Discussion Intel current and future Lakes & Rapids thread

[adroc_thurston]

AMD is not taking Intel market share anymore in CPU 🤣🤣

it's never been more over for me

 

[Geddagod]

AMD is not taking Intel market share anymore in CPU 🤣🤣

https://twitter.com/x/status/2014685594895438168

Prob not client esp after NVL launch but server results from AMD are going to be.... interesting.

Those have no customers.

Even if they don't it's still very weird Intel isn't saying anything....

While they mention Coral Rapids time and again they don't make any mention of
DMR-HD(RRF) or their E-core server roadmap.

Hmph

RTB(Rialto) was incremental/~30%? over PVC, even worse wrt timing.

Apparently PVC being recognized as one GPU was scuffed af so maybe RTB improved that situation

 

[adroc_thurston]

Prob not client esp after NVL launch

NVL is garbage with no 1t, DOA for 2027.

Even if they don't it's still very weird Intel isn't saying anything....

It's a dead product when it haez no customers.

Apparently PVC being recognized as one GPU was scuffed

Really hard to pretend it's a single GPU when the d2d link was like a few hundred GB/s.

 

[511]

Prob not client esp after NVL launch but server results from AMD are going to be.... interesting.

i don't wanna bet against inverse cramer he is a force of nature

 

[Geddagod]

i don't wanna bet against inverse cramer he is a force of nature

I actually got wooshed I didn't even realize it was Cramer who said it lol

 

[Geddagod]

NVL is garbage with no 1t, DOA for 2027.

It's a dead product when it haez no customers.

Really hard to pretend it's a single GPU when the d2d link was like a few hundred GB/s.

Yes I was informed of that a few weeks ago I did (edit:not) know the interconnect bandwidth was that poor originally

 

[adroc_thurston]

Yes I was informed of that a few weeks ago I did (edit:not) know the interconnect bandwidth was that poor originally

they just didn't talk about tile-to-tile interface until the 11th hour (close to Aurora acceptance).

 

[regen1]

Microscopic volume footprint.

Chinese seem to like it, majority of it went to them iirc. Good for video(web/media) for what matters to them. Even the 288 core version of SRF did go to them which wasn't generally available.

It was optimistically Turin-D perf at hella money.

Probably will win in some and lose in others, still could've been better than now it's coming when Venice is around the corner.

See, that's the thing, Graviton and friends are not really core spam platforms per se; they focus very much on scale up perf.
Ampere is dead for the same reasons Atom Xeons are dead.

Graviton 4 was easy for SRF and easier for CWF. CWF would be fine against Graviton 5. Atom/E-core Xeon even if dead could be temporary as it's supposed to be the main thing("Unified") in the future.

It gets vivisected by the SP7 Venice.
Never had any chances.

Doesn't always matter what AMD does. They have been winning for long but there's enough room for Intel, (ICL, SPR, EMR, GNR) Xeons still get sold plenty.
DMR-HD/RRF is way more complete than CWF, SRF wrt fabric, CPU perf, SIMD, etc. and it would've(theoretically) significantly better PPA than DMR-AP.

 

[regen1]

Hmph

Cancelling the nearby products(DMR-SP and more? for whatever reasons) and talking something up that's quite distant. All the Coral Rapid hype build up and there might be Zen 7 not too distant(timing-wise). Lot of the strategy seems dependent on CPU demand-rise and Nvidia-deal.

Apparently PVC being recognized as one GPU was scuffed af so maybe RTB improved that situation

RTB had same D2D as PVC iirc but have to check. Only Xe-Link B/W was doubled(same Xe-Link as PVC but fully enabled) plus other improvements. In any case RTB volume would've been in 2024, not much chance then( easy to forget how delayed PVC was, somewhat similar to SPR ).

Yeah, FLAT mode(2T) became default for PVC, composite mode(1T) wasn't much practical. Argonne(updated Jan 2026) gets very low Tile-to Tile(bidirectional) B/W.
All that fancy packaging and then this, just imagine!

 

[adroc_thurston]

Chinese seem to like it, majority of it went to them iirc. Good for video(web/media) for what matters to them. Even the 288 core version of SRF did go to them which wasn't generally available.

yeah but that's still meme volumes vs what Intel aimed for at Meta.

Probably will win in some and lose in others, still could've been better than now it's coming when Venice is around the corner.

yeah it's dead.
LBT ignoring it this last ER call was very indicative, kek.

Graviton 4 was easy for SRF and easier for CWF. CWF would be fine against Graviton 5.

Kind of.
Graviton3/4/5/6/whatever focus on per-core perf.
Socket-level throughput isn't all there is to a server CPU.

Doesn't always matter what AMD does.

of course it does. Intel is very honest about the comp situation.
hence why Coral TTM acceleration.

All the Coral Rapid hype build up and there might be Zen 7 not too distant(timing-wise).

it has to touch venice first.

 

[Joe NYC]

yeah it's dead.
LBT ignoring it this last ER call was very indicative, kek.

The non-response to question about Clearwater Forest was quite telling. Instead of answering, LBT listed what the priorities are (and not including Clearwater Forest in the list of priorities).

of course it does. Intel is very honest about the comp situation.
hence why Coral TTM acceleration.

If Coral was originally scheduled to be on 14A, then either Coral will have to (instead) target 18A or 14A needs some acceleration too. Or at minimum, no delays.

 

[regen1]

Graviton3/4/5/6/whatever focus on per-core perf.
Socket-level throughput isn't all there is to a server CPU.

Arctic Wolf's 1T perf should be ok. BoM wise lot better vs the humongous mono-dies of P-core. Again DMR-HD shouldn't be as bottlenecked as SRF/CWF wrt fabric in theory. Shouldn't judge Milan(Zen 3) at Rome's(Zen 2) release, necessarily.

of course it does. Intel is very honest about the comp situation.
hence why Coral TTM acceleration.

Nah, there are so many areas in server even aside from various perf parameters. AMD is not everything otherwise according to some GNR(and some previous Xeons) won't have meaningful demand which is very far from reality.

 

[regen1]

Prominent Intel Compiler Engineer Heads Off To AMD - Phoronix

www.phoronix.com

More jump off the Intel ship. May not be directly relevant to this person but seems to be a common feature of this leadership?

 

[adroc_thurston]

Arctic Wolf's 1T perf should be ok

Yeah but GV5 in particular pushes for per-core perf.

Nah, there are so many areas in server even aside from various perf parameters.

This does not apply to Venice and Venice-E.

May not be directly relevant to this person but seems to be a common feature of this leadership?

no, Intel is just next on the scooping list (every Intel gfx PE+ with "raytracing" in bio is now at AMD). AMD already pillaged ImgTec to the bone.

 

[DrMrLordX]

yeah it's dead.
LBT ignoring it this last ER call was very indicative, kek.

That's really quite sad. At least Clearwater Forest was supposed to reach shipping volumes. It was Rogue River Forest that was (effectively) canceled. Now even Clearwater is basically dead.

If Coral was originally scheduled to be on 14A, then either Coral will have to (instead) target 18A or 14A needs some acceleration too. Or at minimum, no delays.

18A-P maybe? It really depends on where Diamond Rapids falls doesn't it?

 

[Joe NYC]

18A-P maybe? It really depends on where Diamond Rapids falls doesn't it?

I wonder if Coral Rapids turns into minimalistic upgrade of Diamond Rapids with SMT. On 18AP

 

[adroc_thurston]

At least Clearwater Forest was supposed to reach shipping volumes

Death by hybrid bonding.
DMR is much the same.

 

[Joe NYC]

Death by hybrid bonding.
DMR is much the same.

That's one reason to keep watching for Clearwater Forest launch. If Intel packaging can handle hybrid bond in volume, and on time.

The official reason for the delay was "packaging problem" without specifying what part of packaging, but one would have to suspect hybrid bond.

 

[adroc_thurston]

If Intel packaging can handle hybrid bond in volume, and on time.

dawg they're already eons behind.
CWF was an H1'25 product once.

 

[DrMrLordX]

Death by hybrid bonding.
DMR is much the same.

That does not bode well for future products.

 

[adroc_thurston]

That does not bode well for future products.

They don't have HBI outside of server for a looooooong while.
Don't worry about it. Server they weren't alive to begin with.

 

[511]

That's really quite sad. At least Clearwater Forest was supposed to reach shipping volumes. It was Rogue River Forest that was (effectively) canceled. Now even Clearwater is basically dead.

Rouge River Forest is not cancelled it's renamed to DMR-HD as for HBI should have tried on a meme volume part that is relatively small.

 

[DrMrLordX]

Rouge River Forest is not cancelled it's renamed to DMR-HD

Why? And if Clearwater is effectively dead then why would they bother with Diamond Rapids-HD?

as for HBI should have tried on a meme volume part that is relatively small.

Such as? Where would they fit that into their lineup?

Don't worry about it. Server they weren't alive to begin with.

That has become all the more-evident.

 

[511]

Why? And if Clearwater is effectively dead then why would they bother with Diamond Rapids-HD?

Cause DMR-HD fixes the issue with Clear water forest of not having AVX-512 and botched Mesh/L3

Such as? Where would they fit that into their lineup?

Broadwell eDRAM Lakefield for foveros and Kaby Lake G for EMIB

 

[DrMrLordX]

Cause DMR-HD fixes the issue with Clear water forest of not having AVX-512

Was that really a problem for cloud compute? Likely not.

and botched Mesh/L3

You sure about that?

Broadwell eDRAM Lakefield for foveros and Kaby Lake G for EMIB

. . . nah.