Discussion Intel current and future Lakes & Rapids thread

[AtenRa]

Here’s some info. Looks like a de-lid has been done.

https://twitter.com/x/status/1453396753701539844

Hell from the picture it seams the P-cores take an enormous die size since the 10 core 14nm Core i9 10900K has the same die size as the 10nm 12900K with just 8+8 cores.

 

[insertcarehere]

Hell from the picture it seams the P-cores take an enormous die size since the 10 core 14nm Core i9 10900K has the same die size as the 10nm 12900K with just 8+8 cores.

Not particularly, 8C GC cluster with 24MB L3 is "only" around 10% larger than 8C TKL with 24MB L3, way smaller than Rocket Lake by any measure.

https://twitter.com/x/status/1453524285260247046

https://twitter.com/x/status/1453479770105262086

 

[AtenRa]

Not particularly, 8C GC cluster with 24MB L3 is "only" around 10% larger than 8C TKL with 24MB L3, way smaller than Rocket Lake by any measure.

https://twitter.com/x/status/1453524285260247046

https://twitter.com/x/status/1453479770105262086

What I mean is that GC Cores are almost double the size compared to Tiger lake Comet Lake at the same node (for example both at 14nm).

We could have a 20 Core 10900K at 10nm at the same size as current AL 8+8.

 

[insertcarehere]

What I mean is that GC Cores are almost double the size compared to Tiger lake at the same node (for example both at 14nm).

We could have a 20 Core 10900K at 10nm at the same size as current AL 8+8.

Tiger Lake is already on the same node at Intel 10nm and GC is not double the size (~10% w/ L3 cluster) , no products with Willow Cove are out at 14nm. Maybe 20 cores Comet Lake could fit in the same size as AL 8+8 but frankly we wouldn't know since no products* with similar cores are out on 10nm.

*cannon lake excepted.

 

[AtenRa]

Tiger Lake is already on the same node at Intel 10nm and GC is not double the size (~10% w/ L3 cluster) , no products with Willow Cove are out at 14nm. Maybe 20 cores Comet Lake could fit in the same size as AL 8+8 but frankly we wouldn't know since no products* with similar cores are out on 10nm.

*cannon lake excepted.

Sorry comet lake not tiger lake, my bad

Well since 10nm has almost 2x the density vs 14nm it could be made.

 

[Hans Gruber]

Sorry comet lake not tiger lake, my bad

Well since 10nm has almost 2x the density vs 14nm it could be made.

For what it's worth. The intel CPU's have at least 50% more chip density than the 7nm Ryzen CPU's. Intel likes to point that out. The power consumption of intel chips, that's another story.

 

[arandomguy]

Seems like no information was given about the GPU. Not sure if that simply means the current K series releases are just focused on target audiences that they feel won't care or that there's essentially 0 changes over Rocket Lake.

Would've hoped for full HDMI 2.1 (4k/120hz) and an updated encode/decode bloc. Not expecting significant performance differences. On the outside chance maybe something to further help leverage the GPU as an accelerator, we're talking about hybrid architectures as the future after all.

The only thing of note I've found parsing various information is that HDMI 2.1 support is now there, however the huge caveat is that it doesn't seem to support full bandwidth as 4k is still limited to 60hz. As there is no other information I could find it's unclear what features, such as VRR, are supported.

No mention of Quick Sync in the launch slides either. So the video decode/encode bloc I guess would be unchanged from TGL/RKL.

 

[Arkaign]

I realize I am in the minority (a big minority), but I honestly wanted a 16 core/24 thread or better that was efficient. My 142 watt 5950x's are killing it in DC, and my EPYC's do geat, but the cheap ones have dried up, so now I want cheap and lots of threads. BTW, I run mine 24/7/365@100% load, so efficiency is a must.

Hell yeah, I am well aware of your beasties and find them invariably impressive, and for VERY good purpose. I'd take one MarkFW over all the miners in the world, you're putting that badass hardware to very real benefit for all of us.

I see a lot of parallels between this recent series of CPUs and the previous tit for tat eras in the industry. The 10900K and 11900K especially remind me of 'Emergency Edition' P4EE lol. A very decent uarch and mature process pushed beyond sanity in an effort to offer at least *something* against the megacore competition on better process tech, and the results were entirely questionable (arguably the only reason for them to exist was extremely poor availability of Zen3, and even then they're still poor value/perf for the vast vast majority of people over an i7 or perhaps even i5 in the case of typical gamers).

The 12900K, I'm hesitant to put in the same category one way or the other until I see what it's like in more sane guise (say set at 5Ghz/4Ghz all core, with good DDR4, and on a decent air cooler). There seems to be something akin to a parabolic efficiency loss in chasing those last few hundred Mhz, though I know for a lot of 'Halo' enthusiast types, they're much less concerned with raw efficiency vs 'look at this record score on whatever'.

Interesting times.

 

[DrMrLordX]

Is DDR4 3200MHz 14-14-14-34 JEDEC standard?

No.

DDR4-3200 CAS/CL 20 is JEDEC standard.

 

[coercitiv]

The 12900K, I'm hesitant to put in the same category one way or the other until I see what it's like in more sane guise (say set at 5Ghz/4Ghz all core, with good DDR4, and on a decent air cooler). There seems to be something akin to a parabolic efficiency loss in chasing those last few hundred Mhz, though I know for a lot of 'Halo' enthusiast types, they're much less concerned with raw efficiency vs 'look at this record score on whatever'.

I would definitely put ADL-S in the same category: very decent uarch and mature process pushed beyond sanity to get top spots in the performance charts. There is no sane reason to push 240W TDP on a 8+8 consumer desktop product when the 6+8 mobile product works well at 45W / 65W. Do that on HEDT with a 20+ core SKU, that's something I would like to see.

This thread is very long. You just responded to a post from 2017. That particular user hasn't even posted anything since August.

Don't talk to bots.

 

[jpiniero]

Intel must have gotten some amount of density increase with I7 versus 10SF given the core sizes.

I would definitely put ADL-S in the same category: very decent uarch and mature process pushed beyond sanity to get top spots in the performance charts. There is no sane reason to push 240W TDP on a 8+8 consumer desktop product when the 6+8 mobile product works well at 45W / 65W.

From Intel's own slide, the 12900K gets an extra 20% in performance going from 125W to 240W in Blender. Might even be a bigger gap with faster memory. That's why. It's not beating anything Zen 3 at 125 W in high threaded apps.

 

[DrMrLordX]

Intel must have gotten some amount of density increase with I7 versus 10SF given the core sizes.

That's curious. Going from Cannonlake to IceLake, didn't they reduce density just to achieve a workable node?

What's density look like between 10nm+ and 10SF?

 

[Abwx]

Intel must have gotten some amount of density increase with I7 versus 10SF given the core sizes.



From Intel's own slide, the 12900K gets an extra 20% in performance going from 125W to 240W in Blender. Might even be a bigger gap with faster memory. That's why. It's not beating anything Zen 3 at 125 W in high threaded apps.

So much about the stated efficency breakthrough...

Beside 241W peak is the same as the 5950X 142W peak provided you read it from right to left...

 

[mikk]

Seems like no information was given about the GPU. Not sure if that simply means the current K series releases are just focused on target audiences that they feel won't care or that there's essentially 0 changes over Rocket Lake.

Would've hoped for full HDMI 2.1 (4k/120hz) and an updated encode/decode bloc. Not expecting significant performance differences. On the outside chance maybe something to further help leverage the GPU as an accelerator, we're talking about hybrid architectures as the future after all.

The only thing of note I've found parsing various information is that HDMI 2.1 support is now there, however the huge caveat is that it doesn't seem to support full bandwidth as 4k is still limited to 60hz. As there is no other information I could find it's unclear what features, such as VRR, are supported.

No mention of Quick Sync in the launch slides either. So the video decode/encode bloc I guess would be unchanged from TGL/RKL.

GPU clock speeds are known: https://download.intel.com/newsroom/2021/client-computing/12th-gen-processor-product-brief.pdf

Up to 1550 Mhz, this is a 250 Mhz increase from i9-11900K.

 

[mikk]

Where did you get the information from?

I don't think so as the backside shot of the CPU shows otherwise.

Everyone who did a calculation say that and apparently this number has been confirmed by Intel.

https://twitter.com/x/status/1453429070084120577

 

[Hulk]

I'm waiting to see some workloads run at the same power for ADL and Zen 3 before I make any judgements.

For example, in a highly MT workload 16 Zen 3 cores running at 4GHz are probably going to be as fast or faster than 8GC's at 5GHz and 8GM's at 3.9GHz. We all know that for MT workloads lots of cores at lower frequency is very efficient. But the 5950X costs more up front so if you don't use the compute like that a lot then it might not be a good buy for you. The 12900K might be a good option if you generally need less compute but then have "bursty" workloads now and then. Like me when I load up a multitrack audio project and need really good ST performance to avoid dropouts. Or editing video and a certain preview effect needs full compute for like 4 seconds.

On the other end of the spectrum cost-wise, we could have the 12900K vs 5800X, now the tables are reversed and perhaps at the the same power performance could swing the other way?

Let's just be intellectually honest and explore all facets of performance and efficiency without bias.

 

[Abwx]

On the other end of the spectrum cost-wise, we could have the 12900K vs 5800X, now the tables are reversed and perhaps at the the same power performance could swing the other way?

Let's just be intellectually honest and explore all facets of performance and efficiency without bias.

5800X has way less cores and is stretched TDP wise, on AMD own settings going from Eco Mode at 80W to default will increase power by 40% for 6% better perfs.

Best comparison is against the 5900X that is also 24T, guess that it s actually Intel s own real contender for the 12900K if we are to look at their pricing.

 

[Hulk]

5800X has way less cores and is stretched TDP wise, on AMD own settings going from Eco Mode at 80W to default will increase power by 40% for 6% better perfs.

Best comparison is against the 5900X that is also 24T, guess that it s actually Intel s own real contender for the 12900K if we are to look at their pricing.

Yes. I agree 12900K vs. 5900X will be quite interesting.

 

[nicalandia]

No. Using 12400 CPUz I come up with 34%.
Check me here.
ST score 682 @4.4GHz, MT score 4984 @4ghz.
(4984/6)(4.4/4)=914 (Single core score @ 4.4GHz)
~34%
yeah, seems high

Yeah, I was expecting that from such a Fat Core

 

[eek2121]

Intel must have gotten some amount of density increase with I7 versus 10SF given the core sizes.



From Intel's own slide, the 12900K gets an extra 20% in performance going from 125W to 240W in Blender. Might even be a bigger gap with faster memory. That's why. It's not beating anything Zen 3 at 125 W in high threaded apps.

IMO, If you are looking to compare efficiency of the cores, the correct thing to do would be to disable the gracemont cores, and half the 5950X (8 cores vs 8 cores), cap them both at 125W, and test them both and see which one performs better. On the Intel side we don’t have a 16 core GC part. I suspect this isn’t due to power, and more to do with die area. A 16 core GC chip would be huge.

That is why, IMO, the future is the atom core. Atom cores appear to be much more efficient in terms of die area and perf/watt.

 

[AtenRa]

That is why, IMO, the future is the atom core. Atom cores appear to be much more efficient in terms of die area and perf/watt.

At the expense of Single Thread Performance

 

[NostaSeronx]

At the expense of Single Thread Performance

Jasper Lake-6W has pretty good STP&MTP against Dali-6W.
Alder Lake-N(0+8+(1?)) at a given TDP probably will do what Jasper Lake does against Dali to Mendocino at a given TDP.

Atom is good enough against Zen. Especially at the low-end.

 

[nicalandia]

That is why, IMO, the future is the atom core. Atom cores appear to be much more efficient in terms of die area and perf/watt.

Atom is good enough against Zen. Especially at the low-end.

Not where money matter as in Xeons, you wont find e-cores on 10nm and 7nm Xeons because they need the absolute best performance to compete with ZEN4/ZEN5 EPYC. They can't do that with Skylake type performance Xeons(they are already getting mauled by Zen3)

 

[lobz]

AFAIK, The size of Tiger Lake H 32EU was like ~190mm2. Sounds like an ADL-P is a real deal. Almost the same die size, similar manufacturing process but much more powerful and power-efficient. Tiger Lake H wasn't that bad either BTW.

No, it was most certainly not! 🙂

 

[NostaSeronx]

Not where money matter as in Xeons, you wont find e-cores on 10nm and 7nm Xeons because they need the absolute best performance to compete with ZEN4/ZEN5 EPYC. They can't do that with Skylake type performance Xeons(they are already getting mauled by Zen3)

Rather than have 15*4 GC-Server cores in mesh, would it not be better to have 15*4 Gracemont-Server quad-core clusters in mesh?
Then, Intel won't be sniped by 192C N2 & 96C V1 as well as 64C/96C/128C Zens.

Mind that third "Vec" ALU looks pretty ideal for AMX.