Discussion Intel current and future Lakes & Rapids thread

[nicalandia]

Funny how we've come a long way from toothpaste as TIM.

It Would seem they are using "gold" tooth paste now and even trying to patent it(not sure how they are trying to patent such an old process)




My interpretation of What a BSM Raptor Lake Might look like.

 

[nicalandia]

Intel Patent Pending application for Package-level backside metallization (bsm) Filed on 2019

US20200395307A1 - Package-level backside metallization (bsm) - Google Patents

Embodiments may relate to a microelectronic package that includes a die and a backside metallization (BSM) layer positioned on the face of the die. The BSM layer may include a feature that indicates that the BSM layer was formed on the face of the die by a masked deposition technique. Other...

patents.google.com

Ponte Vecchio using it






Sapphire Rapids too





If there was a case for a Desktop CPU to use this to enhance thermal conductivity it's Raptor Lake. So here I am hoping for that.

 

[mikk]

Meteor Lake uses LGA 2552 says MLID. Other than that nothing really new to me.

 

[uzzi38]

Meteor Lake uses LGA 2552 says MLID. Other than that nothing really new to me.

Really?

I don't really care much for MLID, but from my understanding of the screenshots posted in a Discord chat I'm in, what he's claiming for Meteor Lake and Arrow Lake absolutely is new. As least compared to what he was saying before. But then again, I really don't follow him, maybe someone here can take a look at those slides and come to their own judgement.

In fact, the Raptor Lake/Meteor Lake split he's suggesting sounds an awful lot like the Rembrandt/Cezanne split a certain someone (don't look at me, I don't want to point fingers) has been mocking for the last few months.

How curious.

 

[Exist50]

Really?

I don't really care much for MLID, but from my understanding of the screenshots posted in a Discord chat I'm in, what he's claiming for Meteor Lake and Arrow Lake absolutely is new. As least compared to what he was saying before. But then again, I really don't follow him, maybe someone here can take a look at those slides and come to their own judgement.

In fact, the Raptor Lake/Meteor Lake split he's suggesting sounds an awful lot like the Rembrandt/Cezanne split a certain someone (don't look at me, I don't want to point fingers) has been mocking for the last few months.

How curious.

Care to post them here as well? Save us the trouble of wading through.

 

[mikk]

What is new? Arrow Lake on Lion Cove+Skymont, 20A and 8+32 isn't new, although he confirmed the older leaks. Meteor Lake using Redwood Cove+Crestmont isn't new either, IPC estimates are too vague at the moment. To me it isn't new but of course not every is aware of all leaks. LGA socket number is definitely new and he even posted a picture which once again shows he has real sources (when it comes to Intel). I haven't seen the screenshot from discord, I cannot comment on this.

 

[uzzi38]

Care to post them here as well? Save us the trouble of wading through.

I would were I not on mobile and stuck posting poor resolution images (limitation of the app I'm using, both horitontal and vertical resolution of images are severely cut for no reason other than plz buy membership). Unless you guys are fine with a link instead?

https://media.discordapp.net/attachments/786263341270040639/982414596142952508/unknown.png

Here's also the Raptor Lake one, but not sure how much is new here at all. Enjoy taking potshots at it though

https://media.discordapp.net/attachments/786263341270040639/982414595815792650/unknown.png

 

[uzzi38]

What is new? Arrow Lake on Lion Cove+Skymont, 20A and 8+32 isn't new, although he confirmed the older leaks. Meteor Lake using Redwood Cove+Crestmont isn't new either, IPC estimates are too vague at the moment. To me it isn't new but of course not every is aware of all leaks. LGA socket number is definitely new and he even posted a picture which once again shows he has real sources (when it comes to Intel). I haven't seen the screenshot from discord, I cannot comment on this.

Q2/Q3 for mobile Meteor Lake with Q4 for desktop Meteor Lake? (Which puts it more in Zen 5 competition territory than Zen 4 competition territory)

Mobile MTL being reserved for high end only with mainstream getting Raptor Lake?

Arrow Lake landing in 2H 2024 rather then 1H like I'm pretty sure was the expectation till date (although the comment on Zen 6 competition is definitely premature given how little is actually known about Zen 5, and trust me, very little that's actually correct is known about Zen 5).

You shouldn't need to comment on any image from Discord, because it's one of the slides in this video you supposedly watched.

 

[moinmoin]

So aside of a 10 cores glitch with Comet Lake Intel is stagnating with 8 P-cores for years to come and hoping people won't notice because eventually 32 E-cores?

 

[mikk]

Q2/Q3 for mobile with Q4 for desktop? (Which puts it more in Zen 5 competition territory than Zen 4 competition territory)

Mobile MTL being reserved for high end only with mainstream getting Raptor Lake?

Arrow Lake landing in 2H 2024 rather then 1H like I'm pretty sure was the expectation till date (although the comment on Zen 6 competition is definitely premature given how little is actually known about Zen 5, and trust me, very little that's actually correct is known about Zen 5).

You shouldn't need to comment on any image from Discord, because it's one of the slides in this video you supposedly watched.

For me it isn't new. MTL for desktop never was realistic before late 2023 or even 2024, I'm sceptical about 2023 to be honest. Meteor Lake mobile was late Q1/early Q2 last year but was doubtful from the beginning, it definitely might come in mid 2023 and I think we should differentiate between MTL-M and MTL-P. MTL-M 2+8 should come out quite a bit earlier than MTL-P 6+8 and if MTL-M comes mid 2023 I'm not sure where MTL-P lands.

He isn't sure about the ARL release, it's written in white letters. He expects it will come in H2 2024 which seems logical for him considering he expects MTL mainly in H2 2023+. It's the same with the assumed Zen 6 competitor in H2 2024, so this part isn't really meaningful at this point. However it's unlikely that ARL is ready in H1 2024 because Intel earlier this year told 20A is ready for manufacturing in H1 2024 and therefore it was unlikely that ARL could hit the market in H1 2024. The technical stuff in this video isn't new apart from the LGA 2552.

 

[Henry swagger]

Higher R&D doesn't speed up development, unless you are severely lacking in such. A competent team does.

Most of Intel is not related to CPUs. They have the process technology team, they have massive amount of software engineers(more than VMWare), and they make networking devices. They are still involved in making motherboards for SFF devices like NUC, and for servers. Don't forget about the research team either.

They are also actively involved in development of PCs in all areas. AGP, PCI Express, USB standards are all from them.

Also, they go above and beyond supporting new designs. When companies like GPD first start out, they will support them in all areas of design, such as making a new motherboard. It's very comprehensive, taking you almost step by step even if you are a tiny startup. They certainly take a very long term view when it comes to development of the PC.

An extreme example: Contra revenue wasn't just about giving them serious discounts to the manufacturers. They took this customer support for developing a new system to a whole new level. They essentially made an entire Tablet of countless variations and pitched it to anyone interested for pretty much no money.

That's why you had so many no name vendors popping up with these things. You could take any of Intel's designs and just rebrand them, and they'd do almost everything for you. No wonder it cost them so much!

Well said.. i was massively impressed with ponte vechio using fevoros it looka so advanced..

 

[Exist50]

I would were I not on mobile and stuck posting poor resolution images (limitation of the app I'm using, both horitontal and vertical resolution of images are severely cut for no reason other than plz buy membership). Unless you guys are fine with a link instead?

https://media.discordapp.net/attachments/786263341270040639/982414596142952508/unknown.png

Here's also the Raptor Lake one, but not sure how much is new here at all. Enjoy taking potshots at it though

https://media.discordapp.net/attachments/786263341270040639/982414595815792650/unknown.png

Thanks. Was on mobile at the time. Didn't want to scrub through a video. Or give it a view, tbh. And quality is plenty fine.

So I'll just give my take item by item.

MTL:

• Socket - No clue. I thought MTL/ARL would use LGA 1800 1700 (Edit: Typo?), but I also knew they would not be platform compatible with ADL/RPL. Could see this being true, but little difference either way. Though a new socket typically hurts motherboard prices.
  
  
• IPC - Bull. Maybe bigger than Raptor Lake, but 12-21%? Nah, he's just making stuff up.
  
  
• Clock speed regressions - Probably bull. Maybe they'll lose 100-200MHz or so, but large enough to compare to ICL? Nah. I fully expect the node shrink to make up any regression from arch/design, and personally guess that clocks will ultimately be higher between comparable SKUs.
  
  
• VPU - sure
  
  
• 2+8, 6+8 - sure, 8+16 - no
  
  
• RPL/MTL volume split - Not sure, but certainly suspicious.
  
  
• Timing - Sounds reasonable enough.

ARL:

• 8+32 on 20A - Think so?
  
  
• LNC IPC - I'm thinking comparable-ish to Golden Cove's gains. Expecting >>GLC gains (like his previous "at least 30%" claim) is just nonsense. But make no mistake, LNC is probably the most important evolution of Core since its inception. Much better in a whole host of ways.
  
  
• Lion Cove is not Royal. Royal is Royal. How hard is this to understand? Clearly no clue what he's talking about.
  
  
• Skymont - We're in for a treat with this one.
  
  
• Timing - Sounds reasonable enough.

In short, I think all the "new", important details range from suspect to nonsensical, and the rest just reiterating well established rumors.

 

[igor_kavinski]

LGA1800 could allow Raptor Lake to take advantage of DLVR. Put it in LGA1700 and you lose this functionality.

 

[Exist50]

LGA1800 could allow Raptor Lake to take advantage of DLVR. Put it in LGA1700 and you lose this functionality.

I guess that's possible. Though it seems like a lot of effort for a fairly minor feature (for the target market) in a refresh gen. Would be cool to see how that plays out though. You'd think DLVR would mean less pins needed.

 

[igor_kavinski]

You'd think DLVR would mean less pins needed.

I think the extra 100 pins ensure no one tries to ram ADL into a Z790/B760/H710 mobo.

 

[Exist50]

I think the extra 100 pins ensure no one tries to ram ADL into a Z790/B760/H710 mobo.

But...why would that be a problem? Idk, but breaking socket compatibility, increasing costs of new motherboards, etc. It all just seems like so much work for so little gain.

And what about the low end? And they really going to use 6+8 to supply the entire mainstream market? Feels like the ability to reuse ADL's 6+0 die could be handy.

 

[LightningZ71]

Have we legitimately seen a DESKTOP consumer grade application that would really benefit from having more than 8 P cores as opposed to having 4 times as many additional E cores? There are precious few apps out there that are threaded enough to go beyond 8 threads without savage diminishing returns, and almost everything that is more highly threaded (embarrassingly parallel by nature) takes better advantage of dedicated cores than SMT threads.

The only thing that's preserving the 5900/5950x leads over the 12700/12900k in highly threaded applications is the significantly higher per-thread L3 cache resources that they have to better enable their SMT to give better overall performance uplift over not having it. We can see in the few apps that don't need that that their performance is much closer.

None of this is to say that ZEN4 may not change this, just that E cores certainly have their uses, and that more P cores won't necessarily be faster.

 

[igor_kavinski]

But...why would that be a problem? Idk, but breaking socket compatibility, increasing costs of new motherboards, etc. It all just seems like so much work for so little gain.

Maybe the bendgate issue is also fixed in LGA1800.

 

[Exist50]

Maybe the bendgate issue is also fixed in LGA1800.

I doubt Intel really cares about a few degrees difference for enthusiast watercooling. Maybe if there were compelling enough reasons for a socket change otherwise, they might look into the mounting hardware, but again, think it's all just too minor. Hell, would probably be cheaper to ship one of those bracket things with every high end board than retool for a whole new socket.

 

[IntelUser2000]

Skymont - We're in for a treat with this one.

So aside of a 10 cores glitch with Comet Lake Intel is stagnating with 8 P-cores for years to come and hoping people won't notice because eventually 32 E-cores?

So I think the gap between E and the P cores will further close, which will serve to reduce the scenarios that result in unstable or jittery experience.

This is a big part of the reason it took them until now to get the hybrid config.
P vs E

Haswell vs Airmont - 137%(or 2.37x) faster per clock, plus lower clocks
Skylake vs Goldmont Plus - 70%
Willow Cove vs Tremont - 60-65%
Golden Cove vs Gracemont - 50% per clock

I wouldn't be surprised if they are aiming just a 30% gap.

You'd think DLVR would mean less pins needed.

I don't see how it'll result in less. Remember I said DLVR is a separate, alternate power phase that kicks in when the load goes higher. So maybe they can save a few pins on lower current, but you need extra for concurrent current delivery.

DLVR won't matter for high end desktop anyway. Who would care?

Q2/Q3 for mobile Meteor Lake with Q4 for desktop Meteor Lake? (Which puts it more in Zen 5 competition territory than Zen 4 competition territory)

I don't know how you can have Raptorlake mobile launch in Q1, and Meteorlake in Q2. It has to be products that has no presence, like MTL-M. Even Q4 is pushing it. Just a Q4 mobile launch would need Ice/Comet like mix, nevermind two quarters earlier!

Since Exist50 is saying Meteorlake won't have much mobile presence, they can technically speed up Arrowlake to come before Q4.

 

[Exist50]

So I think the gap between E and the P cores will further close, which will serve to reduce the scenarios that result in unstable or jittery experience.

This is a big part of the reason it took them until now to get the hybrid config.
P vs E

Haswell vs Airmont - 137%(or 2.37x) faster per clock, plus lower clocks
Skylake vs Goldmont Plus - 70%
Willow Cove vs Tremont - 60-65%
Golden Cove vs Gracemont - 50% per clock

I wouldn't be surprised if they are aiming just a 30% gap.

Huh, I'm kinda looking at this the opposite way. The smaller the gap between them, the less reason there is to have two separate lines at all. If, for example, Atom was within 10% of Core at 1/3 the area, then there'd be little reason to offer Core in the server market. Likewise, you increasingly stray away from a big.little type of arrangement.

Feel like it would be better for each to remain uniquely specialized. Efficiency and/or area for Atom, single core performance for Core. Though maybe it wouldn't hurt to unify the ISA...

Like, I've heard very good things about Skymont, but I'm not sure exactly how good, especially relative to Lion Cove. But I wouldn't be surprised if they were under that 30% IPC gap of yours.

Since Exist50 is saying Meteorlake won't have much mobile presence, they can technically speed up Arrowlake to come before Q4.

Wait, hold on. I said what now? I'm suspicious of MLID's claim that there will be an ICL/CML-like split, if anything. Or did you mean desktop?

 

[Exist50]

I don't know how you can have Raptorlake mobile launch in Q1, and Meteorlake in Q2. It has to be products that has no presence, like MTL-M. Even Q4 is pushing it. Just a Q4 mobile launch would need Ice/Comet like mix, nevermind two quarters earlier!

Oh, missed this. I think Q3 seems far more likely than Q2 for actual products, but that's more of a guess than a prediction. But at the end of the day, I think it would be good for Intel if they could quickly phase out an uncompetitive lineup (Raptor Lake mobile). Pheonix is going to destroy it if they go head to head, so the sooner Meteor Lake comes, the better. Better for Intel 4 ramp as well. The question is whether it's actually ready on time.

 

[IntelUser2000]

Wait, hold on. I said what now? I'm suspicious of MLID's claim that there will be an ICL/CML-like split, if anything. Or did you mean desktop?

Yes desktop haha.

But at the end of the day, I think it would be good for Intel if they could quickly phase out an uncompetitive lineup (Raptor Lake mobile)

I don't think the CPU comparisons will be bad, at least performance. The GPU is a problem.

But it doesn't really matter if Raptorlake mobile is having a comprehensive lineup. Then Meteorlake mobile is going to come 12 month later, period.

You still have Tigerlake products coming out. The cheap Aya NEO and AYN products are Alderlake devices coming in late 2022. Manufacturers are simply going to skip either if one comes much earlier than 12 months.

It makes no sense investing all that money designing, optimizing, and making a laptop* just to replace them in 6 months. Even 9 months doesn't happen. They have about 2 months gap between product announcement and delivery for laptops. For desktops you can buy it the day it's launched.

You have Alderlake mobile coming now. Actually it's been about a month but you get my point. And say you are launching Raptorlake next month. Zero sense for laptops.

*The optimization in laptops are far, far greater than what desktops guys can imagine. Lots of the laptops until recently couldn't support C10 properly because of driver/firmware/BIOS limitations. It's just hard, because if you don't do it right, you might result in an unstable system. BIOS programming, driver programming, OS settings, circuit design, all have to come seamlessly in order to create the post-Haswell efficient laptops today. It's not "Oh just connect them together and done. Voila!" like in desktops. That's why drivers are heavily customized in laptops and not just downloads from hardware guys.

Feel like it would be better for each to remain uniquely specialized. Efficiency and/or area for Atom, single core performance for Core. Though maybe it wouldn't hurt to unify the ISA...

So bear in mind I am pro-hybrid approach. That said,

It's always a tradeoff. So apples-to-apples the square root law says 30% performance needs a core that's 70% larger. Also the single core focused design will have higher frequencies, so that'll result in the core larger as well.

Even if the E core is far superior in both perf/watt and perf/mm2 metric, a core that's better in single thread will be very valued.

Also in theory, the hybrid approach allows oversizing the P core to be larger and more performant than it would be otherwise. You don't have many cores, thus you can make it bigger than if you are trying to balance max MT and max ST performance. I think this part will take several generations to be realized, and part of the reason for the rapid growth for ARM cores.

Rather than:
-16x P with 1x area/perf

You'd have
-8x P with 1.15x perf at 1.5x the die area, and 32x E cores

 

[Hougy]

Thanks. Was on mobile at the time. Didn't want to scrub through a video. Or give it a view, tbh. And quality is plenty fine.

So I'll just give my take item by item.

MTL:

• Socket - No clue. I thought MTL/ARL would use LGA 1800 1700 (Edit: Typo?), but I also knew they would not be platform compatible with ADL/RPL. Could see this being true, but little difference either way. Though a new socket typically hurts motherboard prices.
  
  
• IPC - Bull. Maybe bigger than Raptor Lake, but 12-21%? Nah, he's just making stuff up.
  
  
• Clock speed regressions - Probably bull. Maybe they'll lose 100-200MHz or so, but large enough to compare to ICL? Nah. I fully expect the node shrink to make up any regression from arch/design, and personally guess that clocks will ultimately be higher between comparable SKUs.
  
  
• VPU - sure
  
  
• 2+8, 6+8 - sure, 8+16 - no
  
  
• RPL/MTL volume split - Not sure, but certainly suspicious.
  
  
• Timing - Sounds reasonable enough.

ARL:

• 8+32 on 20A - Think so?
  
  
• LNC IPC - I'm thinking comparable-ish to Golden Cove's gains. Expecting >>GLC gains (like his previous "at least 30%" claim) is just nonsense. But make no mistake, LNC is probably the most important evolution of Core since its inception. Much better in a whole host of ways.
  
  
• Lion Cove is not Royal. Royal is Royal. How hard is this to understand? Clearly no clue what he's talking about.
  
  
• Skymont - We're in for a treat with this one.
  
  
• Timing - Sounds reasonable enough.

In short, I think all the "new", important details range from suspect to nonsensical, and the rest just reiterating well established rumors.

12-21% IPC increase for MTL over RPL? Locuza and Semianalysis already did a great die shot analysis of MTL. It seems there were few changes to the core, so doesn't that make this IPC prediction completely impossible?

I've seen MLID be proved wrong many times before, but never as soon as he made the prediction 😆

 

[IntelUser2000]

12-21% IPC increase for MTL over RPL? Locuza and Semianalysis already did a great die shot analysis of MTL. It seems there were few changes to the core, so doesn't that make this IPC prediction completely impossible?

I've seen MLID be proved wrong many times before, but never as soon as he made the prediction 😆

It's not small. Here's what Raichu said:

About Meteor lake. MTL focus on how to improve the efficiency of the instruction execution, it will not widen the microarchitecture crazy like Alder lake.

More improvements maybe will focus on branch prediction, Micro-operation fusion, instruction dispatch, register remake, and EU execution efficiency.

Some are expecting Haswell-like gains, which is low 10%. 12% seems to be a good average. Haswell was also efficiency focused, so some have said they reduced potential performance to lower power.

Maybe they are doing small/Medium gains rather than zero/Large gains they are doing now. Sunny = Large, Willow = tiny/zero, Golden = Large

to,

Raptor = small, Redwood = Medium, etc. Which would be better for planning I guess?