Discussion Intel current and future Lakes & Rapids thread

[mikk]

If the l2 cache is the same, then it's an Alder Lake.

There is also a 13400 Raptor with only 1.25MB. It's not so easy.

 

[nicalandia]

There is also a 13400 Raptor with only 1.25MB. It's not so easy.

13600 non-K is a "Raptor Lake" CPU with Alder Lake L2$ for both P and e Cores.

 

[jpiniero]

13400 looks like it can be either. Presumably they will cut the L2 cache on the Raptor Cove models.

The other locked i5's are Golden Cove.

 

[nicalandia]

13400 looks like it can be either. Presumably they will cut the L2 cache on the Raptor Cove models.

Reports say that there will be two dies. So far we have only seen the Alder Lake type(same L2$ for P and e cores)

 

[DrMrLordX]

There is also a 13400 Raptor with only 1.25MB. It's not so easy.

You sure that's Raptor Cove? Sounds like Golden Cove.

 

[jpiniero]

Reports say that there will be two dies. So far we have only seen the Alder Lake type(same L2$ for P and e cores)

Two dies yes but the L2 will be the same.

 

[nicalandia]

Dr. Ian Cutress saw DVLR and CBR23 Demo back on September being demoed on Raptor Lake. For reasons only known to Intel they have decided to hold back. Likely for the 13900HX to provide a good performance/Efficiency against what's a seriOus threat from AMD Zen4 Dragon Range.

 

[igor_kavinski]

Promises promises and failing to keep them. What do we do with you, Intel?

 

[HurleyBird]

As is reasonably well known in the business, each "last 5%" of performance costs an additional 50% of the total size of the processing core.

Citation needed on that one.

 

[IntelUser2000]

DLVR and other power management features are not a critical feature for desktop.

The "20% power saving or 7% clock increase" and thinking it would apply to a 200W+ 5GHz+ CPU is misunderstanding of the patent. I doubt it'll get anywhere near the claim in the datasheet at those figures.

The early roadmap leaks had DLVR for mobile, not desktop.

Hence, it's likely not worth it implementing it for desktop chips. DLVR unlike FiVR increases motherboard costs, and we know motherboards are a low margin business.

I was doubting the potential performance gain of a Raptor Lake Refresh for desktop, however the addition of DLVR can provide a 20% gain in efficiency or 7% gain in performance.

Again this is a misunderstanding of the patent. DLVR itself doesn't mean anything - it stands for Digital Linear Voltage Regulator. The patent says you could use FiVR for the same purpose.

The patent tells us the advantage is implementing the second regulator in addition to the main one. The second regulator happens to be a digital linear regulator. Digital, because it can be software controlled, and Linear because it's not switched mode.

The advantage of the second regulator is in the potential, instead of when the second linear regulator is active. The most efficient mode of the "DLVR" is when the second regulator is inactive! Oh the irony! That's why I said "potential" is the advantage. It allows the main regulator to be more precise in predicting voltage ranges, which means you can lower the voltage.

The patent explicitly tells us if the second DLVR is more active, the less gain you'll have since linear regulators are inefficient - and this is simple knowledge, almost Electronics 101.

 

[nicalandia]

DLVR and other power management features are not a critical feature for desktop.

The "20% power saving or 7% clock increase" and thinking it would apply to a 200W+ 5GHz+ CPU is misunderstanding of the patent. I doubt it'll get anywhere near the claim in the datasheet at those figures.

Hence, it's likely not worth it implementing it for desktop chips.

Will they backport the DLVR technology to those SKUs that are using Golden Cove/OG Gracemont Cores?

 

[IntelUser2000]

Will they backport the DLVR technology to those SKUs that are using Golden Cove/OG Gracemont Cores?

The patent says the reason for using the otherwise inefficient DLVR instead of FiVR is because latter noticeably increases die size, while the former does nothing. There's probably some firmware that needs to be adjusted in the CPU, but changes are likely very little if anything.

I am almost tempted to post a separate thread about this topic.

 

[Exist50]

The patent says the reason for using the otherwise inefficient DLVR instead of FiVR is because latter noticeably increases die size, while the former does nothing. There's probably some firmware that needs to be adjusted in the CPU, but changes are likely very little if anything.

DLVR certainly does impact die area. You can even see it in the MTL die shots. And my understanding is that it supersedes FIVR. Might be more to it than that one patent.

 

[IntelUser2000]

DLVR certainly does impact die area. You can even see it in the MTL die shots. And my understanding is that it supersedes FIVR. Might be more to it than that one patent.

That's possible cause I don't know what they are actuallly doing. And just like the patent says, it could be used with FiVR. So you'd have second FiVR. Just saying don't take reference numbers from the patent and apply it to everything, like your dear K CPU. You are just regurgitating the idiot press.

 

[Exist50]

That's possible cause I don't know what they are actuallly doing. And just like the patent says, it could be used with FiVR. So you'd have second FiVR. Just saying don't take reference numbers from the patent and apply it to everything, like your dear K CPU. You are just regurgitating the idiot press.

Well I'm certainly not parroting those numbers. If it replaces FIVR, I expect it to be in the same general ballpark.

 

[IntelUser2000]

I think what might have happened is that they found out the gains are far far less in desktop, probably in the ~5% range or even less and there was pushback among motherboard manufacturers. And this is only if it's being used with stock settings.

This is essentially manufacturer certified undervolting. It's better, but just like 5GHz CPUs meant you pretty much had no OC headroom, after this you'll have no undervolting headroom.

Conversely if you overclock, you are essentially eliminating the voltage headroom set at the factory so you can clock higher. So how would this new feature benefit you? Even if they made it work with overclocking? It wouldn't.

 

[Exist50]

I think what might have happened is that they found out the gains are far far less in desktop, probably in the ~5% range or even less and there was pushback among motherboard manufacturers. And this is only if it's being used with stock settings.

This is essentially manufacturer certified undervolting. It's better, but just like 5GHz CPUs meant you pretty much had no OC headroom, after this you'll have no undervolting headroom.

Conversely if you overclock, you are essentially eliminating the voltage headroom set at the factory so you can clock higher. So how would this new feature benefit you? Even if they made it work with overclocking? It wouldn't.

I don't see why this would complicate things for mobo manufacturers. FIVR with Haswell made things significantly easier for them. And the advantage of on-die power regulation is to more quickly respond to changes in demand, and on a more granular basis (i.e. no shared voltage for every core). Don't see how that directly correlates to overclocking headroom.

 

[dullard]

The advantage of the second regulator is in the potential, instead of when the second linear regulator is active. The most efficient mode of the "DLVR" is when the second regulator is inactive! Oh the irony! That's why I said "potential" is the advantage. It allows the main regulator to be more precise in predicting voltage ranges, which means you can lower the voltage.

The patent explicitly tells us if the second DLVR is more active, the less gain you'll have since linear regulators are inefficient - and this is simple knowledge, almost Electronics 101.

Yes, when the second regulator is active, it will be less efficient. But, you are leaving out that in all other cases it will be more efficient. This is because the power supply can be run at a level that better matches those use cases. So, the efficiency depends on the use. Heavy 24/7 turbo uses will be less efficient. Light uses would be more efficient. In between would be anyone's guess.

In other words, you don't put a bus engine on a motorcycle or a motorcycle engine on a bus and hope to be efficient. And for CPUs they have to be both a motorcycle and a bus to different people.

 

[Kaluan]

Dr. Ian Cutress saw DVLR and CBR23 Demo back on September being demoed on Raptor Lake. For reasons only known to Intel they have decided to hold back. Likely for the 13900HX to provide a good performance/Efficiency against what's a seriOus threat from AMD Zen4 Dragon Range.

View attachment 71937

Holding back? lol
13th gen is pushing the boundaries of what sensible wattage/heat a consumer x86 chip should output and you think they had the luxury of "holding back?"

This reads more like another ADL/RTL "AVX512 but not really" debacle.

Then again, maybe it's partially true, and they really needed it for the alleged RTL refresh some time next year, Meteor Lake/ARL not meeting deadlines again, so becoming a plan C (current Raptor Cove being plan B).

 

[IntelUser2000]

Heavy 24/7 turbo uses will be less efficient. Light uses would be more efficient. In between would be anyone's guess.

Alder/Raptor in light/mid load is not bad at all for efficiency. It's under high load that's efficient, which DLVR the gains might be vanishingly small. Of course for laptops it's a different story.

@Exist50 They already have FiVR. Tell me how adding a second one will help simply things? Also FiVR helps with faster transitions, which apply for all things, including undervolting/overclocking.

The biggest thing that goes against desktop having it is the leaked roadmap with DLVR only for mobile. I feel this is same as people arguing "why can't it be true?" when the only source that said anything remotely true was MLID. Except in this case it's based on misinterpretation of the patent(because of modern press which are braindead), that the 20% power reduction/or 7% clock increase will apply everywhere, all the time.

So the only reliable source(regardless how early it was) says it's a mobile feature, while everyone else arguing for desktop DLVR is likely piggybacking it on misleading and widespread information from press.

 

[Exist50]

@Exist50 They already have FiVR. Tell me how adding a second one will help simply things? Also FiVR helps with faster transitions, which apply for all things, including undervolting/overclocking.

I'm saying that DLVR replaces FIVR/brings the concept back for compute in MTL.

 

[Geddagod]

The "20% power saving or 7% clock increase" and thinking it would apply to a 200W+ 5GHz+ CPU is misunderstanding of the patent.

I'm sorry but this is a dumb differentiation to make. This applies to basically everything, performance claims don't scale well at the top. For example, the 20% performance gain of frequency from Intel 4 doesn't mean anyone thinks that MTL-S is clocking at 7.2 Ghz. Which is why I also listed a bunch of also rumored/potential performance uplifting updates that RPL-update might contain to boost performance.

The early roadmap leaks had DLVR for mobile, not desktop.

I don't think anyone here is suggesting other wise... obviously it is a mobile centric feature as power efficiency is more important in mobile. However there is no reason to believe it won't be implemented in desktop in the future. It's already in the die of RPL-S regardless, so it's essentially an added benefit that's just sitting there.

 

[Geddagod]

So the only reliable source(regardless how early it was) says it's a mobile feature, while everyone else arguing for desktop DLVR is likely piggybacking it on misleading and widespread information from press.

This is just wrong. An Asus Z790 mobo had the option of DLVR bypass, which is weird since RPL-S doesn't have DLVR. However, ASUS in-house overclocker Shamino confirmed that this feature doesn’t work because Intel fused (disabled) DLVR on Raptor Lake processors. He mentioned that they left it there in case future processors use it. And the only potential processor that could use it in the future is a rpl-refresh, since MTL uses a different mobo (though I'm not putting it past Intel that a potential RPL refresh would use a one time new mobo). And considering how DLVR is already integrated into the die of RPL itself, it would be kinda dumb to not enable it in a future refresh.
QXE, who had a 13900k ES, confirmed this by claiming it was present on the ES but disabled by the time QS were rolling out.

 

[Geddagod]

They already have FiVR. Tell me how adding a second one will help simply things? Also FiVR helps with faster transitions, which apply for all things, including undervolting/overclocking.

Are you asking how DLVR is better than FIVR?

 

[ashFTW]

@IntelUser2000 @Exist50 This slide in the MTL/ARL HotChips 34 talk is quite interesting. It points to the next disaggregated design (albeit on the server side) that the Intel teams are working towards. It’s how Falcon Shores has also been described. Clearly MTL/ARL detailed in the talk is the client disaggregated platform and Falcon Shores is the common server platform equivalent, how the future modular CPU, GPU, and xPU will be built.



I still believe Granite Rapids (and hence Sierra Forest, since they share the same platform) are being built like this. If the server design is not moved to this methodology with GNR, it will be another 2-3 years since Intel wants to do two server releases for each platform. Diamond Rapids is supposed to follow GNR on the same platform.