Discussion Intel current and future Lakes & Rapids thread

[Exist50]

Based on what calculation? A 13700k pulls in 200 watts at 5.0Ghz on Prime95 small (with E-cores active @ 4.0Ghz). For argument sake, lets say it pulls 200W with the E-cores off. They are on the same process node as SPR and the P-cores essentially the same cores (minus AVX512). To me, 600W seems more likely than 1000W. Now, if the test were being run using AVX512, and there was no offset in frequency for AVX512, I could see it using more than 600W, but 1000 still seems very unlikely. I will wait until I see real reviews/tests before jumping to conclusions.

Don't think SPR uses the Raptor Lake improvements, but Alder Lake should make for a decent reference.

 

[Edrick]

Don't think SPR uses the Raptor Lake improvements, but Alder Lake should make for a decent reference.

You may be correct, I don't know for certain. I would have hoped they would, but it is Intel we are talking about. What was an Alder Lake 12700k (minus E-cores) pulling at 5.0Ghz?

Edit: I am seeing numbers online in the 270-280W range for 5.0Ghz with E-cores active). That's a fairly big jump. So 700-800W doesn't seem out of the question if SPR is not leveraging any Raptor Lake improvements.

 

[BorisTheBlade82]

You may be correct, I don't know for certain. I would have hoped they would, but it is Intel we are talking about. What was an Alder Lake 12700k (minus E-cores) pulling at 5.0Ghz?

Edit: I am seeing numbers online in the 270-280W range for 5.0Ghz with E-cores active). That's a fairly big jump. So 700-800W doesn't seem out of the question if SPR is not leveraging any Raptor Lake improvements.

That all depends on the question, if there is a TDP cap set.
Please have a look at the ST result table in the linked post, line 99.

An unlimited ADL 12900K draws 41w in ST CB23. Yes, there is some uncore offset, to be considered. But 800-1000w for a 24c SPR is absolutely not out of this world.

Discussion - PES | Assessing Power and Performance Efficiency of x86 CPU architectures

Dear Community, so this is my first thread here as a long-time lurker - but I felt the desire to share a small hobby-project of mine from the last couple of months with you... Performance Efficiency Suite - What is it about? Most Reviewers solely focus on what they consider to be the most...

forums.anandtech.com

 

[Just Benching]

So Intel HEDT is finally back. But only for those that have super cheap electricity or steal it from their rich neighbors

You realize - you are not supposed to be running 56 cores at 5ghz - right? As per the techpowerup review, competitors cpus pull more power per core for similar clockspeeds. They would also be drawing 1000 watts. It's common sense, the reason you go for 50+ core CPUS is because you are running multithreaded workloads - therefore you don't need 5ghz clockspeeds. Duh

The 12900k seems to have pretty great single threaded efficiency btw.. 26w for the 12900k

 

[semiman]

So Intel HEDT is finally back. But only for those that have super cheap electricity or steal it from their rich neighbors

It must be a High Energy DeskTop

 

[Markfw]

You realize - you are not supposed to be running 56 cores at 5ghz - right? As per the techpowerup review, competitors cpus pull more power per core for similar clockspeeds. They would also be drawing 1000 watts. It's common sense, the reason you go for 50+ core CPUS is because you are running multithreaded workloads - therefore you don't need 5ghz clockspeeds. Duh

The 12900k seems to have pretty great single threaded efficiency btw.. 26w for the 12900k

When you are buying a workstation, an HEDT or a server cpu, single threaded efficiency means nothing. How about looking at multi-threaded power efficiency, and at stock and ECO mode ????

 

[Edrick]

An unlimited ADL 12900K draws 41w in ST CB23. Yes, there is some uncore offset, to be considered. But 800-1000w for a 24c SPR is absolutely not out of this world.

At what frequency? 5.3, 5.4? And a 12900K runs at what, 17W-19W idle? What you posted does not conclude that a single core is using 41W by itself at 5.0Ghz.

 

[BorisTheBlade82]

At what frequency? 5.3, 5.4? And a 12900K runs at what, 17W-19W idle? What you posted does not conclude that a single core is using 41W by itself at 5.0Ghz.

No, the Package power alone is never anywhere near that in idle for ADL. I have none myself, but it should be more around 2-5w.
The used method does not capture frequencies, but as this is an ST run it should be pretty much at max clock.
Here is the detailed result, which was provided to me at ADL release by A. Schilling from HWLUXX: https://forums.anandtech.com/thread...f-x86-cpu-architectures.2597905/post-40624976

As you can see, this is 41w on average. So 24 GLC cores with the right cooling and no TDP limit are at least theoretically able to consume 800-1000w, which is all I am saying.

 

[Just Benching]

No, the Package power alone is never anywhere near that in idle for ADL. I have none myself, but it should be more around 2-5w.
The used method does not capture frequencies, but as this is an ST run it should be pretty much at max clock.
Here is the detailed result, which was provided to me at ADL release by A. Schilling from HWLUXX: https://forums.anandtech.com/thread...f-x86-cpu-architectures.2597905/post-40624976

As you can see, this is 41w on average. So 24 GLC cores with the right cooling and no TDP limit are at least theoretically able to consume 800-1000w, which is all I am saying.

It idles at around 3w with balanced power plan - 1.7w with power saving power plan and 15w with high performance..

 

[nicalandia]

An unlimited ADL 12900K draws 41w in ST CB23. Yes, there is some uncore offset, to be considered. But 800-1000w for a 24c SPR is absolutely not out of this world.

The W7 2495X also has a Mesh Interconnect, OI and PCI lane to feed. That all core 5+ Ghz is trully pulling close to 1000 Watts

 

[nicalandia]

I am not surprised by this

Intel Core i5-13400F Review - Force of Efficiency

The Intel Core i5-13400F is only $200, but comes with six Performance Cores and four E-Cores, bringing the total thread count to 16. Performance is good, as expected, but what's even more impressive is the energy efficiency as confirmed by the detailed testing in our review.

www.techpowerup.com

 

[Edrick]

The W7 2495X also has a Mesh Interconnect, OI and PCI lane to feed. That all core 5+ Ghz is trully pulling close to 1000 Watts

Again, source?

 

[tamz_msc]

Der8auer had the 56-core W9-3495X pulling 500W in Cinebench at stock with 2.9 GHz all-core frequency.

 

[nicalandia]

LTT Be like..

https://twitter.com/x/status/1636065464601411584

 

[Hitman928]

LTT Be like..

View attachment 78199

https://twitter.com/x/status/1636065464601411584

It's unfortunate that these stupid thumbnails and click bait articles/tweets keep working for these youtubers, but that is where we are at I guess.

 

[nicalandia]

Again, source?

Skatterbencher, pulling 650 Watts at only 4.4 Ghz all core.

 

[Edrick]

Skatterbencher, pulling 650 Watts at only 4.4 Ghz all core.

Holy crap! If that is indeed accurate then I am happy I decided against a SPR build. Intel still has work to do.

 

[SiliconFly]

Let me rephrase - it's literally marketing. Intel could call it's nodes whatever it likes. I don't think Intel's made any actual claims yet as to density or quality.

Not at all. It's not marketing. Node comparison:

* Intel 7 is on par with TSMC 7nm (slightly better actually)
* Intel 4 is on par with TSMC 4nm (slightly better actually and comparable to TSMC 3nm in some aspects)
* Intel 20A is superior to TSMC 3mn (on par with upcoming TSMC 2nm)
* Intel 18A is superior to the upcoming TSMC 2nm

Intel's 20A is a pretty futuristic node. The first of it's kind actually! It uses GAAFET transistors (called RibbonFET) & Backside-Power-Delivery (BPD called PowerVia). TSMC is still developing these technologies and will release it with TSMC N2 only in 2026. Whereas, Intel 20A is already taped-out & ready to go.

20A has 15% PPW over Intel 3. And Intel 3 has 18% PPW over Intel 4. It's huge!! In short, Intel 20A has a whopping 36% PPW over Intel 4.

In other words, Arrow Lake will have a massive 63% PPW over the current Raptor Lake!!! These kinda numbers are not easy to even visualize! Its nothing short of a silver bullet for Arrow Lake. This makes ARL extremely power-efficient!

Just imagine a 250W 13900K running at just 90W with same performance! Thats the power of 20A. Not marketing at all.

And if they add the rumored ULP (ultra low-power) E-Core to ARL, its idle power consumption will be a new industry record for desktop class processors!!! (and laptop)

 

[Exist50]

Not at all. It's not marketing. Node comparison:

* Intel 7 is on par with TSMC 7nm (slightly better actually)
* Intel 4 is on par with TSMC 4nm (slightly better actually and comparable to TSMC 3nm in some aspects)
* Intel 20A is superior to TSMC 3mn (on par with upcoming TSMC 2nm)
* Intel 18A is superior to the upcoming TSMC 2nm

Yeah... I'd suggest reigning in your expectations a bit. If Intel pulls that off, great, but as things stand, we don't have any real reason to believe those matchups will materialize.

 

[SiliconFly]

Let me rephrase - it's literally marketing. Intel could call it's nodes whatever it likes. I don't think Intel's made any actual claims yet as to density or quality.

Let me rephrase - it's literally marketing. Intel could call it's nodes whatever it likes. I don't think Intel's made any actual claims yet as to density or quality.

Actually, node names haven't been real in the past 25 years. It was TSMC which first started the whole fake naming trend. Intel is just catching up.

For example, Intel 10nm was far superior than TSMC 7nm (Intel 10nm - 100 MTr/mm2, TSMC 7nm - 92 MTr/mm2)

And TSMC 10nm was actually comparable to Intel 14nm and not Intel 10nm (Intel 14nm - 44 MTr/mm2, TSMC 10nm - 52 MTr/mm2).

TSMC has been using fake marketing names for far more time than Intel. Just that nobody cares when TSMC does it.

 

[SiliconFly]

Yeah... I'd suggest reigning in your expectations a bit. If Intel pulls that off, great, but as things stand, we don't have any real reason to believe those matchups will materialize.

May or may not. No one knows. But then again, only the future has the answers.

 

[LightningZ71]

You know enough to understand that density is but one measure of a process node! Process nodes are best thought of as triangles with the three vertices as follows: Density, Power efficiency and Performance. While Intel 10nm (og) had a theoretical achievable density that was higher than TSMC N7, it sacrificed a whole lot of power effiency, performance and most of all yield efficiency to even exist in a shipping product. Intel 10nm (Ice Lake) managed to recover some yield, but was notably defficient against N7 when measure on power and clock performance. Intel 10nm-sf was likely a solid match for N7 in performance, nearly there for power efficiency at comparable clocks, and could achieve higher production density, but still, judging by market appearance, struggled with qualifying yields vs N7. By the time Intel 7/ Intel 10nmESF hit the market, TSMC was on to N6, which was still at an advantage in power efficiency, didn't differ significantly in achievable production density, and only trailed slightly in clocks at comparable power levels, though Intel was able to throw power efficiency out the door to clock higher.

As for both company's actual leading production parts now, you have Intel 7 (+?) 10nmESF+ in Raptor Lake vs. TSMC N5. Given what we're seeing in shipping products, no one has a clear lead in any category as Intel still trades power for clocks.

What remains to be seen is what Intel's yields look like on Intel 4. TSMC is in production now on N3 and N3Efor AMD and others will be in production in a couple of quarters. In the mean time, TSMC's various N4 products are apparently on target.

It's one thing to have stellar specs, its another to actually produce them.

 

[SiliconFly]

You know enough to understand that density is but one measure of a process node! Process nodes are best thought of as triangles with the three vertices as follows: Density, Power efficiency and Performance. While Intel 10nm (og) had a theoretical achievable density that was higher than TSMC N7, it sacrificed a whole lot of power effiency, performance and most of all yield efficiency to even exist in a shipping product. Intel 10nm (Ice Lake) managed to recover some yield, but was notably defficient against N7 when measure on power and clock performance. Intel 10nm-sf was likely a solid match for N7 in performance, nearly there for power efficiency at comparable clocks, and could achieve higher production density, but still, judging by market appearance, struggled with qualifying yields vs N7. By the time Intel 7/ Intel 10nmESF hit the market, TSMC was on to N6, which was still at an advantage in power efficiency, didn't differ significantly in achievable production density, and only trailed slightly in clocks at comparable power levels, though Intel was able to throw power efficiency out the door to clock higher.

As for both company's actual leading production parts now, you have Intel 7 (+?) 10nmESF+ in Raptor Lake vs. TSMC N5. Given what we're seeing in shipping products, no one has a clear lead in any category as Intel still trades power for clocks.

What remains to be seen is what Intel's yields look like on Intel 4. TSMC is in production now on N3 and N3Efor AMD and others will be in production in a couple of quarters. In the mean time, TSMC's various N4 products are apparently on target.

It's one thing to have stellar specs, its another to actually produce them.

Exactly my point too! It's all about PPA. Power, Performance & Density (area).

Time for AMD/TSMC fans to acknowledge that. Already I've heard a thousand stories saying TSMC is wow, Intel 4 is bad, Intel 20A is bad, TSMC 3nm is superior, Meteor Lake is finished, Zen 5 is light years ahead, etc. All I'm saying is it's time to set the record straight.

The Intel's +++++++ generations are already over. We're at the dawn of a new era in node and processor technologies.

Just imagine this, next year Q4, we're going to have a ton of CPUs built on 2nm & 3nm process nodes!

 

[Doug S]

Whereas, Intel 20A is already taped-out & ready to go.

Nodes don't "tape out", designs do. And no designs have taped out on 20A when Intel isn't even shipping any Intel 4 or Intel 3 stuff yet LOL

 

[SiliconFly]

Nodes don't "tape out", designs do. And no designs have taped out on 20A when Intel isn't even shipping any Intel 4 or Intel 3 stuff yet LOL

Nodes don't tape out. Test chips do. LOL.

Reading helps... see this link and this link.

The first link is about the tapeout in 20A. The second is about tapeout itself.