[Tom's Hardware] Intel Xeon 7 ‘Diamond Rapids’ CPUs officially launching in 2027 on Intel 18A-P — next-gen P-core Xeon features PCIe 6.0, 50% higher core counts, and twice the memory bandwidth
Posted by Noble00_@reddit | hardware | View on Reddit | 45 comments
Geddagod@reddit
Idk if Intel officially ever said DMR would launch in 2H 2026, but this is late compared to Intel's general \~2 year cadence. I have no idea what is going on in DCAI, but it's insane how screwed the roadmap is.
The messaging about core counts is pretty weird. Dell also claimed DMR would have 50% more cores, but also officially claimed DMR would have 256 cores. It also sounds like it's going to be later than Q1 2027.
If it is 192 cores, I'm not sure how they plan on competing with Venice with a core count disadvantage, a node disadvantage, while launching like half a year later, if not more.
igenicoOCE@reddit
Dell made a typo, it was always 192
Exist50@reddit
Not a typo. They did intend to have 256c at one point.
CopperSharkk@reddit
yields must be so bad if they are disabling 16 cores per tile.
Geddagod@reddit
That's the semianalysis theory lol
Exist50@reddit
I don't think that's it. Just for whatever reason, they decided to scale it back.
And the construction should be quite resilient to 18A yields.
CopperSharkk@reddit
the render in the article has it as 4 big compute tiles. is it actually smaller 12 core cpu tiles like CWF?
Geddagod@reddit
Rip
rezaramadea@reddit
Without SMT, Venice would lap DMR easy. You should've focused your excitement for Coral Rapids, as they're planning to re-introduce SMT & address the abandonment of 8-channel Mainstream Server Market. With Venice SP7 & SP8, & Verano, I think AMD won't need to worry until Coral.
Geddagod@reddit
I agree. And I believe even iso core count and without SMT, Venice would still beat
I don't think I've been excited for DMR since like the past year lol.
The number of comments you would find from me repeating the Intel CEO quote telling investors that DMR won't be competitive would be very high.
Frankly I don't think AMD would be too worried even then, unless they manage to get unified core in coral rapids (which I don't think they will).
Intel pushing to keep DCAI internal will mean they will very likely be behind on nodes, by maybe even a full node jump, at least till post 14A. Though the arch side should be, hopefully, more than just competitive when they manage to get unified core into DCAI products.
6950@reddit
Geddagod@reddit
Both arch and node are a problem.
6950@reddit
Node is fine it's arch issue for the most part of DMR
Geddagod@reddit
When one looks at products "iso node" from Intel, they manage to compete fine. SPR/EMR vs Milan, Genoa and Turin iso core count vs GNR.
Problem is they get completely outclassed by the leading edge node product that AMD has out at the same time. Turin dense during GNR, Genoa during SPR/EMR.
I think DMR 192T will be better than Turin 192T with SMT off. CWF on 18A also manages to match Turin Dense in total nT perf. On the same node, Intel manages to compete fine.
6950@reddit
Well DMR is on 18AP and it's like 9% improvement so like few % behind N2P? It should have been 256C they even cut the SMT from the original 192C/384T Config I don't get why is it only 192C/T
Geddagod@reddit
I mean you know how I feel about Intel's perf/watt claims, especially for sub node shrinks lol.
If you actually start multiplying out Intel's perf/watt claims for their nodes, you start getting very weird results, very fast.
6950@reddit
I known that but Intel's 2nd gen node usually fixes the problems with the first gen node.
SirActionhaHAA@reddit
It was rumored for late 2026 then there were rumors about a delay to 2027. Venice is launching in july, dmr would be at least 3quarters behind it and be slower. It'd be horrible if not for ai companies buying up every cpu they can get their hands on.
Geddagod@reddit
Intel managed to luck out tremendously here. The market share losses that would have occurred if the AI boom never happened, or the AI boom didn't shift significant resources away from GPUs and back into CPUs recently, would have been immense.
anhphamfmr@reddit
it's weird to me that people can claim with absolute certainty about products that don't even exist. E.g Intel 18A-P vs Tsmc 2nm. Not sure how they can claim 18AP to be an inferior node when there isn't even a single Tsmc 2nm chip in production yet, while Intel 18A chips are selling like hot cake.
Geddagod@reddit
That's why I say "I'm not sure how..."
If you are talking about specific nodes....
This is exactly how I can claim, with very high certainty, that N2 will have a node advantage vs 18A-P.
Intel 18A chips might be selling like hot cakes (what little volume they have out), and have been reviewed. PTL's P-core freq/power curve is esentially the same as ARL on N3B. And PTL's P-core should have a ton of caveats that should make it look better. And it doesn't.
Unless you expect 18A-P to gain an entire nodes worth of PPA against 18A, then N2 will be better than 18A-P.
anhphamfmr@reddit
any source for this? based on what I read a few months ago the PTL P-core's IPC is quite similar to ARL's P core, but PTL achieves the same perf at noticeably lower power.
Geddagod@reddit
Linked in my comment right here. I also go a bit more into the 18A vs N3B comparison.
It is surprisingly better. Like high single digits low teens IIRC in specint2017. The memory latency improvement is very high relative to ARL.
The total perf/watt curve uplift is pretty good for PTL. Huang's data shows that.
However, Huang also points out that when you look at sub-benches in spec2017 that don't benefit much at all from PTL's higher IPC and run mostly in the core private caches, the perf/watt curves vs ARL and PTL are very similar. Showing that PTL's perf/watt uplift is not from the node, but the IPC uplift from the uncore improvements.
That's also shown by my second link in my linked comment, that shows simply a freq/perf curve. The curves between ARL and PTL are pretty much the same.
hakim37@reddit
You comment seems to be deleted.
A point on your uncored benchmark is that it compared single core workloads which I think would neutralise any uplift from power via as the system voltage would be too low for significant droop.
Also comparing 18AP to N2 I think it's going to be closer than you're expecting. Intel is claiming 9% performance uplift / 18% power efficiency which is basically the jump from N3E > N2
Geddagod@reddit
Nope, it's not lol. That comment is still there.
I linked the two PTL single core freq/watt tests further down this thread again though.
At Fmax the voltage of a P-core is pretty high. This wasn't tested exclusively at Vmin or anything.
Intel also claimed a 10-15% bump from Intel 10SF to Intel 7, all while claiming Intel 10SF had TSMC N7 perf/watt. Does that make Intel 7 almost equal N5 perf/watt?
Intel also claimed Intel 3 had a \~18% perf/watt bump over Intel 4. Intel 3 products have around N5 perf/watt. Is Intel 4 actually a N7 class node, or is Intel's arch team so bad that they made a N3 class node look like a N5 one?
This doesn't just apply to Intel either. Multiplying out TSMC's N4P claims would suggest it has almost the same perf/watt as N3E, and yet we can see the same cores on N3E have much better perf/watt than N4P counterparts than what just multiplying out the numbers would suggest.
Exist50@reddit
The gap is large enough that Intel felt forced to use it to compete. We're looking at the better part of a full node's worth.
hakim37@reddit
Intel probably chose N2 for a number of reasons. They had no idea where the 18a family would land and so would need to hedge their bets with N2. Also I'm comparing 18AP to N2 so we're not talking Nova Lake anyways. Razor Lake still uses some N2P over 18AP but there will be full 18AP SKUs.
My point is Intel is slowly catching up with these node generations. 18a > ap is adding nearly a full node jump in capability. N2 > N2P is expected to be like 4%. When we get to 14A vs A14 I think Intel is expected to regain another 5% of the gap.
Exist50@reddit
Or consider that even if it did land exactly where they were hoping it to, that still wasn't good enough to compete with N2, and 18A only fell further behind since. Remember, despite lying about this to the public, Intel knew that p1278 was a shitshow since at least ARL-20A.
If they were merely "hedging their bets", why did they go all-in on PTL, only to backtrack with the next gen, when everything about 18A should have been long-since ironed out?
By the time 18AP is available, so will N2P, btw. But whichever pair you want. Doesn't change things.
Yes, so even with RZL, N2P is commanding a sizable lead vs 18AP. The only 18AP SKUs we see seem to be NVL rebrands, or maybe one new tapeout using NVL IP. Unlike NVL, RZL IP seems to be exclusively on N2P, which is arguably moving even further away from internal nodes.
Quite frankly, you're maintaining realistic expectations for TSMC while treating Intel slides as gospel, despite a much worse track record of accuracy. Maybe the gap will close eventually, but if so, it's not happening with 18A.
hakim37@reddit
I'm not trying to downplay TSMC they obviously are exceptionally talented I'm only citing both companies statements on what they expect from the next node jumps. TSMC states 5% (I incorrectly remembered 4%) and intel recently reiterated 9%.
In the A14/14a generation Intel is again predicting a 5% catch-up to TSMC. Although admittedly on base 18a/N2 so maybe that's a wash.
Geddagod@reddit
Before NVL launched, they would have used 18A on PTL, CWF, WCL, and then the precursor to this node on 20A for ARL. Using 18A-P on Nova Lake would have made this the third iteration of this node. DMR should have launched on this node in a similar time frame as well.
The full 18AP skus is only rumored to be a mid range bLLC sku, isn't it?
If you look at the high end client die, once again, Intel is rumored to be going external.
That's what they claim in perf/watt, but not in density.
I mean we will see. Intel made similar grand claims about 18A perf/watt before they had to scale back.
What they choose for HML is also going to be pretty telling.
Interesting-Rock2474@reddit
9% performance uplift [OR] 18% power efficiency
anhphamfmr@reddit
I can't find any link. only claims. Anh real source that can back your claims about the perf/watt curves of PTL vs ARL in spec2017 subbenches?
Saranhai@reddit
He doesn’t have any real source. He only makes claims and backs them up with more claims others have made 🤷🏻♂️
Geddagod@reddit
The claims "others have made" are backed by data they got with chips they tested on hand.
I'm confused, when you originally saw those exact same links a month ago you admitted you were wrong there too.
Saranhai@reddit
lol good, called me out that's good. I admit I blindly read your twitter sources and believed you knew what you were talking about. Unfortunately...yeah, linking videos or twitter posts making these claims really can't be considered as a real "source". I've learned from my mistakes.
Exist50@reddit
Intel themselves paying a premium to use it isn't proof enough?
anhphamfmr@reddit
it tells more about the production capacity than anything else. remember 18AP is an external node for other customers as well.
Geddagod@reddit
You see the links yet for 18A perf/watt yet? Lmao.
When the choice was made to use N2, Pat Gelsinger was CEO. You know, the dude who massively overbuilt fab shells for supply assuming 18A was going to get used for a lot more volume then Intel ended up needing.
Also the dude who kept on claiming that 18A would bring back node leadership before they had to down scale the node.
For pretty limited volume.
Exist50@reddit
No, they didn't pick N2 for lack of 18A capacity. That's complete and obvious nonsense. Remember this was when they were actively cancelling all their expansion plans.
SirActionhaHAA@reddit
imaginary_num6er@reddit
Has Intel actually launched any workstation CPUs on-time since Sapphire Rapids? It always sounds like their roadmap is never on-time and chips like Falcon Shores eventually get canceled.
996forever@reddit
Sapphire rapids was not on time by the original roadmap. They haven’t been on time since skylake SP.
Exist50@reddit
SPR was supposed to be a '21 product at one point. It was horribly delayed.
996forever@reddit
Their actual last truly on time architecture was ivy bridge.
AutoModerator@reddit
Hello Noble00_! Please double check that this submission is original reporting and is not an unverified rumor or repost that does not rise to the standards of /r/hardware. If this link is reporting on the work of another site/source or is an unverified rumor, please delete this submission. If this warning is in error, please report this comment and we will remove it.
I am a bot, and this action was performed automatically. Please contact the moderators of this subreddit if you have any questions or concerns.