Discussion Intel current and future Lakes & Rapids thread

[majord]

Just looking over die shots - something of note, In the Annotated shots, the 2MB shared L2 for Gracemont is about 2.5x the Density of the 1.25MB L2 on GC .

 

[Thala]

There are plenty of markets that would gladly tolerate Skylake level single core (or worse) for significantly higher multicore performance.

It might be acceptable in some markets, but not in the hyperscaler (aka cloud) domain.

 

[zir_blazer]

Thanks to the guy that posted the link of the Reddit Intel AMA, I could ask and be given a few answers. PCIe Bifurcation tops at 8x/8x/4x (Previously 8x/4x/4x), and the Processor PCIe Root Ports supports ACS. That one is very good news for me.

However, shouldn't it be time already for Alder Lake to get its own, proper Thread, instead of this 562 pages Jupiter? I hate Megathreads, they become way too disorganized.

 

[Hulk]

However, shouldn't it be time already for Alder Lake to get its own, proper Thread, instead of this 562 pages Jupiter? I hate Megathreads, they become way too disorganized.

Much as I have enjoyed this monster I'll give starting an ADL thread a try and see if it takes.

 

[eek2121]

It might be acceptable in some markets, but not in the hyperscaler (aka cloud) domain.

Are you kidding me? Cloud providers would eat a fictional 256 core gracemont product up. No hyperthreading to deal with, massive
core counts, great IPC and awesome power draw.

One can’t seriously argue for an ARM based solution and simultaneously say that an Intel chip with a bunch of atom cores would be a bad idea.

Assuming all the leaks and Intel are right about performance, such a chip would beat a 64 core EPYC milan in multicore performance easily.

Too bad we’ll never see one.

 

[NostaSeronx]

Too bad we’ll never see one.

Are we sure?
Sierra Forest is a thing.

Grand Ridge is the small 5G base station SoC using Gracemont on Intel 4.
Sierra Forest is the big Advanced Performance(-AP solution) server SoC also using Gracemont on Intel 4.

Grand Ridge being 24-cores.
Sierra Forest being 128-cores, with rumors/guesstimates of 256-core and 512-core counts. Do to the sheer size of Birch Stream AP socket(7500+ pins vs SP5@~6100 pins[Zen5-T hv/ 192c&256c options]) and the shrink from actual 7nm(EUV).

 

[insertcarehere]

Skylake level IPC is UNACCEPTABLE When Zen3 EPYC are baby seal clubbing Skylake based Xeons...

It might be acceptable in some markets, but not in the hyperscaler (aka cloud) domain.

Graviton2 is based on cores with Skylake-level IPC and last I heard they have been pretty successful in getting clients to bite as part of AWS, but hey, what would the world's largest cloud computing platform possibly know about client requirements in the cloud and server domains?

 

[eek2121]

Are we sure?
Sierra Forest is a thing.

Grand Ridge is the small 5G base station SoC using Gracemont on Intel 4.
Sierra Forest is the big Advanced Performance(-AP solution) server SoC also using Gracemont on Intel 4.

Grand Ridge being 24-cores.
Sierra Forest being 128-cores, with rumors/guesstimates of 256-core and 512-core counts. Do to the sheer size of Birch Stream AP socket(7500+ pins vs SP5@~6100 pins[Zen5-T hv/ 192c&256c options]) and the shrink from actual 7nm(EUV).

😍wil have to look this up. I was under the impression that Atom was to remain where it is.

Graviton2 is based on cores with Skylake-level IPC and last I heard they have been pretty successful in getting clients to bite as part of AWS, but hey, what would the world's largest cloud computing platform possibly know about client requirements in the cloud and server domains?

Don’t bother, they have no clue. The whole appeal behind EPYC was never about single core performance, it is about perf/watt and core density, combined.

Don’t get me wrong, My 5950X is fantastic, but unless Zen performs well while saving money…luckily AMD is doing well in that regard for now.

 

[IntelUser2000]

Thanks everyone.

Looks like Golden Cove is smaller than expected. That's good. ~30% increase in core area for 19% perf/clock increase and an additional pipeline is a very good result.

Gracemont is 2x the size of Tremont. Larger than expected, but little better performing than expected. Slightly on the disappointing side. Perhaps it's larger because they wanted to reach 4GHz?

 

[coercitiv]

So the 6+0 chip is 163 mm2. Via Andreas Schilling on Twitter.

 

[JoeRambo]

So the 6+0 chip is 163 mm2. Via Andreas Schilling on Twitter.

That is kinda HUGE in a bad way? for 18MB of L3 and 6 large cores + GPU ? 8+8+GPU+30MB of L3 is 209mm. So 46m^2 for 2 clusters of atoms + 2 big cores + associated slices for L3 for 12MB total ?

EDIT: no idea about iGPU sizes, don't roast me if I have missed some 32 vs 96 thing.

 

[rainy]

Looks like Golden Cove is smaller than expected. That's good. ~30% increase in core area for 19% perf/clock increase and an additional pipeline is a very good result.

I disagree with you - Zen 3 core is circa 9 percent bigger than Zen 2 with 19 percent higher IPC which is far more impressive.
Btw, Golden Cove is significantly or even much bigger than Zen 3 and probably would be only a bit faster.

 

[coercitiv]

So 46m^2 for 2 clusters of atoms + 2 big cores + associated slices for L3 for 12MB total ?

Napkin math time

Based on the layout presented by Intel for the 8+8 die it seems that the cores themselves + L3 make up for ~55% of the die, and 6 P cores account for almost 55% of total "core + L3" area. So if we remove around 45% of the total core area which in turn represents around 55% of the whole die... we shrink the die by a (very) rough estimate of 25%.

 

[inf64]

I disagree with you - Zen 3 core is circa 9 percent bigger than Zen 2 with 19 percent higher IPC which is far more impressive.
Btw, Golden Cove is significantly or even much bigger than Zen 3 and probably would be only a bit faster.

I agree, versus Zen3 it looks like a very inefficient use of die space. Versus Cypress Cove it's a nice improvement, no doubt about it. The small cores are a new way of doing more efficient multithreading/handling of non-intensive tasks, so that is a plus as well.

 

[Timorous]

Looking at 11600K vs 11900K performance deltas at 1080p from Computerbase (11900K + 12%), Techspot/HUB (11900K + 8%) it seems reasonable to conclude that with 50% more L3 in the 12900K vs the 12600K vs the 33% difference in RKL chips and the different turbo power limits the 12900K is likely to be around 10% or so faster than the 12600K.

Based on the Intel delta of the 12900K vs the 11900K it seems that the 12600K should roughly match the 11900K with some big wins here and there on average. With Computerbase having the 11900K as 3% faster than the 5600X and HUB having the 11900K being 1% faster than the 5600X it looks like at 1080p the 12600K will only beat the 5600X by a small margin on average which considering the platform cost is not great.

This does not make me feel that confident there will be much price movement by AMD on their stack because really 5600/5800 skus are targeting gamers. The 5900/5950 might see some movement but it depends how perf/watt and productivity performance shakes out IMO.

 

[mikk]

Gracemont is 2x the size of Tremont. Larger than expected, but little better performing than expected. Slightly on the disappointing side. Perhaps it's larger because they wanted to reach 4GHz?

What about AVX, how much area does it need?


6+0 seems roughly at 160 mm² in size. The i5-12400 could be the big star in the mainstream segment. It shouldn't cost more than 200 USD given that the i5-12600 is in the 2xx range.

i5-12600K $ 289
i5-12600KF $ 264

 

[Shivansps]

The I5-12400 will be really awkard for AMD, but so has been the 9400, 10400 and 11400 and they do not seem to care at all. So is not a suprise to me if the best Intel product end up being that one.

 

[JoeRambo]

The I5-12400 will be really awkard for AMD, but so has been the 9400, 10400 and 11400 and they do not seem to care at all. So is not a suprise to me if the best Intel product end up being that one.

They were "faster", "WAY more power efficient" and sold out most of the time anyway. But i think i saw plenty of recommendations for mid-level gaming for 11400. I'd expect 12400 to recieve even more praise.

 

[deasd]

The I5-12400 will be really awkard for AMD, but so has been the 9400, 10400 and 11400 and they do not seem to care at all. So is not a suprise to me if the best Intel product end up being that one.

Next year the AMD's lineup would be filled with non-X Zen3s which were B2 stepping, like 5600, 5900, if they price 5600 at well below 200$ then there's nothing to awkward. Last info about B2 is with better OC headroom and thermal. Ahhh I forgot the mobo price that have to add up though....
I think the real awkward is for the Rocketlakes which were still on the shelve......

 

[TESKATLIPOKA]

Twitter

35% higher score after increasing TDP to 241W.

 

[John Carmack]

Twitter

35% higher score after increasing TDP to 241W.

This essentially tells us why the PL1 on the MSI board is set to be equal to PL2 at 241W because Alder Lake wouldn't win any MT benchmarks otherwise.

 

[tomatosummit]

Twitter

35% higher score after increasing TDP to 241W.

I know this is very rough maths and the i9 is probably power starved, the difference between the i5, i7 and i9 shows adding each extra E core gives roughly 40% of a P core's performance when the core's fully loaded.
I know they were expected to be around 60% of the performance but I wonder if when fully loaded they drop further behind, a limitation of the shared l2 and ring stop perhaps?

 

[insertcarehere]

Twitter

35% higher score after increasing TDP to 241W.

Now...if we're taking these scores at face value, then it means that a 12600k limited to 125w power draw is competitive with a 5800X. Really doesn't speak well for i7 & i9 efficiency vs 5900X however.

Edit: In context, the Chinese post advertising the legion desktop is advertising they can provide all the power/cooling for 12th gen vs other brands that may not, so Lenovo likely has a vested interest to make the performance difference in this chart look big.

 

[Zucker2k]

I know this is very rough maths and the i9 is probably power starved, the difference between the i5, i7 and i9 shows adding each extra E core gives roughly 40% of a P core's performance when the core's fully loaded.
I know they were expected to be around 60% of the performance but I wonder if when fully loaded they drop further behind, a limitation of the shared l2 and ring stop perhaps?

The E-cores are only running at a measly 3.7GHz, remember?

 

[tomatosummit]

The E-cores are only running at a measly 3.7GHz, remember?

yes, which is why i wrote performance instead of ipc.

The <40% relative performance shown on that c23 result for each Ecore compared to Pcore is quite a bit less than expected unless the i9 is still power starved at 240w.
I think this could indicate scaling limitations of the mont cores at desktop power levels.