Intel's make-or-break 18A process node debuts for data center with 288-core Xeon 6+ CPU — multi-chip monster sports 12 channels of DDR5-8000, Foveros Direct 3D packaging tech
Intel unveils x86 CPU with the industry's highest core count.
Intel this week formally introduced its Xeon 6+ processors codenamed 'Clearwater Forest' that pack up to 288 energy-efficient Darkmont cores and are the first data center CPUs made on the company's 18A fabrication process (1.8nm-class). Intel aims its Xeon 6+ 'Clearwater Forest' processors primarily for telecom, cloud, and edge AI workloads as they feature Advanced Matrix Extensions (AMX), QuickAssist Technology (QAT), and Intel vRAN Boost technologies.
Intel's Xeon 6+ processors with up to 288 cores combine 12 compute chiplets containing 24 energy-efficient Darkmont cores per tile that are produced using 18A manufacturing technology, two I/O tiles made on Intel 7 production node, as well as three active base tiles made on Intel 3 fabrication process. The compute tiles are stacked on top of the base dies using Intel's Foveros Direct 3D technology, whereas lateral connections are enabled by Intel's EMIB bridges.
View Original
View Original
Intel's 'Darkmont' efficiency cores have received rather meaningful microarchitectural upgrades. Each core integrates a 64 KB L1 instruction cache, a broader fetch and decode pipeline, and a deeper out-of-order engine capable of tracking more in-flight operations. The number of execution ports has also been increased in a bid to improve both scalar and vector throughput under heavily threaded server workloads.
From a cache hierarchy standpoint, the design groups cores into four-core blocks that share approximately 4 MB of L2 cache per block. At the top of the hierarchy, there is last-level cache across the full package that surpasses 1 GB, roughly 1,152 MB in total. This unusually large pool is intended to keep data close to hundreds of active cores and reduce dependence on external memory bandwidth, which in turn is meant to both increase performance and lower power consumption.
Platform-wise, the processor remains drop-in compatible with the current Xeon server socket, so the CPU has 12 memory channels that support DDR5-8000, 96 PCIe 5.0 lanes with 64 lanes supporting CXL 2.0.
Intel positions Clearwater Forest for telecom and cloud workloads. The company says operators deploying 5G Advanced and future 6G networks increasingly rely on server CPUs for virtualized RAN and edge AI inference, as they do not want to re-architect their data centers in a bid to accommodate AI accelerators. By combining matrix/vector acceleration, vRAN offloads (using the vRAN Boost), large caches, and broad I/O in one platform, the CPU can perform jobs that are normally reserved for various accelerators that consume more power and take up space.
Also, extreme core count of Xeon 6+ 'Clearwater Forest' CPUs — that approaches 288 cores for uniprocessor configurations and 576 cores in dual socket configurations, enabling a single server to host dozens or even hundreds of virtual machines while maintaining power efficiency and low latency.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Systems based on Intel's Xeon 6+ processors will be available later this year.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
-
bit_user Last things first:Reply
The article said:Systems based on Intel's Xeon 6+ processors will be available later this year.
Okay, so not a launch, but rather just a formal announcement.
"as a result" ?? No, 4 MB per 4 cores = 1 MB per core. So, the total amount of L2 cache is just 288 MB.The article said:From a cache hierarchy standpoint, the design groups cores into four-core blocks that share approximately 4 MB of L2 cache per block. As a result, the aggregate last-level cache across the full package surpasses 1 GB
I have seen where it has 1152 of last level cache, but the way it's written makes it sound like you're just talking about L2.
Wait, this has AMX?? Or, is it just the Diamond Rapids members of the family that would have that?The article said:Intel aims its Xeon 6+ 'Clearwater Forest' processors primarily for telecom, cloud, and edge AI workloads as they feature Advanced Matrix Extensions (AMX)
I'd be rather surprised if Clearwater Forest has AMX, since that adds a non-trivial amount of die area per core.
Also, what about AVX-512? -
bit_user So, reviewing what they've previously disclosed about, I see that mentions of AMX, AVX-512, and AVX10 are conspicuously absent.Reply
https://www.tomshardware.com/desktops/servers/intel-reveals-288-core-xeon
Another downside is that it's still PCIe 5.0-based (and CXL 2.0). By contrast, AMD's Venice, probably launching around the same time, is moving up to PCIe 6.0 / CXL 3.0.
Here are some other juicy tidbits, from the above article (17% IPC boost on SPECint!): -
bit_user Reply
The article pretty clearly indicated they had it, which is why I was searching for confirmation.abufrejoval said:These aren't meant to replace P-core Xeons, but front-end ARMs.
Yes, there very much is! Vector/FP is the area where Skymont improved the most over Crestmont/Gracemont, and even those had AVX2.abufrejoval said:not even sure there is much of any floating-point in front-ends. -
bit_user Reply
I think that's why mainstream servers have only 2 sockets.abufrejoval said:I am a bit sceptic when it comes to this giagantism. On one hand, even chips much bigger are likely to never catch everyones scale-up requirements, so scale-out needs to be part of the design anyway.
The main benefit of more cores/server is that you get higher density and it requires less infrastructure per core. That seems to be the main argument behind scaling up core counts per (server) CPU. -
thestryker Darkmont, based on slides, appears to be identical to Skymont for the most part. Seems like maybe some additional math instruction.Reply
Videocardz has the slide deck: https://videocardz.com/newz/intel-announces-xeon-6-clearwater-forest-at-mwc-2026-core-200k-likely-the-next-plus -
bit_user Reply
Oh, good point. I see now that Skymont also has a 416-entry OoO window, so they must've been comparing Darkmont to Gracemont. Chips & Cheese's Skymont details also align with the Darkmont slide's claims about increases in load & store AGUs.thestryker said:Darkmont, based on slides, appears to be identical to Skymont for the most part. Seems like maybe some additional math instruction.
https://chipsandcheese.com/p/skymont-intels-e-cores-reach-for-the-sky -
thestryker Reply
These are as close to perfect as it gets for communications edge which is probably why the MWC announcement. Between the density and accelerators I'd imagine these will be hard to beat outside of bespoke solutions.bit_user said:The main benefit of more cores/server is that you get higher density and it requires less infrastructure per core. That seems to be the main argument behind scaling up core counts per (server) CPU. -
bit_user Reply
I work with a guy who used to be on a team at Intel that developed code for I think either DSA or QAT. He claims most of his former co-workers have left and the software support for those engines is virtually abandon-ware. Maybe it was QAT, because Phoronix has been reporting on QAT getting de-supported:thestryker said:Between the density and accelerators I'd imagine these will be hard to beat outside of bespoke solutions.
https://www.phoronix.com/news/Intel-QAT-FSCRYPT-Linux-6.17 https://www.phoronix.com/news/Intel-QAT-Crypto-Linux-6.17
Of course, it could have improved in Clearwater Forest, but they probably have an uphill climb on the software support aspect. Without that, they're just wasted silicon. -
bit_user Reply
You're contradicting yourself, a little bit. On the one hand, you cast doubt on why they're putting so many cores in these CPUs. Then, you turn around and say these workloads need to run on dedicated silicon.abufrejoval said:IMHO its part of the relentless push toward the hyperscaler fabric for much of what used to be CPU based computing: at their scale pretty near every workload becomes something that simply runs better on a bespoke ASIC, that will only call-out to a general purpose CPU in those 10% cases not covered by ASIC IP blocks.
I think the main case for the accelerators is really just about energy & area-efficiency. However, if/when you have enough general-purpose cores and they get small enough + efficient enough, it really undermines the value-proposition of hardwired accelerators.
I think that's why AMD has done so well without the hard-wired blocks that Intel has been adding. AMD's server CPUs simply had enough additional cores & their efficiency was good enough that you could just run those tasks in software and it was fine.
AMX is a big exception to that, but it adds a lot of area per-core and only helps with a small minority of use cases. Arguably, it's because Intel went so far into making their P-cores enormous that they needed to add dedicated accelerators to regain some competitiveness on area-efficiency (and thereby cost-efficiency). And, ultimately, making E-core Xeons. -
smmasongt Reply
I think of things like GitHub workflows. Every pull request fires up a small VM with 1-2 cores. The workflow is dependent on some tool that isn’t compiled for Arm yet so you may still be sticking to x86. This application alone drives a lot of basic core demand for Microsoft and I’m sure there are many like it. This gives you good consolidation so less racks, less interconnect, less of everything. Many things outside of AI and web hosting.abufrejoval said:That contradiction isn't mine, I'm mostly reading what's being deployed by the hyperscalers. And they aren't necessarily talking about the math which makes it work for them. I mostly commented on how difficult it's to estimate where these fault lines with chiplet designs might come to lie. Is it going to be thousands of cores in a socket or will we see some other type of aggregation, perhaps using optical interconnects?
And will that two-class split continue, merge via bigger ASIC IP blocks taking up fixed functions or split into more tiers?
Many have put their ARM based chips in those front ends, where they get something like twice as many cores out of the same transistor and thermal budget, while the back-end designs stuck to "P"-class Intel and AMD CPUs combined with accelerators with the growing importance of IA. In the front-end they are also likely to be more ISA agnostic, with abstracted standard services which are just used via APIs.
AMD has tried to counter with both types of EPYCs, using either compact or big cores to put a foot in the door of a potentail ARM takeover: after all, there is no reason ARM cores couldn't also cover the big core territy, once the scale of the eco-system was big enough, see Fuitsu designs in HPC.
Likewise I don't see these Xeon 6+ chips replacing those big cores, but essentially Intels answer to both custom ARM cores by hyperscalers as well as AMD's compact core EPYCs: I expect big core variants to follow, on a novel 18A node small cores just allow better defect management via full de-activation of redundant cores.