Discussion Intel current and future Lakes & Rapids thread

[uzzi38]

Not looking good. I understand some resigned over at Intel over this back-porting issue, and it's taken them about 2-3 years to come to the same realization. On the bright side, Tigerlake brings easily 25 - 30% performance boost over Skylake so that fight with Zen 3 should be interesting at the same core-count, but then there's going to be a 16 core R9 4950x to contend with.

Forget even that, going up against same core count will be difficult. Zen 3's no slouch - not even slightly so - but the difference in power draw and clocks will be...interesting.

Look at perf/W of ICL-U vs WHL-U. Of course, TGL-U will be a noticable improvement over that, but Intel will be putting cores that wide on 14nm.

Heck... AMD might have the frequency advantage with Zen 3. Just think about that for a second.

 

[exquisitechar]

Not looking good. I understand some resigned over at Intel over this back-porting issue, and it's taken them about 2-3 years to come to the same realization. On the bright side, Tigerlake brings easily 25 - 30% performance boost over Skylake so that fight with Zen 3 should be interesting at the same core-count, but then there's going to be a 16 core R9 4950x to contend with.

Clock speeds will go down the drain if this is true, there will be no 30% performance gain over Coffee Lake. I don't think the performance per clock would be quite as high as TGL's either.

 

[mikk]

I don't think the performance per clock would be quite as high as TGL's either.

Why should the IPC lower when both are Willow Cove based? I doubt the desktop version is slower, it doesn't make sense. Usually the mainstream architecture is exactly the same between the ULV and desktop. IPC should be the same and Willow Cove on desktop probably clocks higher on 14nm than Tigerlake on 10nm.

 

[uzzi38]

Why should the IPC lower when both are Willow Cove based? I doubt the desktop version is slower, it doesn't make sense. Usually the mainstream architecture is exactly the same between the ULV and desktop. IPC should be the same and Willow Cove on desktop probably clocks higher on 14nm than Tigerlake on 10nm.

Do you think they could do the full cache size on 14nm?

Not to mention the size of the full Willow Cove core on 14nm would be huge as well.

Of course they'll be cutting the core down.

 

[TheGiant]

I am trying to find icelake SMT coefficient, anyone ?
based on cinebench with relative scores..

so the i7 1065G7 should have 3,9GHz ST max freq - both 25 and 15W have the same score, looks ok
so 430/3,9=110,25 points per GHz
at 4C /8T and 1,9GHz freq (15W, in the article) it has 614,7points per GHz with 4C and 153,7 per 1C per GHz
so 153,7/110,25= 1,39 and that means 39% SMT

assuming 9900k 5GHz score of 509 points single core that means 101,8 points per GHz
so that with 2,3GHz 10210U means with 1138/2,3=493 points per GHz with 4C and thats 123,7 points per GHz 1C2T- 123,7/101,8 =1,22 so 22% SMT

if I am not wrong, Icelake has much better SMT coeficient

any reasons why is that? except wider core?

 

[BigDaveX]

Do you think they could do the full cache size on 14nm?

Not to mention the size of the full Willow Cove core on 14nm would be huge as well.

Of course they'll be cutting the core down.

Why would they not be able to make it on 14nm? They make HEDT chips with 1MB L2 cache/core on that process, don't they? The only trouble they would have is if the chip ended up so big they couldn't physically fit it into the package, and even that seems doubtful judging by the delidded shots of Coffee Lake 8c (and that's before you get into the possibility that the rumoured LGA1200 will be a bigger package).

 

[Zucker2k]

Why would they not be able to make it on 14nm? They make HEDT chips with 1MB L2 cache/core on that process, don't they? The only trouble they would have is if the chip ended up so big they couldn't physically fit it into the package, and even that seems doubtful judging by the delidded shots of Coffee Lake 8c (and that's before you get into the possibility that the rumoured LGA1200 will be a bigger package).

Plus they trimmed the igp (IGD)

 

[uzzi38]

Why would they not be able to make it on 14nm? They make HEDT chips with 1MB L2 cache/core on that process, don't they? The only trouble they would have is if the chip ended up so big they couldn't physically fit it into the package, and even that seems doubtful judging by the delidded shots of Coffee Lake 8c (and that's before you get into the possibility that the rumoured LGA1200 will be a bigger package).

Sunny Cove cores have 50% more transistors per core than Skylake.

Willow Cover further significantly increases cache size.

Now try putting that on 14nm.

 

[Cardyak]

Clock speeds will go down the drain if this is true, there will be no 30% performance gain over Coffee Lake. I don't think the performance per clock would be quite as high as TGL's either.

If Willow Cove features an 11% IPC increase over Sunny Cove then that will result in ~30% IPC increase over Skylake. (My 11% guess for Willow Cove is complete conjecture by the way, I have no evidence of this. Just making assumptions based on previous architectures and trying to piece together leaks)

However to a degree you are correct, clock speeds may take a hit, so let's take a rough estimate of clock speeds regressing ~10% as an example. Then the overall performance increase will still be in the region of 20% - Still a vast improvement.

 

[uzzi38]

If Willow Cove features an 11% IPC increase over Sunny Cove then that will result in ~30% IPC increase over Skylake. (My 11% guess for Willow Cove is complete conjecture by the way, I have no evidence of this. Just making assumptions based on previous architectures and trying to piece together leaks)

However to a degree you are correct, clock speeds may take a hit, so let's take a rough estimate of clock speeds regressing ~10% as an example. Then the overall performance increase will still be in the region of 20% - Still a vast improvement.

You're underestimating the amount clocks will drop by. Pretty heavily.

 

[Zucker2k]

You're underestimating the amount clocks will drop by. Pretty heavily.

It goes without saying that Intel has done the math and found it worthwhile, even with the upcoming 10900K in play. That 14nm process is still showing quite significant improvement with every iteration. That is the story here.

 

[BigDaveX]

Sunny Cove cores have 50% more transistors per core than Skylake.

Willow Cover further significantly increases cache size.

Now try putting that on 14nm.

I'm really not sure what you're trying to argue here. It's bigger, has more transistors per core and cache, and therefore is impossible to produce because...?

You do realize that Skylake-X also has more transistors per core (on account of the AVX-512 units) and more cache overall than the base Skylake, right? Clearly the base Skylake (and Kaby Lake and Coffee Lake) aren't hitting the limits of the 14nm process.

If you're saying it'll be impossible to produce this chip at the same clockspeeds and power draw as Skylake, well, yeah, I don't think anyone here would argue otherwise!

 

[jpiniero]

You're underestimating the amount clocks will drop by. Pretty heavily.

They can always encourage board makers to make the PL2 settings higher. So the base clock would be lower but close to 5 should be doable but it would be like 200 instead of 160 or so.

 

[uzzi38]

If you're saying it'll be impossible to produce this chip at the same clockspeeds and power draw as Skylake, well, yeah, I don't think anyone here would argue otherwise!

My point is that on 14nm they'll be cutting down the die to ensure it doesn't pull stupid amounts of power. There'll be some improvements to the core that they'll axe for die space savings just to ensure they don't have an absolute monster for power draw.

My point is you're not going to be looking at anything even close to 5GHz. Probably OG Skylake clocks instead. Sure you could do better with extreme cooling, but that's mostly irrelevant. At stock it'll fall below Zen 3.

Also apparently it might be well later than I first thought. This is becoming a bigger meme by the minute.

 

[uzzi38]

They can always encourage board makers to make the PL2 settings higher. So the base clock would be lower but close to 5 should be doable but it would be like 200 instead of 160 or so.

I might be getting PL1 and PL2 confused, but afaik one of CML-S's PL states is rumoured to be ~252W, so erm... yeah.

 

[jpiniero]

I might be getting PL1 and PL2 confused, but afaik one of CML-S's PL states is rumoured to be ~252W, so erm... yeah.

PL2 is what the board allows the chip to draw in Turbo Boost 2.0 mode. The board can set it to whatever it wants.

 

[uzzi38]

PL2 is what the board allows the chip to draw in Turbo Boost 2.0 mode. The board can set it to whatever it wants.

https://twitter.com/i/web/status/1150025237037109248

Okay, I was wrong... it's 262W

 

[DrMrLordX]

Plus they trimmed the igp (IGD)

For Rocket Lake-S, the iGPU will be off-die entirely, so that's part of where they'll save on transistor budget. Or at least that's the assumption.

That 14nm process is still showing quite significant improvement with every iteration. That is the story here.

Is it now? 14nm doesn't seem to have improved since they launched the 9900k. 9900KS turned out to be a nice bin, and 10980XE didn't demonstrate any new, amazing voltage/clockspeed curve over the 9980XE. Intel may have pulled out all the stops wrt process optimization.

Okay, I was wrong... it's 262W

Yay? At least that post would seem to confirm that 10c Comet Lake-S still exists.

 

[uzzi38]

Yay? At least that post would seem to confirm that 10c Comet Lake-S still exists.

Like Sharkbay's said, it's just late. That's basically the only reason it's mostly unseen in leaks etc unlike the 6 core die.

If the 10 core doesn't exist Intel will literally be conceding the desktop market right there and then.

Oh wait, they are anyway if RKL-S is Q3 2021

 

[liahos1]

Like Sharkbay's said, it's just late. That's basically the only reason it's mostly unseen in leaks etc unlike the 6 core die.

If the 10 core doesn't exist Intel will literally be conceding the desktop market right there and then.

Oh wait, they are anyway if RKL-S is Q3 2021

If Cometlake is Q1 2020, what comes after it and before RKL in Q3 2021? Something doesnt seem right about that.

 

[kapulek]

Oh wait, they are anyway if RKL-S is Q3 2021

Do we know if desktop ZEN4 is 2021 or 2022?

 

[uzzi38]

If Cometlake is Q1 2020, what comes after it and before RKL in Q3 2021? Something doesnt seem right about that.

Oh hey there, didn't realise you're on here as well. But the fact that you're asking here tells me Chia blocked you.

Absolutely nothing afaik.

 

[uzzi38]

Do we know if desktop ZEN4 is 2021 or 2022?

Should be late 2021. Late Q3 or early Q4.

 

[jpiniero]

If Cometlake is Q1 2020, what comes after it and before RKL in Q3 2021? Something doesnt seem right about that.

Yeah. I thought the U version of Rocket Lake is still coming out next year.

 

[liahos1]

Oh hey there, didn't realise you're on here as well. But the fact that you're asking here tells me Chia blocked you.

Absolutely nothing afaik.

no chia didnt block me lol. Intel usually puts out desktop parts on an annual cadence right? So that Q3'21 seems odd if Cometlake is Q1'20.

Especially odd if its just 14nm.