Discussion Intel current and future Lakes & Rapids thread

[desrever]

Seems like Intel will be doing in 2 years what AMD is doing now. Intel won't have an IO die implementation until 2024. AMD will have dense x86 server market for at least a year before intel's e core only server cpus come out. By that time, it could be Zen 5c that intel has to compete with.

All of that is if Intel even delivers on the roadmap. I have doubts that they will even hit 2024 launch for Intel 3 server parts since "new core" and "new process" seems like a renaming of the successor to me. Shelving the old Granite rapid and launching a successor a year later reeks of desperation and the pressure to execute that is going to be huge, usually when companies tries to cram too much into 1 generation in engineering, they end up with something that doesn't work at all.

 

[Hitman928]

Guys we are getting some numbers from Sapphire Rapids with and without HBM memory.

Intel Claims Sapphire Rapids-SP Supercharged HBM Xeon CPUs Over 2x Faster Than AMD EPYC Milan & Milan-X Chips

Intel has revealed new benchmarks of its upcoming data center products including the HBM-Supercharged Sapphire Rapids-SP Xeon CPUs.

wccftech.com

View attachment 57562


But those numbers don't add up according to Microsoft test on the same Benchmark OpenFOAM Motorbike 28M cells

View attachment 57563

For 1 VM the Milan numbers match up (for Milan and MilanX anyway). In the footnotes, Intel used Microsoft's 1 VM performance for their own chart and ran their own tests for the SPR, ICL, and Milan. Intel's builds for their CPUs enabled AVX-512 so if this benchmark can effectively use AVX-512 and scale across many cores (and if so it most likely also likes a lot of memory bandwidth), then Intel's numbers are believable. Need 3rd party reviews of course, but they don't seem out of line given the work load and setups.

 

[Joe NYC]

For 1 VM the Milan numbers match up (for Milan and MilanX anyway). In the footnotes, Intel used Microsoft's 1 VM performance for their own chart and ran their own tests for the SPR, ICL, and Milan. Intel's builds for their CPUs enabled AVX-512 so if this benchmark can effectively use AVX-512 and scale across many cores (and if so it most likely also likes a lot of memory bandwidth), then Intel's numbers are believable. Need 3rd party reviews of course, but they don't seem out of line given the work load and setups.

Is that what Intel did? Got account on Azure to test Milan-X?

 

[Hitman928]

Is that what Intel did? Got account on Azure to test Milan-X?

No, they literally used Microsoft's numbers from Microsoft's test.

Edit: In the notes, all the non-Milan X score notes are prefaced with "Test by Intel" but the Milan-X score notes are "Test by Microsoft".

 

[tamz_msc]

OpenFOAM CFD is almost totally bound by memory bandwidth, so it's not surprising to see SPR doing well in it.

 

[Saylick]

Seems like Intel will be doing in 2 years what AMD is doing now. Intel won't have an IO die implementation until 2024. AMD will have dense x86 server market for at least a year before intel's e core only server cpus come out. By that time, it could be Zen 5c that intel has to compete with.

All of that is if Intel even delivers on the roadmap. I have doubts that they will even hit 2024 launch for Intel 3 server parts since "new core" and "new process" seems like a renaming of the successor to me. Shelving the old Granite rapid and launching a successor a year later reeks of desperation and the pressure to execute that is going to be huge, usually when companies tries to cram too much into 1 generation in engineering, they end up with something that doesn't work at all.

I agree with everything you've said; Intel has a tall order ahead of them and it will be one of those "see it to believe it" moments, as was the case when Intel first laid out their "4 nodes in 5 years" roadmap. The long-term vision is aggressive, and they have to be in order to catch up, but it doesn't help when their execution track record has been shaky at best. AMD has been executing flawlessly (not that the products launched flawlessly) on a roadmap that involved a lot of technical firsts, nonetheless (e.g. MCM, chiplets). AMD has credibility on their side. Intel? Not so much.

 

[IntelUser2000]

Intel won't have enough volume on Intel 4 to serve the mobile and desktop market. Canning Granite Rapids in 2023 and moving it to 2024 will open up some wafer supply, but still. The old rumour was Arrow Lake on TSMC N3. We may see Meteor Lake on TSMC N3 instead.

Arrowlake might still use TSMC N3 for some tiles. Graphics come into mind. But compute is Intel 20A.

 

[Exist50]

Arrowlake might still use TSMC N3 for some tiles. Graphics come into mind. But compute is Intel 20A.

I'm rather curious how that all works out. Seems like the original plan was to use TSMC for a late '23 launch, but they've been pushing products back onto Intel's process. The compute tile would be more significant, but also riskier than a graphics tile. Are they confident enough to commit to a new Intel process?

 

[Joe NYC]

Arrowlake might still use TSMC N3 for some tiles. Graphics come into mind. But compute is Intel 20A.

So, possibly, 2 troubled nodes. Hopefully, at least the IO tile will be solid.

 

[Exist50]

Few miscellaneous things I've noticed.

1) Meteor Lake is powering on in Q2, and they seem to no longer be promising H1 '23 release. With that in mind, and the fact that Intel's never managed a 4Q cadence of this sort before (hell not 5Q either), it's almost certainly delayed to Q3, if not later. Not surprised, given some of my previous remarks on the project.

2) Between this depiction of Arrow Lake as identical to Meteor Lake.



And this with the known MTL topology on the left:



I think it's safe to say that Arrow Lake == Meteor Lake with updated compute and graphics.

3) Combining that last image with this one showing the GNR topology on the right -



- it looks like they're using actual products as at least the basis for these images. What then would the ones under 18A be? Lunar Lake or Nova Lake for client? Diamond Rapids for server? That hub and compute tile approach looks interesting, though what would the narrow dies be?

 

[DrMrLordX]

Arrowlake might still use TSMC N3 for some tiles. Graphics come into mind. But compute is Intel 20A.

Yeah their products are gonna be a mixed up hodgepodge starting in 2023. I'm not holding my breath waiting for 20A to be ready by 2024 though! It's 2022 and we haven't even seen commercial availability of Intel 4 yet.

 

[jpiniero]

Are they confident enough to commit to a new Intel process?

That's why I think they are using TSMC as a second source. Better that than having to refresh Raptor Lake a bunch of times.

 

[DrMrLordX]

That's why I think they are using TSMC as a second source. Better that than having to refresh Raptor Lake a bunch of times.

Wasn't there an internal plan at Intel to make Meteor Lake 100% N3 until management pushed an Intel 4-based design instead? It's not like they couldn't revert to that.

 

[eek2121]

As usual, all vendor-provided benchmarks should be taken with a grain of salt. Especially from unreleased hardware.



Pretty sure 3rd Gen Xeon Scalable is IceLake-SP. That MS benchmark features either Skylake-SP or Cascade Lake-SP.



Intel won't have enough volume on Intel 4 to serve the mobile and desktop market. Canning Granite Rapids in 2023 and moving it to 2024 will open up some wafer supply, but still. The old rumour was Arrow Lake on TSMC N3. We may see Meteor Lake on TSMC N3 instead.

I'm rather curious how that all works out. Seems like the original plan was to use TSMC for a late '23 launch, but they've been pushing products back onto Intel's process. The compute tile would be more significant, but also riskier than a graphics tile. Are they confident enough to commit to a new Intel process?

Intel will be using TSMC for the GPU. They will not have volume issues.

Few miscellaneous things I've noticed.

1) Meteor Lake is powering on in Q2, and they seem to no longer be promising H1 '23 release. With that in mind, and the fact that Intel's never managed a 4Q cadence of this sort before (hell not 5Q either), it's almost certainly delayed to Q3, if not later. Not surprised, given some of my previous remarks on the project.

2) Between this depiction of Arrow Lake as identical to Meteor Lake.

View attachment 57595

And this with the known MTL topology on the left:

View attachment 57596

I think it's safe to say that Arrow Lake == Meteor Lake with updated compute and graphics.

3) Combining that last image with this one showing the GNR topology on the right -

View attachment 57597

- it looks like they're using actual products as at least the basis for these images. What then would the ones under 18A be? Lunar Lake or Nova Lake for client? Diamond Rapids for server? That hub and compute tile approach looks interesting, though what would the narrow dies be?

tick tock tick tock. ADL-S is a tick. RPL-S is a tock, MTL-S is a tick…

The ticks will have major changes, the tocks will have minor changes.

 

[jpiniero]

Wasn't there an internal plan at Intel to make Meteor Lake 100% N3 until management pushed an Intel 4-based design instead? It's not like they couldn't revert to that.

I thought that was what Arrow Lake was.

Intel will be using TSMC for the GPU. They will not have volume issues.

Using TSMC for the IGP is dumb, but shouldn't rule it out. You do still have the issue of how many N3 wafers could Intel get once TSMC starts on the 2023 iPhone.

 

[eek2121]

I thought that was what Arrow Lake was.



Using TSMC for the IGP is dumb, but shouldn't rule it out. You do still have the issue of how many N3 wafers could Intel get once TSMC starts on the 2023 iPhone.

No it isn't. They developed their IP on TSMC. It costs lots of time and money to switch fabs.

EDIT: Ian announced he is leaving AnandTech. RIP.

 

[lobz]

https://twitter.com/x/status/1494432860383715331

No Arrow Lake in H2 2023. Because of that I can't rule out a desktop version of Meteor Lake. Actually they have to release a desktop version, otherwise they have no chance.

https://twitter.com/x/status/1494448784440496137

20A for Arrow Lake?

Dude, on the same day and quasi in the same presentation they literally say about Ponte Vecchio 'on track' for Aurora supercomputer.

But 20A is magically appearing in 2 years. Okey dokey.

 

[DrMrLordX]

Intel will be using TSMC for the GPU. They will not have volume issues.

That's just the iGPU. They still need compute tiles.

I thought that was what Arrow Lake was.

It was rumoured that Arrow Lake would be entirely on N3, but based on the roadmaps that Intel has been releasing over the past week, it looks like they've pushed Arrow Lake out and are trying to put it on Intel processes. Basically Intel 4 and 3 look to be low-volume, and it appears that they're trying to push through to 20a (which may coincide with them finally getting some major shipments of EUV equipment from ASML).

That being said, the idea that Intel is going to go full-production on 20a in 2024 seems very optimistic.

EDIT: Ian announced he is leaving AnandTech. RIP.

Probably merits its own thread. And yeah, ugh. Anandtech is sinking.

 

[dullard]

So this is how Intel is delivering "5 nodes in 4 years".

Intel 7: Alderlake
Intel 4: Meteorlake and Custom networking ASIC in 2023
Intel 3: Granite Rapids and Sierra Forest in 2024
Intel 20A: Arrowlake in 2024, Diamonds Rapids in 2025
Intel 18A: Lunar Lake, Future server and Foundry Customer

Intel 3 and 18A are half nodes, or plusses. The server skips Intel 4 because PC will use them. PCs skip Intel 3, and is pretty much a server node.

That is quite a change from past rumors. Especially for Arrow Lake. Rumors were flying about both TSMC N3 and Intel 4 for Arrow Lake. Putting it on 20A would explain the delay from late 2023 to 2024. Lunar Lake going from the rumored Intel 3 to 18A is skipping a node. The rumors were way off!

 

[eek2121]

That's just the iGPU. They still need compute tiles.



It was rumoured that Arrow Lake would be entirely on N3, but based on the roadmaps that Intel has been releasing over the past week, it looks like they've pushed Arrow Lake out and are trying to put it on Intel processes. Basically Intel 4 and 3 look to be low-volume, and it appears that they're trying to push through to 20a (which may coincide with them finally getting some major shipments of EUV equipment from ASML).

That being said, the idea that Intel is going to go full-production on 20a in 2024 seems very optimistic.

The compute tiles are very small without the GPU. Intel has plenty of capacity. I will try and provide more info later.

 

[lobz]

Oh God... I have to admit, I have JUST watched the rearview mirror video itself from Pat Gelsinger right now.

This is nothing short of terrifying if you hold Intel shares, and I also think, he's gonna do more harm to the company in the next 5 years than the last 3 CEOs together.

Edit: he's gonna achieve that in the next 5 years, DESPITE the fact that technologically Intel will be in a 2 to 3 times better position the whole time than before.

 

[Joe NYC]

That's why I think they are using TSMC as a second source. Better that than having to refresh Raptor Lake a bunch of times.

I don't think Intel has a compute die on TSMC anywhere on the roadmap. And if they did, TSMC could not provide the volume.

Instead, Intel is using TSMC on frivolous things, like GPU (dGPU and iGPU tile).

 

[Joe NYC]

Wasn't there an internal plan at Intel to make Meteor Lake 100% N3 until management pushed an Intel 4-based design instead? It's not like they couldn't revert to that.

But TSMC N3 is so bad TSMC already announced a replacement for it, some time down the line.

 

[dark zero]

I saw the news about the Raptor Lake in Notebookcheck and seeing that they are going 24 core and 32 thread, they will retain Big Octa and going full 16 efficient cores.

My question is... the config of those efficient cores will be on Quad Cores or will move to Octa Cores?

If the latter happens (being in the configuration of Octa Cores), a potential Raptor Lake -N might be at minimun octa? And considering the improvement it might get, it will definately put any Big Dual Core solution (aka current Celeron and Pentium) out of the bussiness unless it includes efficient cores.

Also, seeing a potential 16 core Raptor Lake -N would be brutal to say the least since would be on the Kabylake tier. That jump would be on par of the Conroe jump Intel pulled a lot of time ago.

What I mean with this post is to think... Intel is pulling big jumps since Alder Lake and seems that they won't stop, but... what might be their limits on that?

 

[DrMrLordX]

And if they did, TSMC could not provide the volume.

TSMC is allegedly providing 40 kwpm of N3 to Intel. Meanwhile, Intel will (supposedly) have half that capacity for Intel 4. Intel 3 is an unknown, but it is likely that any Intel 3 wafers will come at the expense of Intel 4 wafers.