Discussion Intel current and future Lakes & Rapids thread

[Exist50]

If Intel 18A is that great and able to be pushed up that far great, then why waste your time with TWO short lived nodes ahead of it? ... So why not skip 20A entirely and use Intel 3 for a year?

Intel 3 will probably have a pretty long life. If nothing else, they need a reasonably mature node for their 2024 server products, and neither 20A nor 18A fit that bill, clearly. But I assume they'll at least use it for IO dies and stuff going forward, even if IFS never really picks it up.

As for 20A vs Intel 3, 20A will presumably be better for anything that can afford the schedule risk and lower yield, i.e. client, not server. Which is why they just have a single, relatively small die planned for it.

Now sure 20A can make a good "stepping stone" to 18A as far as the first node to use the new transistors and PowerVIA, but if it is only going to last six months they might have what, one product able to use it? Which would presumably be replaced/updated quickly to use 18A?

I think it's more useful to think of Intel 4 and 20A as subsets of Intel 3 and 18A respectively. So they're not really different nodes so much as they are early versions that would be expanded/refined in the successors. Given Intel's produce mix, it's probably not a terrible idea.

Though I think you do have a point in that if Intel wasn't an IDM, stuff like ARL 20A wouldn't make any sense. That appears to be an attempt to force the design teams to provide the ramp vehicle the foundry needs to get the node working. Wonder how that works with their new accounting scheme...

There is no "man, should have designed for 20A instead of Intel 18A" since they are designing for both (or more) nodes and cancelling one (or more).

I doubt they're actually designing everything for multiple nodes at once just as a backup plan. That would be extremely inefficient, and they probably don't have the staffing for it at this point.

 

[jpiniero]

I doubt they're actually designing everything for multiple nodes at once just as a backup plan. That would be extremely inefficient, and they probably don't have the staffing for it at this point.

That's actually a good argument for dropping anything they were planning on dual sourcing and only doing it at TSMC (ie: the stuff rumored to be using 20A)

 

[Exist50]

That's actually a good argument for dropping anything they were planning on dual sourcing and only doing it at TSMC (ie: the stuff rumored to be using 20A)

Presumably most of the work there is already done, and if the company as a whole is able to force them to try in the first place, there's still probably enough motive to force them to see it through. But ARL 20A clearly exists for the fabs' benefit alone. If they're to be taken seriously as a foundry, that strategy won't work going forward.

 

[Geddagod]

I doubt they're actually designing everything for multiple nodes at once just as a backup plan. That would be extremely inefficient, and they probably don't have the staffing for it at this point.

Maybe they don't develop the same arch for multiple nodes as a backup plan, but Intel appears to have 'break in case of emergency' products in development in near parallel with other products. RKL was developed a bit after ADL got into 'product definition' and RPL prob started development a bit after MTL started 'product definition' as well. Considering GNR was meant to be a 2023 product originally as well, I wouldn't be surprised if EMR development started only a bit after GNR, in case GNR gets delayed, which... it did.
Intel having contingencies like this was once voiced by Bob Swan in an earnings report as well iirc.
ARL might not have had an official 'contingency' plan, because it was already being developed on both 20A and TSMC 3nm, so if one node/development plan fails, they could just use the other one.

With the recent cutbacks, it would make sense for Intel to focus on one product at a time. Hopefully this would mean that the products they are focusing on would come out on time.

 

[jpiniero]

Presumably most of the work there is already done, and if the company as a whole is able to force them to try in the first place, there's still probably enough motive to force them to see it through. But ARL 20A clearly exists for the fabs' benefit alone. If they're to be taken seriously as a foundry, that strategy won't work going forward.

IMO, I4/I3/20A are just PR nodes. 18A is the only node that really matters. That's probably the node that's going to decide whether Intel spins off or not...

 

[Geddagod]

Just for fun : )

 

[SpudLobby]

IMO, I4/I3/20A are just PR nodes. 18A is the only node that really matters. That's probably the node that's going to decide whether Intel spins off or not...

I3 definitely matters IMO and is probably going to end up their most cost effective node in a long time. I wish they'd use it for ARL dice tbh.

 

[Exist50]

Just for fun : )

So, what's the over-under on deleting this vid and pretending it never existed? 6 months, maybe?

 

[mikk]

MLID says they couldn't get 2-way HT working and instead of delaying ARL they are launching with disabled HT. Maybe this is the reason?

 

[Exist50]

MLID says they couldn't get 2-way HT working and instead of delaying ARL they are launching with disabled HT. Maybe this is the reason?

MLID is making it all up, as usual.

 

[uzzi38]

I find it especially hilarious he's making things up just to try and "prove" that both what Igor's posted AND his own nonsense are both correct at the same time.

 

[Exist50]

I find it especially hilarious he's making things up just to try and "prove" that both what Igor's posted AND his own nonsense are both correct at the same time.

Yeah, doubling down just makes it that much funnier when he nukes it all later. Think he's going to try claiming B-step silicon didn't meet projections when he's inevitably wrong?

 

[H433x0n]

Yeah, doubling down just makes it that much funnier when he nukes it all later. Think he's going to try claiming B-step silicon didn't meet projections when he's inevitably wrong?

I really hope you’re wrong otherwise that’s pretty depressing. I was hoping to see at least a 10% ST performance improvement over RPL-R.

 

[Darkmont]

It might just be me but I never heard of B stepping silicon from him until this very video to conveniently show why his projections haven't changed.

 

[SpudLobby]

I had figured it was N3B, I didn’t know for sure. The only thing I know for sure is that internally there doesn’t seem to be any change of plans and it’s going to be ARL as the next client product. So I’m guessing this info wasn’t that concerning? I’m really unsure.

I couldn’t get any info on how the 20A version was clocking so that’s news to me.

Nah they’ll use N3E.

 

[SpudLobby]

I really hope you’re wrong otherwise that’s pretty depressing. I was hoping to see at least a 10% ST performance improvement over RPL-R.

Honestly all I care about is power consumption, their peak performance as it stands is fine. On a Single Thread basis, where the package power + DRAM power is being measured and normalized for idle, I want to see a good 40% improvement in performance iso-power for whatever power the 2-4GHz range — whatever that power figure is for RPL mobile for those ranges on an integer workload — with Arrow Lake.

E.g. take package + memory power in the 5-15W range for a Raptor Lake mobile sku (which I guess is ADL for most SKUs but rebranded on mobile) I’d like to see 30-40% more performance throughout most points in that curve iso-power from RPL to ARL mobile. Not lower power iso-performance, that’s a really low bar for these figures frankly and a single node alone could do it. 30-40% *more performance* iso-power on ST workloads in some mid frequencies.

That’s not unreasonable moving from Intel 7 to N3E which is really two node jumps, and a mature well yielding TSMC process (Apple will go for millions even on N3B) that wouldn’t be any worse than I7 — nor is it unreasonable to expect even moving to 20A. Intrinsic process transistor improvements alone should be able to buy in that range. If it falls short, then additional L3 or L2 cache (reduces DRAM movement and keeps things closer to the core) and wider or smarter architectures they can run at lower voltages will do to bridge the gap.

 

[H433x0n]

Honestly all I care about is power consumption, their peak performance as it stands is fine. On a Single Thread basis, where the package power + DRAM power is being measured and normalized for idle, I want to see a good 40% improvement in performance iso-power for whatever power the 2-4GHz range — whatever that power figure is for RPL mobile for those ranges on an integer workload — with Arrow Lake.

E.g. take package + memory power in the 5-15W range for a Raptor Lake mobile sku (which I guess is ADL for most SKUs but rebranded on mobile) I’d like to see 30-40% more performance throughout most points in that curve iso-power from RPL to ARL mobile. Not lower power iso-performance, that’s a really low bar for these figures frankly and a single node alone could do it. 30-40% *more performance* iso-power on ST workloads in some mid frequencies.

That’s not unreasonable moving from Intel 7 to N3E which is really two node jumps, and a mature well yielding TSMC process (Apple will go for millions even on N3B) that wouldn’t be any worse than I7 — nor is it unreasonable to expect even moving to 20A. Intrinsic process transistor improvements alone should be able to buy in that range. If it falls short, then additional L3 or L2 cache (reduces DRAM movement and keeps things closer to the core) and wider or smarter architectures they can run at lower voltages will do to bridge the gap.

Power scaling

I don’t want it to scale it like the 7950X does in the above article. My worry is that ARL will basically end up similar to what IgorsLab showed with meager performance improvements over RPL-R but achieving all of that performance at 120-150W. It’ll behave like Zen 4 and run into a brick wall with basically no room for overclocking or tinkering. It’ll be a boring product that has no resemblance to previous Intel products. I don’t know if I’d even bother purchasing it in that scenario.

Most of what you’re saying with regards to mobile perf/watt will be available in October. I don’t think it’ll be 40% performance aT iso power but it’ll be a full node of efficiency improvements when comparing a 13700H vs Core 7 Ultra (not a fan of the new branding).

 

[Abwx]

Power scaling

I don’t want it to scale it like the 7950X does in the above article. My worry is that ARL will basically end up similar to what IgorsLab showed with meager performance improvements over RPL-R but achieving all of that performance at 120-150W.

In these slides ARL stated perf is at a stated 250W, how could it achive those perfs at 120-150W.?..

253W for RPL and 250W for ARL :

Intel Arrow Lake-S to be 6% to 21% faster than Raptor Lake-S, Xe-LPG graphics show up to 2.4x improvement - VideoCardz.com

Intel Arrow Lake-S up to 21% faster than Raptor Lake CPU Igor’sLAB recently obtained an internal presentation from Intel revealing performance projections for the upcoming Arrow Lake-S, the next-generation client platform. Although the platform’s launch is still some time away, the discussed...

videocardz.com

 

[H433x0n]

In these slides ARL stated perf is at a stated 250W, how could it achive those perfs at 120-150W.?..

253W for RPL and 250W for ARL :

Intel Arrow Lake-S to be 6% to 21% faster than Raptor Lake-S, Xe-LPG graphics show up to 2.4x improvement - VideoCardz.com

Intel Arrow Lake-S up to 21% faster than Raptor Lake CPU Igor’sLAB recently obtained an internal presentation from Intel revealing performance projections for the upcoming Arrow Lake-S, the next-generation client platform. Although the platform’s launch is still some time away, the discussed...

videocardz.com

We’ll see when it comes out. My hunch is that it will have power scaling similar to the 7950X or the 4090. ARL-S can probably use 250W, just like the 7950X can use 230W or the 4090 can pull 600W. It just makes zero sense to run it that way since the node doesn’t scale at all.

 

[Abwx]

We’ll see when it comes out. My hunch is that it will have power scaling similar to the 7950X or the 4090. ARL-S can probably use 250W, just like the 7950X can use 230W or the 4090 can pull 600W. It just makes zero sense to run it that way since the node doesn’t scale at all.

If it use an Intel process it will likely scale like RPL, as for the 7950X scaling AT numbers are not accurate for this CPU, you can check the real scaling here for both the 7950X and RPL :

Core i9-13900K, i7-13700K & i5-13600K: Gaming-Könige im Test: Leistungsaufnahme und Effizienz

Intel Raptor Lake im Test: Leistungsaufnahme und Effizienz / Leistungsaufnahme in Anwendungen ab Werk / Der Verbrauch steigt

www.computerbase.de

Edit : The 7950X doesnt exhaust its max TDP, in Handbrake it use 189W (and 205W in CB) while the 13900K get to 250W in most benches.

 

[H433x0n]

If it use an Intel process it will likely scale like RPL, as for the 7950X scaling AT numbers are not accurate for this CPU, you can check the real scaling here for both the 7950X and RPL :

Core i9-13900K, i7-13700K & i5-13600K: Gaming-Könige im Test: Leistungsaufnahme und Effizienz

Intel Raptor Lake im Test: Leistungsaufnahme und Effizienz / Leistungsaufnahme in Anwendungen ab Werk / Der Verbrauch steigt

www.computerbase.de

Edit : The 7950X doesnt exhaust its max TDP, in Handbrake it use 189W (and 205W in CB) while the 13900K get to 250W in most benches.

It’s not planned to use an Intel process, so it’ll probably overclock horribly.

What’s wrong with the anandtech numbers? They both show similar results.

 

[Abwx]

It’s not planned to use an Intel process, so it’ll probably overclock horribly.

What’s wrong with the anandtech numbers? They both show similar results.

It s about accurate for the 13900K but not always for the 7950X, FI their 65W chart for this latter is actually 88W as demonstrated by thelr score wich match the 88W score at Computerbase.

As for DT ARL it is supposed to be fabbed with Intel 4 or something like this, so it should scale differently than a TMSC s fabbed CPU, perhaps that the mobile variant, wich require eventualy greater efficency, could be fabbed at TSMC.

 

[SpudLobby]

Power scaling

I don’t want it to scale it like the 7950X does in the above article. My worry is that ARL will basically end up similar to what IgorsLab showed with meager performance improvements over RPL-R but achieving all of that performance at 120-150W. It’ll behave like Zen 4 and run into a brick wall with basically no room for overclocking or tinkering. It’ll be a boring product that has no resemblance to previous Intel products. I don’t know if I’d even bother purchasing it in that scenario.

Most of what you’re saying with regards to mobile perf/watt will be available in October. I don’t think it’ll be 40% performance aT iso power but it’ll be a full node of efficiency improvements when comparing a 13700H vs Core 7 Ultra (not a fan of the new branding).

I (personally) don't care about desktops and these really aren't even especially granular nor do I think it's wise to extrapolate too much from them to AMD and Intel's mobile lineups other than insofar as AMD is obviously ahead.

RE: Perf/Watt: if Intel 4 lives up in the moderate frequency ranges, and those PL1 figures are interesting, then you "should" (lol) see up to 40% lower power iso-performance, but 20% higher performance (technically 20% more frequency, iso-arch) iso-power, and this is a pretty standard node jump - actually a bit larger. That's why I said provided Intel totally does basically nothing, then two real node jumps (which N3 basically would be from ADL, or 20A if it's even decent yield) should be closer to that 30-40% perf jump iso-power. Since MTL isn't bringing major arch changes with Redwood cove and I don't think any (well not massive) cache boosts, then the optimistic Intel 7 -> Intel 4 20% performance gain iso-power is largely what we have assuming MTL's packaging doesn't eat into it, and again assuming I4 really pans out that way.

Again in practice I doubt we'll see any of this at the upper ranges before someone harps on that again, but at modest clocks and wattages I do think we'll see some improvement.

 

[SpudLobby]

I should just ignore it, but I can't believe you all even bother to comment on MLID - I am assuming no one watches other than as satire. Yeah, he's an influential moron so I'm sure that gets old what with the rumors he promulgates, but you really don't need to watch to refute every last bullet point like it's a Snopes C-tier takedown.

IMO: Bepo is unintentionally funny in a way MLID could only dream of.

 

[SpudLobby]

I missed the full rumor cycle on Arrow Lake’s power and IPC, and I’m not going to bother with specifics or reading back: can someone catch me up to speed here?

Someone mentioned Raichu saying that ARL-U was a refreshed MTL. That might be an opportunity, assuming it's just not a straight up rebrand, which is very possible.

Going off of rumors, isn't ARL's GPU still 12.7-based? Sounds like they might be straight up reusing it from MTL. Or maybe a minor tweak at best (Alchemist+?). Which would be weird considering LNL seems to be on 12.9/Battlemage.

I think a more interesting question is comparing marketshare in the segments where they compete. AMD has some big holes in the lineup that PHX2 and the lower end Ryzen 7000 chips should help fill. Not a high margin market, but certainly high volume.

Honestly, an ARL-U that’s just MTL mobile with some tweaks on Intel 3 with better yield and such has potential. Sign me up for +++.