Discussion Intel current and future Lakes & Rapids thread

[uzzi38]

The exceptions are "hard IPs" like Intel's current cores and analog logic like PHYs, FIVR, etc. These are not synthesizable, and tied directly to the process. These too can be backported, but there's significantly more effort involved, and in the worst case it can resemble most of a redesign than a port. Still not impossible, however, if you're willing to throw the time, and manpower into it.

Those are kind of the biggest problem at hand with the idea of a Willow Cove backport though. The cores and analog logic surrounding it are precisely the issue. Though, to be fair, regardless of what Rocket Lake is it does appear some of that analog logic is being brought back - iirc Sharkbay did talk about IMVP9 for Rocket Lake, which has only been a thing for the Cove products thus far. Skylake and it's derivatives used IMVP8, but as for the cores on top of that I'm not sold.

In any case, time is still a bit of an issue if they started in 2019.

There would be a power penalty, surely, but again, this greatly exaggerates it. It's no secret that Cdyn scaling between nodes has slowed in recent years, and that is particularly apparent for 10nm. I imagine the penalty would be in the ballpark of 25% more power (roughly in line with the core difference between Rocket Lake and Comet Lake), not something obscene like 100%.

Not much else I can say except those were the numbers they gave as part of their own estimations. They definitely felt it would become a serious problem in the 4+GHz range anyway. Though at this point they were definitely thinking of all-core frequencies looking back at it - not ST. Which, to be fair, is a bit on me for misreading.

 

[lobz]

Obviously Intel feels like they can do it, or they wouldn't have bothered.

Now I wouldn't be surprised if they don't end up releasing it saying it's due to the virus.

I think this way it's the sentence looks more realistic, looking back on they track record on honesty this past 3-4 years.

 

[coercitiv]

Taking away his exaggeration, what’s the more reasonable performance estimate?

The numbers witeken mentioned may not be far off, based on the fact that RKL is a CML successor. Any further flagship product will need to beat 10c/20t in ST performance, MT performance, or both.

So we know RKL is 8c/16t, and we know MT performance must be at least equal to CML 10t/20t. That means RKL must somehow make up for a 20% core count deficit, which in turn requires 25% more performance per core. It's unlikely it will clock higher under heavy workloads, so we're looking at 25% IPC uplift just to match CML. And here we are, right in the middle of witeken's 20-30% IPC estimate.

Having said that, Intel needs to design the **** out of this 14nm port to make a worthwhile product against their own 2019 2020 SKUs. (in terms of MT perf)

 

[SAAA]

That: 25% better IPC is a plausible average after 5 years of wait. Golden cove would have been better but they have to keep something for the 10 or 7 nm node jump, right?.

Also If anyone here seriously believes a full backport of Willow isn't possible to do and won't reach 5GHz on 14nm +++ they are wrong. I mean every single arch Intel did worked on -two- nodes at every time:

- Sandy Bridge 32 nm then Ivy Bridge 22 nm
- Haswell 22 nm then Broadwell 14 nm
- Skylake 14 nm then… uhh... Cannonlake 10 nm

They were all mostly forward ports to a new node, with some minor changes or addition to the core itself.

Skylake on 10 nm, Cannonlake, clocked risibly and yielded even worst meanwhile on 14 nm +++ it's pulling 5.3 GHz and 10 cores dices with Comet lake. Yes, Icelake is faster, Tiger lake even more but they are on 10 nm +/++.
Is there any reason for why a backport, on a relaxed node and with less hotspots won't clock as well? It's all better except for power and area consumption, the only reason to move back to 8 cores honestly.
Think like if they backported Haswell on 32 nm: it would have clocked to 5GHz as easy as Sandy… sure AVX2 would have burned terawatts then, but on desktop who cares if you get top performance?

 

[DrMrLordX]

meanwhile on 14 nm +++ it's pulling 5.3 GHz and 10 cores dices with Comet lake.

That's not the all-core turbo for Comet Lake-S. Not even the 10900K. Care to guess how much power it'll burn doing that?

I think there's far too much optimism surrounding Rocket Lake-S.

 

[jpiniero]

The numbers witeken mentioned may not be far off, based on the fact that RKL is a CML successor. Any further flagship product will need to beat 10c/20t in ST performance, MT performance, or both.

I will say that is not entirely true. Intel could have cut back to 8 cores because of the shortage and because they decided against 10 cores in mobile. Of course the shortage should be gone because of the virus but when Intel started they wouldn't have known that obviously. It would mean a tatical retreat from DIY but it's not like they haven't already done that from not supplying the market for several months now. Need to see how big the Comet Lake-S supply ends up being.

 

[beginner99]

I will say that is not entirely true. Intel could have cut back to 8 cores because of the shortage and because they decided against 10 cores in mobile. Of course the shortage should be gone because of the virus but when Intel started they wouldn't have known that obviously. It would mean a tatical retreat from DIY but it's not like they haven't already done that from not supplying the market for several months now. Need to see how big the Comet Lake-S supply ends up being.

Yeah your reasoning does make some sense but I'm still doubtful.

That's not the all-core turbo for Comet Lake-S. Not even the 10900K. Care to guess how much power it'll burn doing that?

I think there's far too much optimism surrounding Rocket Lake-S.

Theoretically it could be golden/willow cove on 14nm without AVX512 or very high avx offset. After all intel did made a point that willow cove and forward will be more decoupled from process. That was in fall 2018 if I remember correctly. So RKL end of 2020-early 2021 would fit into such a planning rather well. Still if that is the case I wonder what took intel that long to backport. It seems a little late and I can only think that intel has limited 10nm and will use that for server and mobile only. Desktop might actually never see a 10nm product (best case scenario) or intel starts to shift desktop one node behind everything else (worst case scenario).

 

[DrMrLordX]

Theoretically it could be golden/willow cove on 14nm

Most early bets were Sunny Cove. I would be shocked if it were Golden. Willow seems like the best-case.

After all intel did made a point that willow cove and forward will be more decoupled from process.

Saying that publicly doesn't mean it's a slam dunk development-wise. Intel is struggling to launch Comet Lake-S on time. It's going to be 5 months late, and they aren't even using a new core design. That gives me little to no confidence that Rocket Lake-S will hit the streets in January 2021. If it is also 5 months late, that would put it in June 2021, which would be dangerously close to a Q4 Alder Lake-S launch. Assuming that isn't also late. If Intel can launch Alder Lake-S with something other than embarassingly low volumes in Oct. 2021, launching Rocket Lake-S in June 2021 makes almost no sense. Might as well pull the plug on the entire thing by that point.

 

[SAAA]

That's not the all-core turbo for Comet Lake-S. Not even the 10900K. Care to guess how much power it'll burn doing that?

I think there's far too much optimism surrounding Rocket Lake-S.

As long as Intel themselves binned the cores to reach 5.3 GHz each stable there's no reason to believe anyone won't just turn on MCE or overclock and let it burn 300W at 5.3 all core. If it's too hot maybe pull down to 5.2 or 5.1 but it's all about cooling then. If the same motherboards are supporting 8 cores Rocket Lake they'll be beefy enough to power it, maybe I'm wrong and it ends up with 4.5GHz all core for most but that would make it still much faster in single thread and pretty close in multi to a 10 core granddad. The single core turbo bin will tell us once we hear it: every single CPU over the last 10 years could be overclocked to match it under all core operations, if we hear of a 4.x or 5.x turbo someone will run it at that speed regardless of power.

 

[DrMrLordX]

As long as Intel themselves binned the cores to reach 5.3 GHz each stable there's no reason to believe anyone won't just turn on MCE or overclock and let it burn 300W at 5.3 all core.

Yeah that's what I figured would be the power draw. Go go custom water.

 

[witeken]

In any case, time is still a bit of an issue if they started in 2019.

The presentation that contained the Tweakers Rocket Lake roadmap dates from ~Jan/Feb'19. The roadmap itself could even be from 2018. So they likely did not start in 2019.

 

[uzzi38]

The presentation that contained the Tweakers Rocket Lake roadmap dates from ~Jan/Feb'19. The roadmap itself could even be from 2018. So they likely did not start in 2019.

That same roadmap listed 10 cores for Rocket Lake if memory serves me correctly.

Clearly it went through some changes.

 

[coercitiv]

That same roadmap listed 10 cores for Rocket Lake if memory serves me correctly.

Clearly it went through some changes.

The point was Intel was likely already talking to clients about RKL in 2018. Why start work on it later in 2019?

 

[inf64]

The point is that 14nm 8C/10C even based on a Golden Cove will not be competitive Vs 16C(+) Zen3 derivatives on 7nm+. It will be brutally bad in perf./watt if intel clocks it high enough to get the ST parity and brutally bad in just overall performance if intel clocks it low to constrain power draw issues. We are talking about a core that was targeted at 10nm or even 7nm and backporting it to 14nm. It will be HUGE on 14nm - TGL 4C with iGPU is 150mm^2 , imagine 8 TGL cores with bigger iGPU but on 14nm, yikes.

 

[Zucker2k]

The point is that 14nm 8C/10C even based on a Golden Cove will not be competitive Vs 16C(+) Zen3 derivatives on 7nm+. It will be brutally bad in perf./watt if intel clocks it high enough to get the ST parity and brutally bad in just overall performance if intel clocks it low to constrain power draw issues. We are talking about a core that was targeted at 10nm or even 7nm and backporting it to 14nm. It will be HUGE on 14nm - TGL 4C with iGPU is 150mm^2 , imagine 8 TGL cores with bigger iGPU but on 14nm, yikes.

Well, either Intel is bluffing or they've done the math and the backport checks all the relevant numbers. 14nm is not a magic node, however, if there's ever a process that could deliver, it's it. I'm just waiting patiently to see the RKL-S chip drop, and by Jove, I'd be very surprised if that 8 core chip doesn't best the 10900K in both single/multithread, as well as gaming. Otherwise, what's the use in releasing such a chip? They'd be better off sticking to the CML-S and slashing prices until next year.

 

[inf64]

Well, either Intel is bluffing or they've done the math and the backport checks all the relevant numbers. 14nm is not a magic node, however, if there's ever a process that could deliver, it's it. I'm just waiting patiently to see the RKL-S chip drop, and by Jove, I'd be very surprised if that 8 core chip doesn't best the 10900K in both single/multithread, as well as gaming. Otherwise, what's the use in releasing such a chip? They'd be better off sticking to the CML-S and slashing prices until next year.

Maybe intel figured out that RKL is just better than CML-S, who knows. I think that chances are that 8C RKL-S being better than 10C CML-S are big, BUT the chip will be huge and I'm doubtful about performance draw of such a behemoth as TGL core on 14nm with all that cache will be just huge. I guess we will have to wait and see, it wouldn't surprise me to see AMD even delay Zen3 parts to 2021 due to this virus outbreak crisis.

 

[SAAA]

The point is that 14nm 8C/10C even based on a Golden Cove will not be competitive Vs 16C(+) Zen3 derivatives on 7nm+. It will be brutally bad in perf./watt if intel clocks it high enough to get the ST parity and brutally bad in just overall performance if intel clocks it low to constrain power draw issues. We are talking about a core that was targeted at 10nm or even 7nm and backporting it to 14nm. It will be HUGE on 14nm - TGL 4C with iGPU is 150mm^2 , imagine 8 TGL cores with bigger iGPU but on 14nm, yikes.

What would work better then after Comet lake 10 core: a 12 core with same speeds or a 10 core-5.4 GHz Lava lake extreme ultimate edition?
They could choose to lose on both ST and MT performance or have a chance to get ST lead again, assuming Zen 3 doesn't manage 5GHz boosts, not to talk fixing a lot of security issues in one move with the new arch.

The iGPU will be a bit smaller if they use 32EUs on desktop rather than 96 like mobile TGL, even backported on 14 nm it would be roughly the same area as current IGPs.

7700 is 125 mm^2, 8700 is 151 mm^2 and 9900k is 177 mm^2. The soon to be released 10900k will probably fall in the 200-205 mm^2 range, that's ~13 mm^2 for each additional core on average.

Even with +40% transistors for each core a Willow cove core would end up about 18 mm^2 on 14 nm, the delta is 5 mm^2, times 8 cores makes 40 additional mm^2.

Rocket lake without additional IO and features over Comet Lake could end up between 215 to 225 mm^2.
IMHO a plausible and better choice than a 225 mm^2, 12 core Skylake with inferior ST performance.

 

[dmens]

So we know RKL is 8c/16t, and we know MT performance must be at least equal to CML 10t/20t. That means RKL must somehow make up for a 20% core count deficit, which in turn requires 25% more performance per core

I think your use of the word "must" is fundamentally flawed in this case.

 

[coercitiv]

I think your use of the word "must" is fundamentally flawed in this case.

I think you're confusing the consumer point of view with the manufacturer point of view.

 

[dmens]

I think you're confusing the consumer point of view with the manufacturer point of view.

Consumers still expect competitive products from Intel? 🤪

 

[DrMrLordX]

Well, either Intel is bluffing or they've done the math and the backport checks all the relevant numbers.

. . . or they're desperate, Rocket Lake-S will not meet expectations, and they'll mothball claiming the launch would be too close to Alder Lake-S to be justifiable.

 

[coercitiv]

Consumers still expect competitive products from Intel? 🤪

No, they just go around on forums dropping silly remarks.

 

[exquisitechar]

Well, either Intel is bluffing or they've done the math and the backport checks all the relevant numbers. 14nm is not a magic node, however, if there's ever a process that could deliver, it's it. I'm just waiting patiently to see the RKL-S chip drop, and by Jove, I'd be very surprised if that 8 core chip doesn't best the 10900K in both single/multithread, as well as gaming. Otherwise, what's the use in releasing such a chip? They'd be better off sticking to the CML-S and slashing prices until next year.

Rocket Lake-S is probably an afterthought, Rocket is more important in laptops. After all these years, 14nm (which make up the vast majority of Intel’s volume) parts get a meaningful improvement and new features, which is much needed for Intel when AMD is releasing Renoir and even more competitive parts down the line. On desktop you also get new features and better ST performance than CML-S even if it’s 8 cores, and RKL-S will be flat out superior when it comes to CPUs with 8 or less cores (most of the market). AMD is currently unbeatable when it comes >8 core “mainstream” desktop parts/MT performance, maybe Intel has just given up on that market for now.

I don’t deny the possibility of >=10900K MT performance, but I’m not convinced by this reasoning. Reading docTB’s comments on RKL, I think that people who expect a WC backport with 25% higher IPC than CML and similar clock speeds are in for a disappointment.

 

[dmens]

No, they just go around on forums dropping silly remarks.

Ya, like random claims of double digit IPC uptick.

 

[mopardude87]

Consumers still expect competitive products from Intel? 🤪

I am almost hoping Comet can compete with 3000 series, i do think the 4000 series will have people forgetting Comet in a hurry though. All i care about by end of year is having a 8c/16t chip whichever is fastest in gaming assuming a game warrants a upgrade from my current 7700k. I won't pay a huge premium for it, but if your chip is going to be hotter then Hades and be on 14nm, that thing better at least be priced competitively. I doubt it will but who knows. I got all year to wait, good things come to those who wait.