Discussion Intel current and future Lakes & Rapids thread

[OriAr]

A very reasonable position to take. All I can say is that time will vindicate me. Indeed, if there's a new architecture day this year, it may not even take very long.



The core idea behind Ocean Cove was deemed unnecessary, hence why the team wasn't needed any longer. Future designs may inherit parts of the work on Ocean Cove, but they will not be Ocean Cove. Of course, this was all under BK, so draw your own conclusions about the wisdom of such a move.

Was Ocean Cove supposed to be a revolutionary design that Intel felt eventually wasn't needed? Or was it supposed to be a modified Golden Cove?

 

[lobz]

Was Ocean Cove supposed to be a revolutionary design that Intel felt eventually wasn't needed? Or was it supposed to be a modified Golden Cove?

The first.

 

[Ajay]

Wonder if we are starting to see some of the first moves made by Keller, et al? Intel won’t get ahead of AMD with incremental core changes alone. Especially, while AMD enjoys a process advantage.

 

[IntelUser2000]

The core idea behind Ocean Cove was deemed unnecessary, hence why the team wasn't needed any longer. Future designs may inherit parts of the work on Ocean Cove, but they will not be Ocean Cove.

Thanks, but it really tells us nothing.

The core that we have zero clue about is now cancelled, and replaced with another core that we have no idea about. And they are going to use the same code name.

In another news. Fanlesstech reports that two Tigerlake NUCs are coming 2H of this year.

 

[jpiniero]

In another news. Fanlesstech reports that two Tigerlake NUCs are coming 2H of this year.

There's your 10 nm desktop. Wonder why they wouldn't use Rocket Lake.

 

[uzzi38]

There's your 10 nm desktop. Wonder why they wouldn't use Rocket Lake.

In what'll probably seem like a complete 180 from me (yes I'm fully aware), I kind of doubt this is the desktop 10nm we'll see.

I could very well be wrong, but I do think we'll see a proper desktop chip at some point. Even if low volume.

 

[IntelUser2000]

There's your 10 nm desktop. Wonder why they wouldn't use Rocket Lake.

Why Rocketlake?

 

[jpiniero]

Why Rocketlake?

6 cores and higher frequency.

 

[IntelUser2000]

6 cores and higher frequency.

Also a much crappier GPU, which matters in such small form factors.

 

[Exist50]

Thanks, but it really tells us nothing.

The core that we have zero clue about is now cancelled, and replaced with another core that we have no idea about. And they are going to use the same code name.

In another news. Fanlesstech reports that two Tigerlake NUCs are coming 2H of this year.

Oh, I should clarify that even the Ocean Cove name itself is dead, hence my hope for another Architecture Day as validation. I was told that much, at least.

In any case, I think there's some undue (if understandable) pessimism about Tiger Lake and the state of 10nm in general. Doubt a 28W chip in a NUC is the most we'll see of 10nm desktop. Weren't there rumors of Alder Lake S or something? Though I guess that could also be 7nm.

 

[DrMrLordX]

Weren't there rumors of Alder Lake S or something?

In late 2021, yes.

 

[FriedMoose]

Also a much crappier GPU, which matters in such small form factors.

Rocket Lake uses a Xe iGPU chiplet.

 

[mikk]

Rocket Lake uses a Xe iGPU chiplet.

This is a rumor. Realistically RKL-S with 32 EUs is a monolithic chip just like Tigerlake and beyond that RKL-S is a full backport of Tigerlake-S 8+1. If there is a RKL variant with 96 EUs it could make sense but not for 32 EUs.

 

[uzzi38]

This is a rumor. Realistically RKL-S with 32 EUs is a monolithic chip just like Tigerlake and beyond that RKL-S is a full backport of Tigerlake-S 8+1. If there is a RKL variant with 96 EUs it could make sense but not for 32 EUs.

The word 'realistically' implies something that makes sense.

A backport is not realistic for a multitude of reasons. I'd actually argue it's a worst case scenario.

For Intel's sake, I don't want Rocket Lake to be a backport. I mean, just think of the implications if 2-3 years ago they felt like they needed to invest millions into backporting 10nm architectures. And I mean short-term and - most importantly - long term implications.

 

[uzzi38]

Completely unrelated note to the above: ADL-P has 96EUs. That idea that -P is -U and -H both looks to be a solid one.

https://twitter.com/x/status/1219653665285378055

 

[jpiniero]

A backport is not realistic for a multitude of reasons. I'd actually argue it's a worst case scenario.

Selling Skylake in 2022 is worse. And that's what's going to happen if they didn't backport (or give in and finally dual source at TSMC or Samsung).

 

[Ajay]

Selling Skylake in 2022 is worse. And that's what's going to happen if they didn't backport (or give in and finally dual source at TSMC or Samsung).

I'll fall off my chair if Intel actually produces a 'Core' CPU at a foundry.

 

[uzzi38]

Selling Skylake in 2022 is worse. And that's what's going to happen if they didn't backport (or give in and finally dual source at TSMC or Samsung).

It's not. I'm confident of it.

It'll be a meme for volume, but they'll pump something else out to try and tide over the 10nm flop.

 

[lobz]

Guys.
1: The TGLs in the NUCs are the same mobile chips that come out towards the end of the year in laptops, and not 10 nm desktop.
2: I thought almost everyone here agreed in the past weeks, that Rocket Lake is 14nm (that is 100%) and has 12th gen graphics (that is also 100%), therefore these 2 things rule out a monolithic design. EMIB is also not happening, so most likely we'll see a solution like on the Intel+AMD chip (where EMIB was used only between the GPU and the HBM).

 

[lobz]

This is a rumor. Realistically RKL-S with 32 EUs is a monolithic chip just like Tigerlake and beyond that RKL-S is a full backport of Tigerlake-S 8+1. If there is a RKL variant with 96 EUs it could make sense but not for 32 EUs.

Rumor or not, intel won't do the 12th gen graphics on 14nm, no matter how many EUs. Double-Backport-Double-Jeopardy-Double-Dynamite design

 

[beginner99]

so most likely we'll see a solution like on the Intel+AMD chip (where EMIB was used only between the GPU and the HBM).

That would be terrible because that would take away at least 4 if not 8 lanes of pcie which is already rather scarce on intels platform. I don't see it. You get less cores and less pcie lanes vs comet-lake and AFAIK it's a desktop product? Who the hell would buy it over cometlake? That can only work if has pcie4 and is willow-cove based.

 

[coercitiv]

The word 'realistically' implies something that makes sense.

A backport is not realistic for a multitude of reasons. I'd actually argue it's a worst case scenario.

It's not. I'm confident of it.

It'll be a meme for volume, but they'll pump something else out to try and tide over the 10nm flop.

Every information we have so far points to RKL being a backport: lower than CML core count, better AVX capabilities.

You should know this, you brought the leak translation on the forums:

A friend posted a full translation:

"Apologies, made mistakes copying the spec. RKL-UP3/S should be AVX-512."

"In plain language: RKL = 14nm version of TGL, minus iTBT, with weakened IGD"

"Plus changed VRM scheme to SVID."

This practically confirms a Willow Cove backport.

 

[uzzi38]

Every information we have so far points to RKL being a backport: lower than CML core count, better AVX capabilities.

You should know this, you brought the leak translation on the forums:

Yeah, I did. Doesn't mean it logically makes sense though, does it?

The way I see it, back then Zen was a non-issue - nobody expected anythibg even remotely like Zen 2, ARM had (and really, have) no chances of being relevant in the desktop/mobile space for years to come, and why bother when 7nm should be ready in 2022-2023.

You know what, I'd be interested to hear what you lot have to think as to why such a backport would have been conceived. I can only think of one reason why it would have been, and I really hope it's not the case.

 

[Gideon]

The way I see it, back then Zen was a non-issue - nobody expected anything even remotely like Zen 2

Intel hired a number of AMD people 2-3 years ago (including Jim Keller) and probably did so annually previously. I'd think they could have been quite aware of what Zen 2 was bringig, if they cared. Now I'm sure the management would not have listen prior to 2017's Zen 1 release, but are you certain this was the case after? After all Zen 1 was almost certainly the trigger for 6-, 8- and later 10- core skylake processors.

The oddest thing about this backport to me is, why just the desktop CPUs? They knew that they need Cooper-Lake for servers as well (in parallel to Ice-Lake to fill the volume) why not backport on the server side as-well if they already go through the effort.

EDIT:
As for the GPU - a separate chiplet for only a 24EU GPU? That would be incredibly stupid considering how small Tiger-lake's 96EUs are on 10nm.

 

[uzzi38]

Intel hired a number of AMD people 2-3 years ago (including Jim Keller) and probably did so annually previously. I'd think they were quite aware of what Zen 2 might bring. I'm sure the management didn't care about AMD before Zen 1 release, but are you certain this was the case after?After all Zen 1 was almost certainly the trigger for 6-, 8- and later 10- core skylake processors.

The oddest thing about this backport to me is, why just the desktop CPUs? They knew that they need Cooper-Lake for servers as well (in parallel to Ice-Lake to fill the volume) why not backport on the server side as-well if they already go through the effort.

1. Rocket Lake is for mobile too actually.

2. They weren't away of what Zen 2 would bring. Not fully anyway. They would have only really known near the end of 2018 when Zen 2 started sampling, which isn't quite enough time to get a full backport out for early 2021.

Also, look at Raven Ridge. Why would Intel have thought the mobile sector was in danger? Similarly, Zen 2 is first and foremost designed with servers in mind... why would Intel focus a counter to Zen 2 on mobile and desktop?

Oh, and as an aside:

I'm confident on the fact that it would have taken 2-3 years to fully work through a backport. I heard it from a couple designers I know, and I asked another completely different one as well. Such a backport isn't one that could have been done easily.