch, I would consider Intel, since the P-Cores give the best gaming performance at the moment. But, even if I bought an Intel CPU, I would probably disable all the E-Cores, to make sure the thread scheduler doesn't mess up.

hmm , interesting.E-cores is 3.9% slower than Ryzen 3300x in 1080p and power draw almost on par with 3300x, It's just Ryzen 3300x.

I look at alder lake in general as a high consumption mess
These E-cores only? THESE impress me.

An E-core based laptop for example, would be quite interesting to see


(To be clear, this is sarcasm. Even the E-cores are power hungry monsters)

The best part of this whole review: these maxed-out E-cores consume MORE , so doubling performance/watt by doubling core count is almost impossible

This was addressed in the review, maybe read it fully for a change: "Second, the E-cores on the i9-12900K are saddled with that large 30 MB L3 cache, and other power-hungry processor components which are affecting efficiency".

Lastly in case ya'll haven't noticed ADL CPUs with both P and E cores do not allow to completely disable all P-cores, at least one P-core is required to drive the CPU.

 

[Sabotaged_Enigma]

Impressive by these E-cores and great article!
But I still prefer that it was made "all P-cores" if possible. Reasons are quite simple: There's no need for a scheduler. And as far as I'm concerned, it's virtually impossible to schedule every task into proper P-cores or E-cores. Sometimes it just doesn't work properly. Maybe you want a heavy-load rendering task to be on the P-cores in the back and at the same time at the front you use the browser or something. But it's very possible when a new software comes in and remains to be optimised. This doesn't happen when you make it "all P-cores".
Just saying. I mean, we don't know how to make a chip better they do, do we?

 

[Pilgrim]

My favourite thing about Alder Lake is the Gracemont cores. They're marvels of semiconductor engineering.

 

[birdie]

Impressive by these E-cores and great article!
But I still prefer that it was made "all P-cores" if possible. Reasons are quite simple: There's no need for a scheduler. And as far as I'm concerned, it's virtually impossible to schedule every task into proper P-cores or E-cores. Sometimes it just doesn't work properly. Maybe you want a heavy-load rendering task to be on the P-cores in the back and at the same time at the front you use the browser or something. But it's very possible when a new software comes in and remains to be optimised. This doesn't happen when you make it "all P-cores".
Just saying. I mean, we don't know how to make a chip better they do, do we?

• You can perfectly disable E-cores.
• There will be soon CPUs with no E-cores at all.
• Pinning tasks to P/E cores works beautifully from what I've seen so far (ProcessLasso/Task Manager/etc).
• Very few applications so far have been outright incompatible with ADL and this will be improved/solved sooner rather than later.
• E-Cores are performant enough not to actually rile about it in case W11 mismanages everything.

 

[MikeSnow]

This was addressed in the review, maybe read it fully for a change: "Second, the E-cores on the i9-12900K are saddled with that large 30 MB L3 cache, and other power-hungry processor components which are affecting efficiency".

Lastly in case ya'll haven't noticed ADL CPUs with both P and E cores do not allow to completely disable all P-cores, at least one P-core is required to drive the CPU.

Reading comprehension and severe form of ADD have become a major issue it seems and that's coming from a person who actually suffers from severe ADD/memory issues.

Yes, that exactly what I was saying: Intel's implementation is not ideal. It doesn't matter if the actual piece of silicon dedicated to the e-Core is actually very efficient, if that efficiency doesn't translate into increased system efficiency overall. And, in many situations, it seems it does not.

And considering it complicates things for the operating system and the application developers, its value is even more questionable. For example, I bought a 5800X instead of a 5900X or 5950X not because I couldn't afford them, but because they have 2 CCDs, making thread scheduling more complicated compared to a single CCD.

Reading comprehension and severe form of ADD have become a major issue it seems

I'll try to ignore this part of your comment.

 

[defaultluser]

I told folks who cant stop making love with Alder Lake that the performance in E-core-only servers would be around 60% (these less dense Xeons will all turbo over 4.,5 ghz, so terrible webserver latency). and the lack of SMT means poor database performance

And, when you target those isolated clouds loads on your customized stack , you must simultaneous,y compete with Amazon's Neoverse!

Intel still hasn't put more than 8 of these in a single socket!

Intel® Xeon® E Processors

Intel® Xeon® E processors specifications, benchmarks, features, Intel® technology, reviews, pricing, and where to buy.

www.intel.com

The best part of this whole review: these maxed-out E-cores consume MORE , so doubling performance/watt by doubling core count is almost impossible

 

[Sabotaged_Enigma]

• You can perfectly disable E-cores.
• There will be soon CPUs with no E-cores at all.
• Pinning tasks to P/E cores works beautifully from what I've seen so far (ProcessLasso/Task Manager/etc).
• Very few applications so far have been outright incompatible with ADL and this will be improved/solved sooner rather than later.

You miss my point. We've got to consider the possibility that schedulers mis-schedule tasks.
And I don't appreciate thing like "manually disabling E-cores" or what. That costs me the threads of them.
AND they are launching models with no E-cores at all, as you said. So the question is, why didn't they make it all P-cores in the first place, something like 16c/32t? Because they couldn't. Who would give up that 8 more threads? Either they couldn't make a silicon with 16 full-power P-cores that's small enough, or they can't control power consumption. That's the problem, and that's what I'm tryna say. I prefer 16 full-power "P-cores" which provide me 32 threads, and at the same time I do NOT AT ALL have to worry about mis-scheduling.

• E-Cores are performant enough not to actually rile about it in case W11 mismanages everything.

OK. Next time you try putting your video rendering task on E-cores which takes more time, and live with it.

I may be talking directly, but please bear with me.
I expressed my preference, but it doesn't mean I hate E-cores. And I believe when it comes to a problem, simple ways are always the best answer. The more complex it is, the less reliable it'll be.

 

[birdie]

You miss my point. We've got to consider the possibility that schedulers mis-schedule tasks.
And I don't appreciate thing like "manually disabling E-cores" or what. That costs me the threads of them.
AND they are launching models with no E-cores at all, as you said. So the question is, why didn't they make it all P-cores in the first place, something like 16c/32t? Because they couldn't. Who would give up that 8 more threads? Either they couldn't make a silicon with 16 full-power P-cores that's small enough, or they can't control power consumption. That's the problem, and that's what I'm tryna say. I prefer 16 full-power "P-cores" which provide me 32 threads, and at the same time I do NOT AT ALL have to worry about mis-scheduling.

OK. Next time you try putting your video rendering task on E-cores which takes more time, and live with it.

I may be talking directly, but please bear with me.
I expressed my preference, but it doesn't mean I hate E-cores. And I believe when it comes to a problem, simple ways are always the best answer. The more complex it is, the less reliable it'll be.

16 P-cores will have a TDP of around 400W or something, I'm not sure many people would want that if any. Intel has been rumored to offer even more E-cores in Raptor Lake so MT performance will be higher/what people are looking for. P-Cores make sense only for tasks which are bad at parallelizing and there aren't that many of them, at least not that many that normal people will run simultaneously. OK, Windows Updates could be one of them, maybe your web browser, maybe an old odd app or game here and there, we are looking at most four maybe five. Everything else may use as many E-cores as possible. We've already had this discussion here and here. It's getting tiresome to see the same questions being asked and answer them again.

 

[blu3dragon]

Is it me, or do the down clocked P cores do relatively better in the game benchmarks than the clock speeds suggest they should? Does disabling HT help in games, or is this all down to cache architecture and size?

 

[mechtech]

Nice work W!zz, interesting stuff. One would have thought those e-cores would have used less power than they do.

 

[Wirko]

The best part of this whole review: these maxed-out E-cores consume MORE , so doubling performance/watt by doubling core count is almost impossible

Yes, that's very interesting, and contrary to what I predicted. E-core perf/W will probably be better in the smaller ADL dies, because of less cache, but we won't see those anytime soon.

E-cores still have a solid advantage in perf/mm² so they are not total nonsense, or purely a marketing trick.

It would be interesting to have at least some benchmarks with P-cores only but with HT enabled. Maybe there's a second Wiz2ard around to do all that.

 

[ARF]

I'm interested if it's economical to mine some cryptocurrency (maybe Monero?) on E-cores. Maybe someone in discussion can share his findings.
Thanks for great review W1zzard!

No.

Yeah, "fallen", right there next to AMD.

Intel is in a really bad position now. And the 400W power figures prove it:

16 P-cores will have a TDP of around 400W or something, I'm not sure many people would want that if any.

 

[persondb]

Our testing shouldn't be taken as guidance for any future Pentium Silver product with eight "Gracemont" cores because the E-cores on an i9-12900K have the luxury of that sweet 30 MB L3 cache, a 160-bit DDR5 memory interface (...)

Did you mean 128-bit memory interface? It's only 160-bit when counting ECC bits, which aren't being used anyway and I don't even think the memory controller supports it.

Looking at the last chart, Cinebench energy usage, the "energy efficient" cores are in practice less efficient than the P-Cores, and less efficient than all the Zen 3 CPUs, when you actually need to get some work done.

And when you enable both the P-Cores and the E-Cores, the energy usage drops by just 3% compared to using just the P-Cores. Is this worth the hassle of having to deal with all the quirks of a hybrid architecture?

Even in the single threaded SuperPI energy test, the E-Cores were barely more efficient than the 5800X cores.

While the big/little architecture had some success in the mobile device market, I don't think Intel's implementation is worth the hassle for the time being.

That's because those aren't efficient when going at full speed. Even Intel slides say so. If you see the image below, you will see that Intel is saying that perf/watt for ST is better when there isn't need to activate the P-Core. Because the E-core can perform at lower power modes than the minimum of the P-core.

So this means that you will see effects on power efficiency when you are just lightly browsing the web or using a light office app. You can kind of see the same paradigm in mobile SoCs, where they have kinda of transitioned into a Big+Middle+Little cores model, e.g. that lastest Dimensity 9000 has 1 Cortex X2, 3 Cortex A710, and 4 Cortex A510. One of the reasons that Cortex A510 are there is for light usage and if you run them at full power, their perf/watt is worse than bigger cores.

The other and main reason is MT performance. But this is going off-topic since it was just about efficiency.

 

[chrcoluk]

I dont think the E-Cores have ever been energy efficient as such, they just less powerful, however they are efficient in silicon real estate.

 

[MikeSnow]

That's because those aren't efficient when going at full speed. Even Intel slides say so. If you see the image below, you will see that Intel is saying that perf/watt for ST is better when there isn't need to activate the P-Core. Because the E-core can perform at lower power modes than the minimum of the P-core.

Not denying that. The problem is that the net result, even when the computer is idle, is an improvement of 8W compared to using just the P-Cores: from 64W to 56W used by the system. I guess that is the best case scenario for the E-Cores.

If I ran such a system idle 24x7, my energy bill would be up to 9 euro less, over an entire year, thanks to the E-Cores. So, I have to repeat my question, is it worth the hassle, for less than 1 euro per month in energy savings? Especially when those savings could actually become losses when you need put some load on the system.

And, if you really are that desperate to save those 8W at idle, you could just use a 5950X system, which draws just 51W when idle, so you would actually save 13W.

 

[persondb]

Not denying that. The problem is that the net result, even when the computer is idle, is an improvement of 8W compared to using just the P-Cores: from 64W to 56W used by the system. I guess that is the best case scenario for the E-Cores.

If I ran such a system idle 24x7, my energy bill would be up to 9 euro less, over an entire year, thanks to the E-Cores. So, I have to repeat my question, is it worth the hassle, for less than 1 euro per month in energy savings? Especially when those savings could actually become losses when you need put some load on the system.

And, if you really are that desperate to save those 8W at idle, you could just use a 5950X system, which draws just 51W when idle, so you would actually save 13W.

Well, if it`s at idle then there are many other factors at play. At idle, most of the CPU time is spent sleeping, I can see that behavior in my 3700X when I am just on my desktop and with (almost) nothing running, most of the cpu cores will sleep and 1-2 will sometimes waken up, do something then go back to sleep. So this means that other system factors like SoC uncore is more important for that.

What I meant is for light usage, like just browsing the web and whatever, anything that would cause some load in the CPU but far from being full throttle. Stuff like RGB software, Microsoft Teams, e-mail clients, Anti-virus, and whatever bloat running, could get some power improvements.

For the 5950X power consumption numbers at idle, there are some variations, I guess. The original review list it as 54W. Just very slight test setup variation can have some effects on there.

 

[Why_Me]

When Intel is desperate they make up “gears” and “E Cores” it’s pretty pathetic how far they’ve fallen

 

[RandallFlagg]

The comments about transcoding while gaming (i.e. recording / streaming) with AL E-cores seems spot on. However, it also seems to need DDR5 to effectively both game and transcode.

When you have that combo of AL+DDR5 and in that use case, the results are pretty stunning :

 

[Pepamami]

It would be interesting to see Intel release Ecore only cpu, like 32core or even 64 one, coz look at the size of these ecores, its 4 times smaller

 

[Wirko]

It would be interesting to see Intel release Ecore only cpu, like 32core or even 64 one, coz look at the size of these ecores, its 4 times smaller

They are four times smaller in PowerPoint, and three times smaller under the microscope.

https://en.wikichip.org/wiki/intel/microarchitectures/alder_lake

 

[Shrek]

This seems like a rather important evaluation of the E-cores, but far too long winded in my opinion.

Reminds me of Blaise Pascal
"I'm sorry I wrote you such a long letter; I didn't have time to write a short one."

 

[GURU7OF9]

Intel has to do a bit more work on these e cores cause they are quite inefficient. They should have just used the skylake core instead.

@W1zzard : Can you pretty please test this CPU at various TDP levels? 15,25,35,65, 105,125,150? So that we can see it will do in the mobile space for example.

Yeah that would be great!
So we could find the point of diminishing returns so once you go over x amount of watts performance improvement is negligable, especially for gaming

I think this would be a very interesting test. Looking at stress test results many people say that these CPUs are inefficient; I think they only look so because they are pushed to the limit by default like an overclocker would do. Lowering the power limit to more reasonable levels could make the situation clearer.