Discussion Intel current and future Lakes & Rapids thread

[inquiss]

Relying on brand loyalty only, I would presume? Surely someone has a reason, and Tan probably had to ultimately approve it, so there must be a reason.

Could contracts lock companies into buying from the vendor what's on the roadmap? So they have to buy the 16ch platform. I have no idea how that part works.

Could they just bot have enough good people to design and validate both platforms? Seems dumb too.

Can't think of a reason.

 

[jpiniero]

Could they just bot have enough good people to design and validate both platforms? Seems dumb too.

Obviously they would cancel the 16 over cancelling the 8 if it was a resource issue.

 

[adroc_thurston]

Obviously they would cancel the 16 over cancelling the 8 if it was a resource issue.

Not really because this is where politics gets in play.

 

[Kepler_L2]

Could contracts lock companies into buying from the vendor what's on the roadmap? So they have to buy the 16ch platform. I have no idea how that part works.

Could they just bot have enough good people to design and validate both platforms? Seems dumb too.

Can't think of a reason.

Maybe it's a margins thing

[News] Intel Reportedly Demands 50% Margins for New Products in Profitability Push

Intel's new CEO, Lip-Bu Tan, has been working to turn the company around. According to Tom’s Hardware, Intel has stopped approving new projects unless...

www.trendforce.com

DMR-SP probably couldn't be competitive vs Venice SP8 at 50%+ margins, so they decided to cancel it instead of dragging company margins down.

 

[jpiniero]

Not really because this is where politics gets in play.

So they'll make enough Diamond-AP so send out to reviewers but since nobody will actually buy it, there won't be need to produce much?

They can't cancel everything; "Investors" will freak out.

 

[DrMrLordX]

High margins with near-zero volume is the wrong move.

 

[Tigerick]

CFO said:

But in admitting that it will take a “multi-year process” for Intel to make big competitive moves in the business, Zinser said the company’s upcoming “Diamond Rapids” server product line due for launch next year “doesn’t get us quite there.”

“In certain cases the performance is actually better, but in other cases it’s not. And so we’ve got more work to do to finally get to a place,” he said.

Zinsner called the successor to “Diamond Rapids,” code-named “Coral Rapids,” the “real opportunity” that will let Intel “take a really good step forward.”

Without SMT, DMR does not compete up to Venice SP8 32x4=128c/256T. Let's hope Intel speed up the launching of Coral Rapids which supposedly fabbed by 18A-P node.

 

[511]

CFO said:


Without SMT, DMR does not compete up to Venice SP8 32x4=128c/256T. Let's hope Intel speed up the launching of Coral Rapids which supposedly fabbed by 18A-P node.

DMR us 256C/256T

 

[Joe NYC]

Trying to strongarm margins maybe?

By upselling customers to more expensive, 16-channel platform?

 

[Joe NYC]

Obviously they would cancel the 16 over cancelling the 8 if it was a resource issue.

Intel probably has a contractual obligation to deliver 16 channel platform to NVidia.

 

[511]

Intel probably has a contractual obligation to deliver 16 channel platform to NVidia.

Those are separate the custom sku are not part of this cancellation same as the 288C SRF which was turned into a custom SKU.

 

[Joe NYC]

Those are separate the custom sku are not part of this cancellation same as the 288C SRF which was turned into a custom SKU.

That's kind of underwhelming (especially SRF being an older CPU), but I guess fine compared to NVidia's Grace Arm CPU.

 

[511]

That's kind of underwhelming (especially SRF being an older CPU), but I guess fine compared to NVidia's Grace Arm CPU.

Nvidia is not getting SRF they are getting DMR

 

[regen1]

https://phoronix.com/benchmark/result/xeon-6-latency-optimized-mode/geometric-mean-of-all-test-results-result-composite-x6lom.svgz

This is something I have been speculating for some time. Intel's new architecture seems to be choked by latency. When you free up the latency, it becomes less efficient; however, you gain a good bit of performance.

At lower clocks (which would be the norm in high count core implementations), 18A + lower latency may prove to be a potent combination for Intel's next gen NVL class.

It's a fabric/mesh optimization issue wrt design. There's way too much optimization headroom left. In many instances the gains are disproportionately high in "Latency Optimized Mode" but still improves Perf/W while consuming higher power than default.
DMR is better optimized in this regard(on paper at least) but the problem is decent volume of DMR is considerably later than the competition(Venice) and on top of that they have now cancelled DMR-SP.

 

[regen1]

The leadership is absolutely brain-dead for DMR-SP's cancellation.
Too many leadership and direction changes, excessive lay-offs and now we don't have a next-gen mainstream(-SP) Xeon.
No successor for GNR-SP(esp. for enterprise) for a long time even considering they won't cancel the mainstream line-up(-SP?) next gen.
This also would be a considerable blow to foundry/18A family, could affect foundry breakeven time adversely.
What a disaster. A self-inflicted harm for both "Products" and "Foundry".

 

[Win2012R2]

Seems like very dumb move, but in any case RAM/NAND pricing is going into stratosphere, their share of server BOM will be very high again, so in this situation server buyers might as well get the best next gen CPU to, which is AMD's Zen6.

If AMD gets to well over 50% CPU server share then I can't see Intel ever recovering from it, plus AMD will get sweet AI cash coming in.

It's over.

 

[adroc_thurston]

If AMD gets to well over 50% CPU server share

They'll get there anyway.
But this shart makes the job easier.

 

[DavidC1]

DMR us 256C/256T

It's incompetency holding them back, not feature checkmarks like SMT. Having SMT and core parity didn't get Granite Rapids competitive with AMD. And 256 cores over 192 core Venice should be able to cancel out most of the SMT advantage on the latter. They closed the gap with GNR but the amount might as well be insignificant. I mean, it was a big surprise considering Emerald Rapids did quite well considering what it was. I wonder if an 80 core EMR with a shrink to Intel 3 would have ended up being similar? But it just continues on the big picture, which is a downtrend for Intel.

What if there are legitimate reasons for cancelling -SP? Maybe it really is that uncompetitive and thus they decided to shift all resources into the successor, whereas the big -AP does much better.

 

[511]

It's incompetency holding them back, not feature checkmarks like SMT. Having SMT and core parity didn't get Granite Rapids competitive with AMD. And 256 cores over 192 core Venice should be able to cancel out most of the SMT advantage on the latter.

Tbh DMR is pretty gigantic upgrade over GNR even bigger than GNR was over EMR.

They closed the gap with GNR but the amount might as well be insignificant. I mean, it was a big surprise considering Emerald Rapids did quite well considering what it was. I wonder if an 80 core EMR with a shrink to Intel 3 would have ended up being similar? But it just continues on the big picture, which is a downtrend for Intel

EMR/GNR /SPR use variants of Golden Cove so it wouldn't matter much in terms of the core.
Their server roadmaps had delays upon delays for a while which is the real issues only since SRF/GNR did they launch on time.

 

[DavidC1]

Tbh DMR is pretty gigantic upgrade over GNR even bigger than GNR was over EMR.

EMR/GNR /SPR use variants of Golden Cove so it wouldn't matter much in terms of the core.
Their server roadmaps had delays upon delays for a while which is the real issues only since SRF/GNR did they launch on time.

Yes, but them cancelling -SP is a suggestion their problems are much bigger than they let out even in 2026+.

They can't be that stupid to just cancel -SP. I think it legitimately is that bad so they decided going all in on next gen was better over the long run especially considering their finances. It's not just core sucking for Intel. In the Netburst days, only the core sucked, everything else was fantastic. Nowadays the occasional mediocre core is hobbled down by horrible everything else. Core, fabric, memory, cache, I/O, software all sucks. This is why GNR sucks. Core count and core uarch alone, GNR should have been competitive against Turin.

 

[Joe NYC]

Nvidia is not getting SRF they are getting DMR

I think I misunderstood your post (mentioning SRF).

I did think that NVidia was getting 16-channel DMR, which I think could be (one of the) reasons for cancelling 8-channel DMR - to concentrate the resources.

 

[LightningZ71]

Unless Intel thinks two things: DMR 8ch can not be competitive with like products from other vendors in any way and can not be priced well enough to cover margins. Second, they have seen the writing on the wall in that market with possibly ARM, cloud services, or something else that is going to, or already is, taking significant volume in that segment.

Frankly, I just don't see a lot of mid-range servers anywhere around me these days. It's either tiny stuff for small offices, edge devices for specific industries, or just about the largest thing that they can get their hands on for the purpose of density. I don't know anything about HPC, so maybe that's something, but, other than that, I think they are just bowing out of a market in decline, at least with declining margins.

 

[adroc_thurston]

Frankly, I just don't see a lot of mid-range servers anywhere around me these days. It's either tiny stuff for small offices, edge devices for specific industries, or just about the largest thing that they can get their hands on for the purpose of density. I don't know anything about HPC, so maybe that's something, but, other than that, I think they are just bowing out of a market in decline, at least with declining margins.

Onprem dinosaurs exist and boy there are a lot of them.

 

[QuickyDuck]

2P today is like the 4P of yesterday. Looking forward to 1P DMR replace old 2P system.

 

[DavidC1]

Unless Intel thinks two things: DMR 8ch can not be competitive with like products from other vendors in any way and can not be priced well enough to cover margins. Second, they have seen the writing on the wall in that market with possibly ARM, cloud services, or something else that is going to, or already is, taking significant volume in that segment.

That's why I think they are abandoning it to focus on the next generation, which they see some hope of competing. They did that with Netburst.

The 65nm Presler seemed mostly abandoned in 2006. In fact because they were preparing for Core 2 later that year. And in that case it was the right decision. The company was stable back then though. We'll see whether cancelling -SP means further digging of the hole or Coral Rapids is indeed really good.