Discussion Intel current and future Lakes & Rapids thread

[MoogleW]

The irony will be when Intel launch ARL, you ll see that they ll advertise huge numbers in comparison with RPL perf/watt wise, and yet it wont be more efficient than the 7950X...

I mean, at least GN Steve won't make fun of Intel for 'releasing an AMD CPU' again like when Intel compared against Ryzen in their tigerlake launch

 

[DrMrLordX]

No foundry just gets rid of outdated production nodes.

There are countless fabless companies that work with older nodes which become much cheaper as the newer, cutting edge nodes take prominence.

Case in point Raspberry Pi, who waited until 2019 to pull the trigger on their 28nm design RPi4 SoC, when 7nm was already in production for sometime.

Some customers simply cannot afford state of the art fab node prices.

In the case of Intel 7 and Intel 4, those nodes likely won't be used by foundry customers. Intel 3 and Intel 18a? Far more likely.

 

[soresu]

In the case of Intel 7 and Intel 4, those nodes likely won't be used by foundry customers. Intel 3 and Intel 18a? Far more likely.

Price is key - Intel are not going to leave fabs to rot if they can charge enough to still get a profit while still charging less than something more advanced.

 

[SiliconFly]

When it comes to Intel 20A, what bothers me the most is the max boost clocks (fmax). Intel 4 has only a HP library optimized for power-efficiency which limits MTL's max turbo frequency (which is kinda okay for mobile cpu cos it provides increased power-efficiency).

But Intel 20A has a ppw gain of 36% over Intel 4. Meaning, the cells might be even shorter (vertically). So, is that going to have a crippling effect on ARL's max turbo frequency. I don't think it'll severely impact ARL mobile parts. But what happens to the ARL desktop parts if 20A has poor fmax?

So the real question is, is ARL desktop parts gonna have clock regression & hence performance regression? I know, hitting 6.0 GHz is extremely difficult considering Intel 20A is brand new. But with a smaller cell, it might make things even more worse. I wish Intel share more info about 20A. Any ideas?

 

[DrMrLordX]

Price is key - Intel are not going to leave fabs to rot if they can charge enough to still get a profit while still charging less than something more advanced.

Traditionally they've used older nodes for things like chipset production. I strongly suspect that Intel 4 capacity can be converted to Intel 3 capacity, though I could be wrong.

 

[SiliconFly]

Traditionally they've used older nodes for things like chipset production. I strongly suspect that Intel 4 capacity can be converted to Intel 3 capacity, though I could be wrong.

Even I think Intel 4 will be absorbed into Intel 3 when it's time.

 

[ondma]

It should have the rentable cores feature which could be a superior alternative if Intel engineers have done their homework.

I thought neither of them (hyperthreading/rentable cores) were going to be working for ARL??? I guess it will have to be a big unknown until we get more preliminary results or see the actual product.

 

[SiliconFly]

I thought neither of them (hyperthreading/rentable cores) were going to be working for ARL??? I guess it will have to be a big unknown until we get more preliminary results or see the actual product.

As of now, rumors suggest no hyper-threading and no rentable-cores for ARL. But they're just unconfirmed rumors at this point. Would be interesting if Intel releases any ARL info at the next week event. Or we may have to wait for some hotchips announcements at a later date.

 

[Hitman928]

When it comes to Intel 20A, what bothers me the most is the max boost clocks (fmax). Intel 4 has only a HP library optimized for power-efficiency which limits MTL's max turbo frequency (which is kinda okay for mobile cpu cos it provides increased power-efficiency).

But Intel 20A has a ppw gain of 36% over Intel 4. Meaning, the cells might be even shorter (vertically). So, is that going to have a crippling effect on ARL's max turbo frequency. I don't think it'll severely impact ARL mobile parts. But what happens to the ARL desktop parts if 20A has poor fmax?

So the real question is, is ARL desktop parts gonna have clock regression & hence performance regression? I know, hitting 6.0 GHz is extremely difficult considering Intel 20A is brand new. But with a smaller cell, it might make things even more worse. I wish Intel share more info about 20A. Any ideas?

HD libraries are optimized for density and efficiency, not HP. Traditionally, I believe Intel also offered a UHP library for performance at all costs type of design but it looks like they might be trimming down to just HD libraries for denser/slower/more efficient cells and HP libraries for less dense/faster/less efficient cells.

 

[A///]

I'm not so confident on Intel rapidly converting older nodes... we know what happened last time they went hand over fist like that.

 

[Geddagod]

HD libraries are optimized for density and efficiency, not HP. Traditionally, I believe Intel also offered a UHP library for performance at all costs type of design but it looks like they might be trimming down to just HD libraries for denser/slower/more efficient cells and HP libraries for less dense/faster/less efficient cells.

GLC/RPC was only UHP. RWC is only HP. Zen 4 is (iirc relaxed?) HD as the standard cell with HP mixed in. LNC honestly might be the first CPU core since PLMC from Intel to return to using a majority HD design (on TSMC). Funnily enough, LNC looks to be facing issues just like PLMC did.

 

[Geddagod]

As of now, rumors suggest no hyper-threading and no rentable-cores for ARL. But they're just unconfirmed rumors at this point. Would be interesting if Intel releases any ARL info at the next week event. Or we may have to wait for some hotchips announcements at a later date.

Wouldn't be surprised if we see some die shots of ARL in a week on 20A.
We might get an Intel architecture day return on August 2024 (so like 2 months before ARL launch prob) where they give us more specific details about LNC and SMK. Until then, I expect Intel to be very vague on the exact details abt LNC. At best they repeat their GLC mantra- "wider, deeper, smarter".
We also would most likely get a Intel client roadmap update next week. ARL might (finally) get revealed to use TSMC 3nm as well. We would also prob be able to confirm the rumors about the cancellation of PTL and perhaps what node(s) the gen after ARL uses.
Intel Innovation events are always ridiculously fun due to the sheer amount of information Intel drops during those events. Very excited.

 

[A///]

Wouldn't be surprised if we see some die shots of ARL in a week on 20A.
We might get an Intel architecture day return on August 2024 (so like 2 months before ARL launch prob) where they give us more specific details about LNC and SMK. Until then, I expect Intel to be very vague on the exact details abt LNC. At best they repeat their GLC mantra- "wider, deeper, smarter".
We also would most likely get a Intel client roadmap update next week. ARL might (finally) get revealed to use TSMC 3nm as well. We would also prob be able to confirm the rumors about the cancellation of PTL and perhaps what node(s) the gen after ARL uses.
Intel Innovation events are always ridiculously fun due to the sheer amount of information Intel drops during those events. Very excited.

We'll find out more about arrow around the time Zen 5 launches I'd bet. Intel is very good at releasing news based on what they know AMD will be announcing or releasing to plant a seed of doubt or curiousity about a competitor product na cause people to hold off. we'll see die shots sooner as you say but the leaks will get heavy around q2 which is when most people expect z5 to release. I can see intel releasing info to leakers steadily unless the plan to sideswipe amd and release it earlier in q3 just to spite amd. I've got some blind faith in intel that makes me think they'll try and release arrow in June or July. The perfect window for people to buy zen 5 because the first few weeks or months will make it hard for people to get zen 5 for msrp if it's very good. I really wish there were more performance leaks on arrow lake because what we have seen so far is not hopeful but there's no details on those and it's long since been finished and sent out. The one leak was so sketchy I'm willing to ignore it.

really hope to see arrow be a cooler chip than what we've gotten the last severan gens from intel. maybe some oc room in case Pat gunslinger gelsinger reads here.

 

[igor_kavinski]

Maybe the really, really early samples of ARL don't have rentable cores enabled and they are working out some kinks. That could be the source of the rumors.

 

[SiliconFly]

HD libraries are optimized for density and efficiency, not HP. Traditionally, I believe Intel also offered a UHP library for performance at all costs type of design but it looks like they might be trimming down to just HD libraries for denser/slower/more efficient cells and HP libraries for less dense/faster/less efficient cells.

I'm sure Intel 3's HD library packs more density. But, it was Pat who mentioned that Intel 4's HP cell library is optimized for power-efficiency rather that performance (maybe not as much as HD, but still). Thats probably the reason MTL clocks are capped at 5.1 GHz. I remember reading that the cell height was reduced to improve efficiency but at the cost of higher boost clocks. Intel generally uses HP cells for their client cpus. But that may change in the future.

 

[A///]

Depending on the core design they may not need higher clocks to achieve good performance. Frequency regression is one issue, but it doesn't necessarily explain lower performance. Really goes back to that royal core biz buzz bazz people were discussing with such fervor.

 

[DrMrLordX]

Even I think Intel 4 will be absorbed into Intel 3 when it's time.

That's odd phraseology. It does seem likely that they'll do fab conversion where possible.

 

[A///]

That's odd phraseology. It does seem likely that they'll do fab conversion where possible.

I can see it but only if the yields down the line are good enough. I may be wrong here given it's been almost 9 years but I recall intel having to retool fabs to make more 14nm processors because 10nm wasn'tquite panning out with yields. although intel did come out and state their problems were because of how aggressive they were being and learned their lesson. ofc this was years later with their own admittance and it seems to be fine now. for the time being. but no one has leaked any internal reports for them or tsmc, for example. just a lot of "theory" aka hooey.

 

[SiliconFly]

That's odd phraseology. It does seem likely that they'll do fab conversion where possible.

I think Intel 4 & Intel 3 HP libraries follow the same design rules (with minimal porting). And hence should make things easy for Intel to migrate the process/equipments from 4 to 3.

 

[SiliconFly]

I can see it but only if the yields down the line are good enough. I may be wrong here given it's been almost 9 years but I recall intel having to retool fabs to make more 14nm processors because 10nm wasn'tquite panning out with yields. although intel did come out and state their problems were because of how aggressive they were being and learned their lesson. ofc this was years later with their own admittance and it seems to be fine now. for the time being. but no one has leaked any internal reports for them or tsmc, for example. just a lot of "theory" aka hooey.

Lets hope that at least future processors don't have to run at ridiculous clocks for higher performance.

 

[A///]

Lets hope that at least future processors don't have to run at ridiculous clocks for higher performance.

Highly agree.

 

[SiliconFly]

New MTL geekbench leak along with 13700H for comparison.

MTL Core Ultra 7 1002H: Single = 2439, Multi = 12668
i7-13700H: Single = 2531, Multi = 13360

(It's an ES. So, the final may be higher.)

Its closer to a 13700H, but hopefully with a lot more power-efficiency! 7 more day...

 

[igor_kavinski]

Its closer to a 13700H, but hopefully with a lot more power-efficiency! 7 more day...

Great if it's running at 25W. Not so if it's 35 or 45W.

 

[SiliconFly]

Great if it's running at 25W. Not so if it's 35 or 45W.

Yep. Thats a bit of a worry.

 

[cebri1]

Great if it's running at 25W. Not so if it's 35 or 45W.

Are you talking TDP or real power consumption when being tested? Because the 13700H has a PL2 of 120W. If it can meet 13700H max perf at 80W it would be already a significant step forward for Intel.