Discussion Intel current and future Lakes & Rapids thread

[exquisitechar]

So is SPR using Golden or Willow Cove?

Leveraging native AMX and a massive IPC increase (compared to Sunny Cove) would point to Golden.

It's Golden, yes.

 

[naukkis]

Intel needs massive IPC uplift - and massive performance/watt uplift at the same time. They need completely new uarch to be competitive, if they just keep making those Cove-cores bigger their end product will be useless for about every use case.

 

[Hitman928]

How much IPC increase is Golden Cove expected to have? If the ~50% over Skylake claim is true that would require a 25% minimum IPC increase over Sunny Cove.

 

[yuri69]

The Golden Cove's gain was rumored to be in line with the Sunny Cove/Zen 3 IPC gain => around 20%.

Skylake => Sunny => Willow => Golden should be like: 1.18*1.02*1.2 = 1.44 or 44% gain.

 

[DrMrLordX]

Just like Comet Lake is not only 10C sales, same buyers who are buying 10400F or 10600K or 10850K or 3600X on deals will have a additional option in Rocket Lake-S.

+10% IPC and -2 cores doesn't look like it's going to make the product that much better than just Comet Lake-S, especially if Intel is forced to massively discount everything.

 

[Hulk]

I was having a look back at some Ice Lake (Sunny Cove) and Tiger Lake (Willow Cove) benchmark results. It's so hard to calculate IPC gains with Intel reference systems and unknown clocks. Intel says +18% for Sunny Cove over Skylake. Looking at the numbers this is possible but of course it depends on what applications you use to arrive at this number.

Rocket Lake is Sunny Cove with Xe graphics so Rocket Lake could close the gap on Zen 3 somewhat but they still have the thermal (process) problem.

I'm starting to think Intel is staying with 14nm for the desktop parts because they can't get the clocks high enough with 10SF to compete with Zen 3. They have made the calculation that they will go for performance and yields by staying with their higher leakage/higher clocking nuclear 14nm++++ process.

From a marketing point of view they get to "sell" some cherry picked benches and no one really has to know about the power and thermals... meaning the people buying Dell/HP desktops at Best Buy.

Intel probably could have gotten another 8% or so uptick in IPC by going with Willow Cove but they probably couldn't stand the lower yields/increased die size/lower profits that would have resulted with those large Willow Cove caches.

So basically had to choose from a bunch of crappy options for their next desktop release.

14nm with Sunny Cove, damn the thermals and hope they are competitive with Zen 3.
or
14nm with Willow Cove, damn the thermals and damn the margins, but they will be more competitive with Zen 3.
or
Simply refine 10SF and put current Tiger Lake on the desktop at the best clocks they can manage. Hope the yields and performance work out.

It's hard to say what the right decision is because we don't have all of the data but I think I would have pushed for 14nm Willow Cove cores at highest clocks possible and deal with the nuclear furnace that results until they can transition to 10nm had I been at the big table at Intel.

Finally, one more thing has been on my mind since we started the Handbrake testing with reporting Average Effective Clock from HWinfo. The Zen 3 parts are only managing just under 3700Mhz while the 10900k on the chart is just below 5000MHz. That's a huge difference clockspeed as reported by HWinfo's average effective clockrate.

Zen 3 more than makes up for this clock deficit with IPC but if Intel loses 500-800MHz by moving to 10SF on the desktop they simply don't have the core to make up this performance loss.

While we don't have a lot of data as yet with the new Handbrake bench, using Average Effective Clock and my efficiency calculation Zen 3 has a 50% IPC advantage in this Handbrake test over Skylake.

 

[uzzi38]

Intel needs massive IPC uplift - and massive performance/watt uplift at the same time. They need completely new uarch to be competitive, if they just keep making those Cove-cores bigger their end product will be useless for about every use case.

You can go bigger and still improve perf/W. You just need to ensure that your IPC increase outstrips any increase to power at the same clock.

 

[Shivansps]

+10% IPC and -2 cores doesn't look like it's going to make the product that much better than just Comet Lake-S, especially if Intel is forced to massively discount everything.

Well, if Intel focus on the middle area with good prices and that stops the increase in prices ill say we all win. Otherwise CPUs are going to face the same fate as GPUs.

 

[DrMrLordX]

Well, if Intel focus on the middle area with good prices and that stops the increase in prices ill say we all win. Otherwise CPUs are going to face the same fate as GPUs.

I'm not inclinded to disagree with you there. Rocket Lake-S hitting with a max price point in the $300-$350 range (for the 11900k) would be a massive boon for budget buyers who probably can't find anything but Comet Lake-S on the market today, especially when you consider what would happen to Comet Lake-S prices in such an event. Unfortunately, the last time Intel was the "discount brand" was in late 2005/early 2006 with Pentium D (and the infamous Pentium D 820), which is a position that it took years for them to assume as AMD slowly ate away at their performance margins. Intel still had best-in-class performance as of late 2018 and best gaming performance as of September/October of 2020, so Intel could easily come out with $530 11900k under the assumption that they can snap up sales AMD loses due to CPU shortages. Supply/demand will play a big role here in determining how much, if any, downward price pressure there will be on Rocket Lake-S and Comet Lake-S after January.

I still think that, if prices do plummet, Comet Lake-S (which will still be widely-available) will be attractive enough that Rocket Lake-S may be a hard sell.

 

[uzzi38]

Well, if Intel focus on the middle area with good prices and that stops the increase in prices ill say we all win. Otherwise CPUs are going to face the same fate as GPUs.

I'm expecting Intel to stick with CML-S pricing for respective RKL-S SKUs. With the exceptionally useless i9 on top this time around.

 

[JoeRambo]

I'm not inclinded to disagree with you there. Rocket Lake-S hitting with a max price point in the $300-$350 range (for the 11900k)

Why are we so focused on "Halo" products or their pricing? 11900K could cost 666 eur - i don't care, what is important is 11400 priced at say 180 eur and decent perf and features. What about 10600K with price that is 100 eur lower than 5600X ? AMD reacting and dropping prices or releasing additional SKUs - even better for us customers!

 

[AMDK11]

So is SPR using Golden or Willow Cove?

Leveraging native AMX and a massive IPC increase (compared to Sunny Cove) would point to Golden.

It is logical that WillowCove will not be in the next generation Xeon for a very simple reason. How is x86 SunnyCove different from x86 WillowCove? The difference lies only in the cache subsystem which is Inclusive in SunnyCove and non-Inclusive in WillowCove. We already have an equivalent of WillowCove in Xeon in the form of Icelake-SP which, like WillowCove, has a non-Inclusive cache subsystem, but with different capacity. As for IPC WillowCove, Intel itself emphasized that there is no increase in IPC. And since WillowCove in terms of the x86 core is the same as SunnyCove, the difference will result from favoring a given soft inclusive and a smaller L2 or non-inclusive and a larger L2.

x86 SunnyCove L2 512KB and L3 2MB Inclusive
x86 WillowCove L2 1.25MB and L3 3MB non-Inclusive
x86 Icelake-SP L2 1.25MB and L3 1.5MB non-Inclusive

 

[JoeRambo]

It is logical that WillowCove will not be in the next generation Xeon for a very simple reason. How is x86 SunnyCove different from x86 WillowCove? The difference lies only in the cache subsystem which is Inclusive in SunnyCove and non-Inclusive in WillowCove. We already have an equivalent of WillowCove in Xeon in the form of Icelake-SP which, like WillowCove, has a non-Inclusive cache subsystem, but with different capacity.

Nope. Skylake-SP already had that L2 non-inclusive in L3 structure, but L1 contents were inclusive in L2. Icelake-SP will keep that structure as WillowCove is simply a different core, with different coherency controller that is aware of new cache hierarchy. So that is not the reason for WillowCove being MIA in servers. I think real reasons has to do with timing, 10nm, Icelake-SP already being in pipeline for half of decade.

 

[AMDK11]

Nope. Skylake-SP already had that L2 non-inclusive in L3 structure, but L1 contents were inclusive in L2. Icelake-SP will keep that structure as WillowCove is simply a different core, with different coherency controller that is aware of new cache hierarchy. So that is not the reason for WillowCove being MIA in servers. I think real reasons has to do with timing, 10nm, Icelake-SP already being in pipeline for half of decade.

WillowCove is like Skylake-X. Skylake and SunnyCove are cores with the Inclusive subsystem, and Skylake-X, SunnyCove-SP, and WillowCove are the transition from the same x86 core to the non-exclusive cache subsystem. I don't think the move from SunnyCove-SP to WillowCove will bring about any real performance boost on Xeon, especially since the x86 core itself is the same. Another thing is GoldenCove, which will bring a sizable increase in IPC with the same type of cache subsystem as WillowCove.

Besides, the IceLake-SP is not only a SunnyCove with a non-inclusive cache subsystem, but also a double FMA 512bit execution unit. SunnyCove / CypressCove and WillowCove have two FMA units 1x 512bit + 1x 256bit, i.e. they can perform 1x 512bit or 2x 256bit.

 

[JoeRambo]

WillowCove is like Skylake-X. Skylake and SunnyCove are cores with the Inclusive subsystem

Willow Cove is not like Skylake-X, completely different thing from "uncore" side. Where on Skylake-X it was enough to keep shadow tags of what is in L2, Willow Cove needs to go deeper and also handle cache lines that are L1, since that is no longer inclusive of L2.

 

[Edrick]

Willow Cove is not like Skylake-X, completely different thing from "uncore" side. Where on Skylake-X it was enough to keep shadow tags of what is in L2, Willow Cove needs to go deeper and also handle cache lines that are L1, since that is no longer inclusive of L2.

I think it is safe to say that Willow Cove, in its current form, is only going to be a mobile product. Both desktop and server will go from Sunny Cove (or Cypress Cove) directly to Golden Cove for SPR and ADL. Sure intel could release SPR with a modified Willow Cove like originally planned, but I doubt it since it will be a 2022 product.

 

[AMDK11]

Nope. Skylake-SP already had that L2 non-inclusive in L3 structure, but L1 contents were inclusive in L2. Icelake-SP will keep that structure as WillowCove is simply a different core, with different coherency controller that is aware of new cache hierarchy. So that is not the reason for WillowCove being MIA in servers. I think real reasons has to do with timing, 10nm, Icelake-SP already being in pipeline for half of decade.

By saying "the same" I meant that for Skylake-S and Skylake-X / SP the base is the x86 Skylake core, as well as for Iceland, Rocketlake, Iceland-SP and Tigerlake, the base is x86 SunnyCove, which Intel itself does not hide. Switching from Iceland-SP to Tigerlake-SP would not do much in certain loads and it would not be enough for Zen3. Anyway, Intel itself claims that with WillowCove they had two options:
1. Increase the IPC
2. Increase the clock
They chose the second option. Personally, I think WillowCove does little to SunnyCove / CypressCove only changes the overall load balance where it performs better but at the expense of others (worse) or comparatively with a much larger x86 + L2 + L3 core surface. Nevertheless, changes to the structure of the cache system in WillowCove are the foundation for future generations (GoldenCove) and indicate the direction that Intel has already taken.

 

[naukkis]

They chose the second option. Personally, I think WillowCove does little to SunnyCove / CypressCove only changes the overall load balance where it performs better but at the expense of others (worse) or comparatively with a much larger x86 + L2 + L3 core surface. Nevertheless, changes to the structure of the cache system in WillowCove are the foundation for future generations (GoldenCove) and indicate the direction that Intel has already taken.

As their cache-structure now wastes much more space for same performance reason why they changed it can be speculated - did they change their cache-system to prevent side-channel leaks from their previously used cache scheme?

 

[AMDK11]

As their cache-structure now wastes much more space for same performance reason why they changed it can be speculated - did they change their cache-system to prevent side-channel leaks from their previously used cache scheme?

I think it is quite possible that the WillowCove cache subsystem is a data leak remedy. WillowCove and GoldenCove have a cache subsystem that is very similar to all Zen generations (1, 2, and 3), and possibly this makes much less vulnerable to attack.

 

[jpiniero]

While I don't doubt it, is there anything to suggest that Golden Cove is using non-inclusive L2 and L3?

 

[AMDK11]

While I don't doubt it, is there anything to suggest that Golden Cove is using non-inclusive L2 and L3?

Some time ago there was an entry in the Sisoftware benchmark of the Alderlake sample which shows that the GoldenCove cache subsystem is non-inclusive as well as WillowCove.

WillowCove L1-I 32KB, L1-D 48KB, L2 1.25MB and L3 3MB

GoldenCove L1-I 32KB, L1-D 48KB, L2 1.25MB and L3 3MB

 

[DrMrLordX]

Why are we so focused on "Halo" products or their pricing? 11900K could cost 666 eur - i don't care, what is important is 11400 priced at say 180 eur and decent perf and features. What about 10600K with price that is 100 eur lower than 5600X ? AMD reacting and dropping prices or releasing additional SKUs - even better for us customers!

The price of the 11900k will have an influence on everything below it in the Rocket Lake-S hierarchy. Also bear in mind that the 11900k will only be 8c, so in a sense the 10900k may still be the "halo" product on Z490.

At the present, the "street" scalper price on a 5600x is ~$400 and up (I see plenty for $420). A 10600k is already $130-$150 lower than that. So what you are imaginging already exists.

Willow Cove is not like Skylake-X, completely different thing from "uncore" side. Where on Skylake-X it was enough to keep shadow tags of what is in L2, Willow Cove needs to go deeper and also handle cache lines that are L1, since that is no longer inclusive of L2.

Skylake-X also has more AVX units.

 

[JoeRambo]

At the present, the "street" scalper price on a 5600x is ~$400 and up (I see plenty for $420). A 10600k is already $130-$150 lower than that. So what you are imaginging already exists.

I meant 11600K. 10600K has to be cheaper due to performance and feature deficit. But if 11600K has perf/feature parity and is 100 eur cheaper? Why should person buying it care about 11900K pricing?

Also bear in mind that the 11900k will only be 8c, so in a sense the 10900k may still be the "halo" product on Z490.

I leave caring about those things to Intel marketing department, if they released 1Ghz, 4C, 200W chip that is twice as fast as my current 10900K in ST ( and by definition ~as fast in MT as 8C Skylake ) - I'd be first in line to buy it.

 

[DrMrLordX]

I meant 11600K. 10600K has to be cheaper due to performance and feature deficit. But if 11600K has perf/feature parity and is 100 eur cheaper? Why should person buying it care about 11900K pricing?

That changes things a bit, though if the 11600k comes out at that price, a 10600k is going to be even cheaper . . . and maybe only ~10% slower at the same clockspeed.

I leave caring about those things to Intel marketing department, if they released 1Ghz, 4C, 200W chip that is twice as fast as my current 10900K in ST ( and by definition ~as fast in MT as 8C Skylake ) - I'd be first in line to buy it.

That isn't going to happen though. Rocket Lake-S isn't going to be anywhere near twice as fast as the Comet Lake-S, either in absolute ST performance or in ST performance at isoclocks.

 

[Hulk]

Intel released Skylake in August of 2015 and didn't release Sunny Cove until four years later, August of 2019. Besides their 10nm process problems they really exacerbated the position they find themselves in by basically stopping everything and going into full greed mode.

They stayed with the same architecture and process (except for tweaks) for four years! All the while the rest of the industry kept innovating and advancing. Now they find themselves behind in architecture, behind in process, and slashing prices on outdated parts.

Like IBM in the '80's they are completely misreading the marketplace.

I assume the didn't release new architectures because the normal pattern would be to release a new architecture with a process shrink to offset the increased number of transistors (die size). When they couldn't do that they should have continued to develop new architectures and dealt with the implications of the reduced profits to continue architectural development.

While to a certain extent they didn't have control over the process problems they did when it came to introducing new cores. We should have seen Willow Cove 3 years ago and we should be buying Golden Cove parts today.

It's pretty obvious that each bean counter in charge just wanted to put off the coming disaster "until after the next earnings report" instead of facing reality and proactively heading off where they are now, which is getting smashed by AMD and Apple in the CPU department.