Discussion Intel current and future Lakes & Rapids thread

[scineram]

TGL-H 45W:

https://twitter.com/x/status/1373247155985551365

Hang on, are they retaining the UHD Graphics branding? I thought it will be X^e from now on.

 

[jpiniero]

Hang on, are they retaining the UHD Graphics branding? I thought it will be X^e from now on.

They are for the 32 EU die models.

 

[SAAA]

If that diagram is true scale then the Gracemont cores aren't much smaller than the Golden Cove. I'm wondering if "Skylake" level performance for the Little cores means Skylake cores.

Not sure how schematic it is vs reality, but the fact they added 8 Gracemont cores rather than 4 Golden cores (guess estimating) for that area hints the contrary. They can't be that large or it would be pointless.

2x MT is unrealistic unless it's a workload that is severely memory bandwidth limited. It's not like you are getting any additional big cores.

2xMT on scalable tests like cinebench and similar shouldn't be unrealistic. But it does point to small cores being quite high clocked:

Rocket lake
8 cores 1x (total performance with HT)
8 MT

Alder lake
8 big cores 1.2x (total with HT)
8 small 0.8x (no HT)
16 MT

For the small cores to pull around 67% of the big cores with HT they must have decent IPC and clocks.

Moreover it's upto 20% ST, not even IPC.

Well, that's true, but if the comparison is with Rocket lake on 14 nm I doubt 10 nm will clock much higher at all, we are talking 5-5.3 GHz already, hence up to 20% should be mostly IPC in this case.

 

[lobz]

Rocket lake
8 cores 1x (total performance with HT)
8 MT

Alder lake
8 big cores 1.2x (total with HT)
8 small 0.8x (no HT)
16 MT

I hope you're not a math teacher.

 

[NTMBK]

Not sure how schematic it is vs reality, but the fact they added 8 Gracemont cores rather than 4 Golden cores (guess estimating) for that area hints the contrary. They can't be that large or it would be pointless.



2xMT on scalable tests like cinebench and similar shouldn't be unrealistic. But it does point to small cores being quite high clocked:

Rocket lake
8 cores 1x (total performance with HT)
8 MT

Alder lake
8 big cores 1.2x (total with HT)
8 small 0.8x (no HT)
16 MT

For the small cores to pull around 67% of the big cores with HT they must have decent IPC and clocks.



Well, that's true, but if the comparison is with Rocket lake on 14 nm I doubt 10 nm will clock much higher at all, we are talking 5-5.3 GHz already, hence up to 20% should be mostly IPC in this case.

2X MT might reflect that the big cores clock down less under load, too.

 

[JoeRambo]

Moreover it's upto 20% ST, not even IPC.

At this point, can they really push clocks anywhere? Even on 10nm+++++ process they are on now? If your baseline is ~5ghz, 100mhz is 2% clock difference.

I'd be more worried about "up to" part. Don't really need more IPC with some esoteric FP or even worse - AVX512 workload

 

[jpiniero]

If that diagram is true scale then the Gracemont cores aren't much smaller than the Golden Cove. I'm wondering if "Skylake" level performance for the Little cores means Skylake cores.

My expectation is that the Cluster is about the size of a Golden Cove core. Gracemont being basically Skylake with max density is plausible but would be tough even assuming Golden Cove cores are gigantic.

 

[mikk]

If that diagram is true scale then the Gracemont cores aren't much smaller than the Golden Cove. I'm wondering if "Skylake" level performance for the Little cores means Skylake cores.

Atom and Core are completely different, Atom doesn't even support SMT. If Gracemont reaches Skylake IPC it's a natural evolution, all the previous Atom generations added like 20-30% IPC. Tremont is 1 generation behind Skylake (roughly Haswell IPC).

 

[yuri69]

WTF Intel?

ALD topping at a maximal 20% ST gain? The average IPC gains is thus possibly significantly lower. I expected the 20% figure to be an average IPC gain.

So far it seems, ALD is not that revolutionary Conore-moment stomping AMD to the ground again. Zen 4 will likely have no problem maintaining parity.

 

[mikk]

WTF Intel?

ALD topping at a maximal 20% ST gain? The average IPC gains is thus possibly significantly lower. I expected the 20% figure to be an average IPC gain.

Why it is significantly lower? RKL-S goes up to 5.3 Ghz in the desktop, are you confident Golden Cove can clock significantly higher than this? In fact the IPC gain can be higher, they may not need 5.3 Ghz for ADL-S.

So far it seems, ALD is not that revolutionary Conore-moment stomping AMD to the ground again. Zen 4 will likely have no problem maintaining parity.

ADL is a game changer for Intel if they really can double the MT performance, even for RKL-S (or TGL-U mobile) ST performance is not the real problem. They are behind in multithread performance (beside power efficiency) by a huge margin no matter if deskop or mobile.

 

[LightningZ71]

Hmm, just thinking out loud here. If the big cores on ADL can manage to be roughly the same as RKL + 15% on average, peaking to 20%, and the small cores can be roughly equivalent in performance to six threads on rocket lake now, that'll give an MT performance that's roughly between the 5900x and 5950x, excluding memory effects. I base that on the 5800x being roughly equivalent to the 11900.

If they can effectively produce a product that competes with the 5950x on a reasonable size monolithic die with decent yields, which, after two years since the process it's produced on has been in some sort of volume production seems reasonable, then that's likely enough to bridge the cost gap for Intel and put AMD in the position to have to spend more to manufacture a competing chip.

For AMD, N6 production of a Zen3 might be enough. If they did nothing but double the L2 per core, expand some internal buffers, improved power and thermals enough to increase the average max boost by 100mhz and the sustained all-core by a similar amount, with similar sized CCDs, they would be able to keep ADL 8+8 at or below the 12 core part on average and keep the 16 core part as a stand alone premium product.

 

[jpiniero]

Atom and Core are completely different, Atom doesn't even support SMT. If Gracemont reaches Skylake IPC it's a natural evolution, all the previous Atom generations added like 20-30% IPC. Tremont is 1 generation behind Skylake (roughly Haswell IPC).

Using Skylake as a base seems like a reasonable thing to do. Cut whatever doesn't make sense with a Core without HT, add any additional instructions to get parity with Golden Cove, go max density.

 

[yuri69]

Why it is significantly lower? RKL-S goes up to 5.3 Ghz in the desktop, are you confident Golden Cove can clock significantly higher than this? In fact the IPC gain can be higher, they may not need 5.3 Ghz for ADL-S.

ADL is a game changer for Intel if they really can double the MT performance, even for RKL-S (or TGL-U mobile) ST performance is not the real problem. They are behind in multithread performance (beside power efficiency) by a huge margin no matter if deskop or mobile.

The devil lies in the "up to" phrase. Any figure featuring "up to" doesn't represent any form of central tendency.

An extreme case: the "up to 20%" could represent a set of 100 benchmarks each reaching a +10% gain and a single benchmark reaching an outlier of a +20% gain. So in that case the IPC would really be significantly lower.

Don't focus on math. Focus on marketing.

The scenario you described implies the IPC gain being greater than the total ST performance gain due the frequency regression. If this was the case, marketing would surely present the IPC number. Technical marketing always goes after the bigger, above norm/unprecedented, class-leading figures.

Marketing fluff aside, mobile Tiger Lakes can already reach 5GHz. Desktop (massive VRMs, etc.) ALD-S is based on an improved variant of the TGL's process. My bet is Intel won't really regress in the frequency department.

As for the MT performance, sure, it goes past 10c. But that's still a niche. The same goes for all those mighty laptop workloads requiring 16 threads.

 

[DrMrLordX]

Rocket Lake-S is already "up to" 19% faster than Skylake, and we all saw how that worked out, didn't we?

 

[Hulk]

Hmm, just thinking out loud here. If the big cores on ADL can manage to be roughly the same as RKL + 15% on average, peaking to 20%, and the small cores can be roughly equivalent in performance to six threads on rocket lake now, that'll give an MT performance that's roughly between the 5900x and 5950x, excluding memory effects. I base that on the 5800x being roughly equivalent to the 11900.

If they can effectively produce a product that competes with the 5950x on a reasonable size monolithic die with decent yields, which, after two years since the process it's produced on has been in some sort of volume production seems reasonable, then that's likely enough to bridge the cost gap for Intel and put AMD in the position to have to spend more to manufacture a competing chip.

For AMD, N6 production of a Zen3 might be enough. If they did nothing but double the L2 per core, expand some internal buffers, improved power and thermals enough to increase the average max boost by 100mhz and the sustained all-core by a similar amount, with similar sized CCDs, they would be able to keep ADL 8+8 at or below the 12 core part on average and keep the 16 core part as a stand alone premium product.

I like this theory. Intel could be playing the "long game." Somewhere between 5900X and 5950X performance on a monolithic die. If yields are good the profit margin would be there and no dealing with chiplets. And if this is actually the case they have some transistor budget to improve performance and/or add cores when moving to 7nm in 2055.

 

[Zucker2k]

Moreover it's upto 20% ST, not even IPC.

You expect ADL to clock higher than Willowcove or Cypresscove?

 

[Mopetar]

Rocket Lake-S is already "up to" 19% faster than Skylake, and we all saw how that worked out, didn't we?

The performance honestly isn't that bad, especially for something that had to be back ported which is going to require compromises and possibly outright losing some performance improvements due to the design becoming infeasible on the older process node.

Obviously they're still behind AMD in general, but outside of a few benchmarks, the gap isn't that wide. The main problem is that the clocks are pushed beyond the point where the chip can hope to compete efficiently in terms of power consumption, and that if Intel does charge as much as has been rumored, they won't compare favorably with AMD's similarly priced offerings that have more cores, which makes the value go down considerably for anything that can use all of those extra threads on the AMD CPU.

 

[lobz]

Rocket Lake-S is already "up to" 19% faster than Skylake, and we all saw how that worked out, didn't we?

It crushes geekbench quite relentlessly.

 

[Ajay]

You expect ADL to clock higher than Willowcove or Cypresscove?

Not Cypress, but Willow, sure - high TDP.

 

[dmens]

LOL 20%. Y'all need a gullibility meter.

 

[tamz_msc]

You expect ADL to clock higher than Willowcove or Cypresscove?

10nm ESF should clock higher than 10 nm SF(Willow Cove), at least in theory.

 

[mikk]

Using Skylake as a base seems like a reasonable thing to do. Cut whatever doesn't make sense with a Core without HT, add any additional instructions to get parity with Golden Cove, go max density.

They have a real small core architecture which is a much better fit, they don't need Skylake.

Rocket Lake-S is already "up to" 19% faster than Skylake, and we all saw how that worked out, didn't we?

Up to 19% in Singlethread and yes it looks plausible.

 

[Timorous]

Up to 19% in Singlethread and yes it looks plausible.

Only in Specfp. Was a lot lower in Specint.

 

[RTX2080]

Rocketlake, Cometlake Refresh(10325), Zen3 IPC comparison, all being locked at 4C8T, 4Ghz

我还能再挤一点点，Intel 10105 10325 快速评测

虽然 11 代处理器定档 3 月 30 号，但是评测已经漫天飞舞，作为数码博主的我自然也是不能错过，该买买改借借。我前段时间已经快速评测过了 i5 ，今天就给各位带来几款新的 i3 快速上手评测。 Intel 11 代基本上已…

translate.google.com

 

[jpiniero]

Thought about it some more and I guess the 2x MT is more realistic if you are talking about a power constrained scenario. Base on the 11900T is only 1.5 versus 2.1@35W for the 11900H.