Intel "Diamond Rapids" Xeon CPU to Feature up to 192 P-Cores and 500 W TDP

[AleksandarK]

Intel's next-generation "Oak Stream" platform is preparing to accommodate the upcoming "Diamond Rapids" Xeon CPU generation, and we are receiving more interesting details about the top-end configurations Intel will offer. According to the HEPiX TechWatch working group, the Diamond Rapids Intel Xeon 7 will feature up to 192 P-cores in the top-end SKU, split across four 48-core tiles. Intel has dedicated two primary SKU separators, where some models use eight-channel DDR5 memory, and the top SKUs will arrive with 16-channel DDR5 memory. Using MRDIMM Gen 2 for memory will enable Intel to push transfer rates to 12,800 MT/s per DIMM, providing massive bandwidth across 16 channels and keeping the "Panther Cove" cores busy with sufficient data. Intel planned the SoC to reach up to 500 W in a single socket.

As one of the first mass-produced 18A node products, Diamond Rapids will be the first to support Intel's APX, also featuring numerous improvements to the efficiency of AMX. Intel also plans to embed native support for more floating-point number formats, such as NVIDIA's TF32, and lower-precision FP8. As most of the world's inference is good enough to run on a CPU, Intel aims to accelerate basic inference operations for smaller models, enabling power users to run advanced workloads on CPUs alone. With a 1S, 2S, and 4S LGA 9324 configuration, Diamond Rapids will offer 768 cores in a single server rack, with a power usage of only 2000 W. Supporting external accelerators will be provided via the PCIe Gen 6 connector. Scheduled for arrival in 2026, Intel could time the launch to coincide with its upcoming "Jaguar Shores" AI accelerators, making a perfect pair for a complete AI system.



View at TechPowerUp Main Site | Source

 

[Daven]

It’s good to see that Intel will keep up with AMD.

AMD Namedrops EPYC "Venice" Zen 6 and EPYC "Verano" Zen 7 Server Processors | TechPowerUp}

Both companies will have close to the same specs next gen.

 

[Squared]

This is the first 18A product? That means it truly is 18A and not 20A rebranded as 18A. And it's yielding well enough to produce many large dies. And Panther Lake is perhaps a little further away.

 

[N/A]

48 P core tile on 18A sounds great. Why can't we have that on socket 1954.

 

[Daven]

This is the first 18A product? That means it truly is 18A and not 20A rebranded as 18A. And it's yielding well enough to produce many large dies. And Panther Lake is perhaps a little further away.

Since there never was a 20A product, we can never know if 18A is just a rebrand. The best we can ever know is how well 18A compares to N2 when if ever Intel releases a product using it.

 

[mb194dc]

Seems like we're in another Intel news flow stock pumping phase... Let's see the chips in the real world and we'll talk. Same as the Intel x3d hype competing chips in 2026...

Usual pattern is lots of hope, then it turns out their latest manufacturing process is a disaster and they're going to lose tens of billions more...

 

[agent_x007]

Intel problem in server space isn't performance, it's price you have to pay for that performance (power bill and cooling included).
BUT maybe they surprise me this time and deliver a reasonably priced alternative to EPYC... maybe...
(reality check : this big thing can be many things, but I highly doubt it's going to be cheap - LGA 9324 )

 

[bonehead123]

this big thing can be many things, but I highly doubt it's going to be cheap

"Cheap is as cheap does"

And this AINT that, in any way, shape or form....

Get ready to fire up those massive sub-zero, LH2-powered cooling rigs, which, ICYMI, aint exactly cheap either, hahahaha

 

[ZoneDymo]

It’s good to see that Intel will keep up with AMD.

AMD Namedrops EPYC "Venice" Zen 6 and EPYC "Verano" Zen 7 Server Processors | TechPowerUp}

Both companies will have close to the same specs next gen.

I mean nothing here says they can touch AMD tbh, we will have to see actual benchmark results.

 

[Daven]

I mean nothing here says they can touch AMD tbh, we will have to see actual benchmark results.

I think both companies will have 16 channel memory and 192 of the full cores. I think the 256 core Zen 6 are the dense cores.

 

[Tek-Check]

Intel problem in server space isn't performance, it's price you have to pay for that performance (power bill and cooling included).

Intel's problem currently in server market absolutely IS performance. There is no doubt about it. You just need to follow industry benchmarks. In 2P configuration, EPYCs are whopping 40% faster while using significantly less power and having lower cost of ownership.

Even much faster RAM does not help Xeons, as shown below. Frankly, all three things together are the issue: performance, price and efficiency, especially P SKUs.
These are main reasons why AMD has risen from 0% server market share in 2017 to 40% very soon. Data center folks have been massively abandoning Xeon systems in recent years. As a result, AMD has experienced an unprecedented raise in server CPUs. Intel will need to pull a rabbit out of hat with Diamond Rapids to stop bleeding in server market. 192P cores sounds great to me.

AMD EPYC "Turin" with 192 Cores and 384 Threads Delivers Almost 40% Higher Performance Than Intel Xeon 6

AMD has unveiled its latest EPYC processors, codenamed "Turin," featuring Zen 5 and Zen 5C dense cores. Phoronix's thorough testing reveals remarkable advancements in performance, efficiency, and value. The new lineup includes the EPYC 9575F (64-core), EPYC 9755 (128-core), and EPYC 9965...

www.techpowerup.com

I think both companies will have 16 channel memory and 192 of the full cores. I think the 256 core Zen 6 are the dense cores.

We don't know about Zen6 192 big core SKU. It has not been leaked. This would be a monster 16 CCDs with 12 core each, if AMD decided to release it.

 

[john_]

It’s good to see that Intel will keep up with AMD.

AMD Namedrops EPYC "Venice" Zen 6 and EPYC "Verano" Zen 7 Server Processors | TechPowerUp}

Both companies will have close to the same specs next gen.

AMD needs to start eating from Nvidia's AI pie and Intel needs to start becoming competitive again in CPUs and start oversupplying GPUs for the mid-low end market at good prices. If those happen, we will have three very strong companies fighting each other and a more balanced market in CPUs and GPUs. On the contrary today we have a monopoly in GPUs and a duopoly in CPUs that is preserved thanks to Intel's ties with OEMs.

 

[TheinsanegamerN]

AMD needs to start eating from Nvidia's AI pie and Intel needs to start becoming competitive again in CPUs and start oversupplying GPUs for the mid-low end market at good prices. If those happen, we will have three very strong companies fighting each other and a more balanced market in CPUs and GPUs. On the contrary today we have a monopoly in GPUs and a duopoly in CPUs that is preserved thanks to Intel's ties with OEMs.

AMD is going to have to develop tools to adapt AI to their GPUs if they want to compete with nvidia.

Intel's OEM ties have been degrading for years, another few years of failure and we might see AMD finally break intel's grip on laptops.

 

[Daven]

We don't know about Zen6 192 big core SKU. It has not been leaked. This would be a monster 16 CCDs with 12 core each, if AMD decided to release it.

Here is the slide deck with Zen 6 and 7 info:

https://www.amd.com/content/dam/amd/en/documents/corporate/events/advancing-ai-2025-distribution-deck.pdf

We are definitely getting 256 cores in Zen 6. I think there are notes about full verse dense versions. I’ll look again and update.

Update: to stay on topic, Intel is saying up to 192 P cores but I’m getting conflicting info if the up to 256 Zen 6 cores are full or dense.

 

[truthsayer]

Intel problem in server space isn't performance, it's price you have to pay for that performance (power bill and cooling included).
BUT maybe they surprise me this time and deliver a reasonably priced alternative to EPYC... maybe...
(reality check : this big thing can be many things, but I highly doubt it's going to be cheap - LGA 9324 )

Every engineer in semiconductors would tell you that "POWER IS PERFORMANCE".

With the SAME power budget, A 48C 4Ghz CPU becomes a 48c 5Ghz CPU if development found some power budget. Or if the node is good during planning, a planned 48C CPU becomes a planned 64C CPU.

With INCREASING power budget, the effects are EXPONENTIAL (look at the power curve for TSMC 7nm). From the earlier example, to go from 4Ghz to 5Ghz, a CPU may needs +15 watts, a less efficient CPU can easily need +30 watts, +45 watts -- multiple times the increase in power. When this causes power itself and cost (closely related) to become unfeasible, plans are scrapped, roadmaps changed, products cancelled.


POWER IS PERFORMANCE: Does Intel want more cores/memory channels/accelerators on their Xeons? Yes, but they can't because of power. Does Nvidia want to glue 3 GPUs together instead of 2 for Blackwell? (assuming they can work out the packaging/memory system) Yes, but they can't because of power.
Products with good efficiency but bad performance are rare, but almost all products with bad efficiency are completely screwed in terms of performance.

 

[agent_x007]

"Power is Performance" is SOO Pentium 4/Netburst era thinking, that the idea of it still being alive just sounds hilarious to me (we came full circle ?).

You never should compensate lack of performance vs. previous gen with increasing power on new gen (at least from my point of view).
It's a very fast way to single yourself out into a corner, from which there are no real good options to move from (and eventual
"moving on" will just hit harder than necessary).

People should stop expecting multiple tens of % performance increases until new fab technology takes over (and I mean planar => FinFIT style, not just "refining" what we have previously).
Each CPU generation should be used longer, with new architecture development taking longer.

I hope we get there eventually, but I would prefer before making each gaming PC and server into pretty good winter heaters.

PS. Performance by itself means nothing. It simply needs to be "good enough" to be marketable.
No need to compete for "first place" when they clearly don't have the right product for it (I guess Intel forgot how NOT to do it ?).
Because Intel could still crush AMD IF they sold Xeons with 1/2 of the equivalent Epyc performance, for 1/10th of the price.

 

[lexluthermiester]

Things are heating up between Intel and AMD, pun intended!

These new Xeon series CPUs are going to be monsters.

 

[wickerman]

People should stop expecting multiple tens of % performance increases until new fab technology takes over (and I mean planar => FinFIT style, not just "refining" what we have previously).
Each CPU generation should be used longer, with new architecture development taking longer.

It's worth considering that is what 18A is. Intel has used finfet and some refinement of it (as you rightly point out) since 22nm and then came fancy tri-gate and superfin. But with 18A (20A was supposed to as well) we see intels version of the Gate All Around transistor with horizontally stacked nanosheets (ribbons) instead of fins. This means every one of those billions of transistors gets to be even smaller, faster, and more power efficient. Its also a showcase of intel's backside power delivery. Moving your power rails to the backside of the wafer means you are away from sensitive signal routes and this is a pretty big deal too as if Intel can pull this off it will be ahead of TSMC and Samsung. TSMC was initially testing BPD for N2P but A16 is where they announced it will debut.

Intel also relied on DUV lithography for 10nm and 7nm with extensive multipatterning, with 10nm it was 8-10 masks per layer on critical features and with 7nm things expanded with up to 12 masks. EUV lithography reset all that nonsense, taking critical features down to 4-6 layers. The importance of this cannot be understated as every mask is an oppertunity for things to go wrong. Any tiny shift in any of those masks will impact how out of spec every feature becomes. The consumer facing impact there is everything from poor clock scaling and power draw to chips that just die one day because a flow of eletrons just ate some poorly formed but rather important traces. Intel moved to EUV with Intel 4, so we get to see what they learned there transitions to 18A and we see mask count reduce still (4-5).

So I would argue this is the new fab tech moment. We see matured EUV ready for a new transistor design, and we see backside power delivery. As troubled as the path has been, this is the moment Intel can do things in the fabs at scale that its competitors are not ready to do at scale.

Its also interesting to note that 18A was set to be an internal node to push intel silicon but Micrsoft and Amazon are also on board (with some kind of AI accelerators or cloud service stuff?) so theres a good chance that they hit the ground running and have the volume to refine things for foundry customers showing up a few quarters later.

 

[john_]

AMD is going to have to develop tools to adapt AI to their GPUs if they want to compete with nvidia.

Intel's OEM ties have been degrading for years, another few years of failure and we might see AMD finally break intel's grip on laptops.

They are expanding in AI with their new hardware. I am not following closely or have the technical knowledge to understand the peculiarities in that part of the industry, but from what I understand they are far behind Nvidia because their solutions until now where having some kind of limitation at 8 GPUs? Something like that. They also lack an interconnect like Nvidia's NVLink. They seem to be working in fixing those with their Helios rack and new Pensando cards. It wasn't just CUDA their problem, but everything around their Instinct GPUs including hardware. Nvidia I believe hit a ceiling in performance with Blackwell and their next architecture will show us if they will retain, increase or start losing their advantage over AMD and everyone else.

It will be difficult for AMD to replace Intel in the minds of OEMs. OEMs love premium brands and I think they will prefer to give more chances to Qualcomm (or Nvidia) than making AMD their first option. So, while we can see AMD winning market share in laptops, that market share can evaporate in an instance if Qualcomm releases something really interesting and good for Windows on ARM.

 

[Squared]

Since there never was a 20A product, we can never know if 18A is just a rebrand. The best we can ever know is how well 18A compares to N2 when if ever Intel releases a product using it.

I suppose we can't know for certain, but 20A was always seemed to me to be a pathfinder node for 18A, like 4 was for 3. I assumed 4 and 20A didn't have the density and yields for server CPUs. So 48-core Diamond Rapids tiles as the first 18A chips suggests it's a much more mature node than the original 20A.

Once upon a time, the sequence was going to be a Arrow Lake on 20A (now on TSMC N3), Clearwater Forest on 18A (delayed because of packaging and market changes), then Diamond Rapids on 18A. 4/3 had that sequence (Meteor Lake on 4, Sierra Forest on 3, Granite Rapids on 3). Each chip in the sequence is bigger and requires a more mature node. So 18A out of the gate has to be good enough for Diamond Rapids.

 

[Vayra86]

Things are heating up between Intel and AMD, pun intended!

These new Xeon series CPUs are going to be monsters.

If they actually happen AND work as advertised. Given Intel, that's quite a stretch in 2025, and as it depends on a new node... well. It could just as well be smoke and mirrors.

 

[Tek-Check]

We are still waiting for Sierra Forest Xeon E CPUs above 144 cores, for the socket 7529. Will they ever release it?

 

[Squared]

We are still waiting for Sierra Forest Xeon E CPUs above 144 cores, for the socket 7529. Will they ever release it?

Canceled, maybe. I saw something about that a while back, although it's still not clear to me, since few sources other than Serve the Home have announced the news and that news is behind a paywall there.

Intel Says It Won’t ‘Broadly’ Deploy 288-Core Xeon 6900E Server CPU

Intel said on Monday that it doesn’t plan to ‘broadly’ deploy its 288-core Xeon 6900E server CPU, which it is instead targeting at cloud customers with custom chip needs.

www.crn.com

Intel's New High-Core Count CPU is Making a Major Change

It makes sense, but it is good news for AMD

axautikgroupllc.substack.com

 

[Tek-Check]

Canceled, maybe. I saw something about that a while back, although it's still not clear to me, since few sources other than Serve the Home have announced the news and that news is behind a paywall there.

So, probably not enough of customers want such product, as there are other options on the market.