Intel Cancels its Mainstream Next-Gen Xeon Server Processors

[erek]

“With Oak Stream platforms and Diamond Rapids CPUs, Intel has a better setup than in the current Birtch Stream and Granite Rapids era. The reason is that the Oak Stream roadmap had both a 16-channel and an 8-channel variant. Sixteen channels give the memory capacity of an eight-channel 2DPC platform.

A few months ago, Intel’s CEO Lip-Bu Tan made remarks discussing some of the decisions he was on the path to change, such as removing SMT or Hyper-Threading, and how it made Intel’s upcoming products (like Diamond Rapids) less competitive. Within a few minutes of meeting Intel’s new Data Center Group EVP and GM, Kevork Kechichian, I mentioned that it must be tough for the business to have comments like that out there.

From what I understand, Intel has been analyzing its roadmap over the past few weeks. About a month ago, there was chatter among OEMs that Intel’s next-generation 8-channel platform for Diamond Rapids might end up on the chopping block. More recently, OEMs have been learning that the 8-channel Diamond Rapids platform has fallen off Intel’s roadmap, while the 16-channel Diamond Rapids remained.

We had multiple sources, enough to publish on, so I sent a note to Intel last night, and they quickly sent the following statement:

We have removed Diamond Rapids 8CH from our roadmap. We’re simplifying the Diamond Rapids platform with a focus on 16 Channel processors and extending its benefits down the stack to support a range of unique customers and their use cases. (Source: Intel Spokesperson to STH)

I know there are many rumors in the industry, but this, like removing a Cooper Lake variant from its roadmap, selling its server systems business to MiTAC, and so forth, are all ones that Intel itself has confirmed directly to us.

Final Words​

In many ways, this is the way of the server market. Lower-end server platforms that offered the same number of memory channels as previous generations have gradually been phased out over time. The (Xeon) EN is dead – Long Live EN! from over a decade ago was a classic case study in this, except that the Xeon 6700P/ 6500P series has been quite popular. Perhaps the real takeaway is that in servers, eventually the smaller sockets disappear as the world adjusts to higher server capacities.

For those who want a deeper dive into market dynamics, given this change, you can check our Substack for that and why I think it might be a net positive signal.”

Source: https://www.servethehome.com/intel-cancels-its-mainstream-next-gen-xeon-server-processors/

 

[philb2]

What does this really mean, cutting through the corporate cotton-in-mouth speak? Is Intel properly focusing or (unwillingly) yielding part of the market to AMD?

By the way how many pins on a 16channel CPU socket?

 

[Epyon]

Fail Intel Fail hard. I hope they crash and burn as a company

 

[westrock2000]

Does SMT make subletting out VPS slots more complicated or something? I know under full load SMT didn’t add much, but under partial load isn’t SMT beneficial? Why get rid of it? Or is that because of the switch to power and efficiency cores? But I thought they talked about backing out of power and efficiency cores…well getting away from the efficiency core at least.

 

[KazeoHin]

Fail Intel Fail hard. I hope they crash and burn as a company

They already did?

People think empires crash gloriously with a crescendo and fireworks and an explosion and ewoks dancing to music. But shit RARELY works like that. Cray, IBM, Intel, All the big "Only way to go" names in tech just slowly fizzle out and never truly go away. There is no one pinpoint moment where IBM "died": it was a slow, gradual fade into irrelevancy. Their HQ didn't explode, their CEO didn't crash-out... They just... stopped mattering.

 

[toast0]

Does SMT make subletting out VPS slots more complicated or something?

You can't responsibly have one customer on one thread and another on the other thread. But eh, you get 2 vcpus as the minimum size on cpus with smt, and 1 vcpu on cpus without.

You have to test for your use case, sometimes it's helpful, sometimes it's not. Really depends on what you're doing, I've had packet proxy servers were SMT was useless (even worse, the servers had 28 cores, and I could only use 16, the other 12 did nothing... a 16 core server would have been better, but had to reuse hardware; one thread per NIC queue and the NICs had 16 queues; moving packets across to a different core doesn't make sense). And I've had servers where SMT makes a positive impact, where there's more processing going on and maybe some waiting for memory where SMT helps.

 

[Epyon]

They already did?

People think empires crash gloriously with a crescendo and fireworks and an explosion and ewoks dancing to music. But shit RARELY works like that. Cray, IBM, Intel, All the big "Only way to go" names in tech just slowly fizzle out and never truly go away. There is no one pinpoint moment where IBM "died": it was a slow, gradual fade into irrelevancy. Their HQ didn't explode, their CEO didn't crash-out... They just... stopped mattering.

I like the way you put it. I am going to guess they will always be behind tsmc unless something stalls even tsmc and everyone else can catch up. But yeah i think they are already dead i hope the stock price keeps falling though or they just give up the high end.

 

[LukeTbk]

Is Intel properly focusing or (unwillingly) yielding part of the market to AMD?

AMD epyc line up seem to also go for above 8 channels (12-16), 8 for the highest end could have started to be limiting with how high core counts are getting.

Would be interesting to know in the last few months in the Xeon 6 how popular the 8 channel lane version were versus the 12 one.

 

[Aurelius]

Fail Intel Fail hard. I hope they crash and burn as a company

I don't. Intel was abusive as a monopolist and complacent even after it became clear AMD, Apple, and Qualcomm were eroding its base. However, Intel self-destructing would only reduce competition, limit domestic tech investments, and stifle research. I want Intel to be a phoenix — it might not ever reclaim its former dominance, but it should come out of this alive with renewed focus and a sense of urgency.

Think about it: imagine if the non-Apple computer market was down to AMD and Qualcomm. That'd be fine... up until AMD stagnates or misuses its position. And Qualcomm has a long way to go before it has both a wide variety of mainstream chips and the robust Windows-on-ARM compatibility it needs for people to make the switch.

 

[LukeTbk]

It depends what happen in all the current fab investment in the hope they die, which sound dangerous to be a lot of waste.

even after it became clear AMD, Apple, and Qualcomm were eroding its base.

TSMC, Samsung, AWS, Google (in 2026 how many of TPU/Axion will be deployed versus Xeon at google), Tesla (how much of the robotics from cars to humanoid world will be Intel compute vs Ai6/7 and how much of those will be made by intel fabs instead of TSMC/Samsung)

Competition is already so good and large, that I am not so sure how good it would be to crash and burn, crash and a nibble/agile rebounding Intel seem much better, that make the transition of their current infrastructure build-up into something that work more seamless. Maybe they could burn to ashes and all those patten, talents, know-how, buildings be better use by better orgs that acquire them, but that a gamble, there a very long and worldwide built-up organization know-how of a very complex supply chain.

If the idea than a different in a position of lead would have been less complacent or using their market power in better ways... seem unlikely, that always happen.

Being the next Kodak, how much good that do ? And kodak did not had to crash and burn for digital to conquer all the world of capturing light.

One issue about Intel here, one of the nice things is the over 100,000 superbe jobs and maybe something that need to be let go for that agile rebounding.

Intel in 2022 had more employees than TSMC and Nvidia combined now I think.

 

[cjcox]

They already did?

People think empires crash gloriously with a crescendo and fireworks and an explosion and ewoks dancing to music. But shit RARELY works like that. Cray, IBM, Intel, All the big "Only way to go" names in tech just slowly fizzle out and never truly go away. There is no one pinpoint moment where IBM "died": it was a slow, gradual fade into irrelevancy. Their HQ didn't explode, their CEO didn't crash-out... They just... stopped mattering.

Until all the major suppliers that big companies use have affordable choices other than Intel, they live on. AMD still, to this day, has not been able to climb that wall.

Sure, in a few (emphasis) cases, nested on page 10 of subsection Z, there's a link that will take you to a single SKU that uses AMD, but it is currently "sold out", sure... I guess there's that. And maybe it's "getting better", but IMHO, still has a very very very very long way to go. I realize that people measure "success" by the niche gamer desktop market, and really, I'm not even talking desktop. But, even so, what I said is still very true of the business desktop market, where again, the volumes are huge in comparison.

With that said, Intel is making mistakes for sure. I mean, they use to make HEDT chips that were affordable vs the Threadripper. While high end Ryzen is cute and all, it lacks lanes (intentionally hampered).

What do I want? Everything. Used to be able to get that from Intel, but now they are "cutting back" so as to compete with AMD on "lack of features", sigh. And, AMD's perceived successes in the niche enthusiast (gamer) market, just means they are no longer the discount choice. Sigh again. All this does is "make us all WAIT". So... while perhaps not seen today in the consumer enthusiast gamer world, I think soon, there will be a mountain of unsold AMD minis, etc. Combined with lack of availability of things like, RAM, we could be heading for a storm end user wise. And while "different", the end result is very much like the mega scam GPU crypto farm craze. Which results in, "we WAIT". And as "we wait", it means more people "get out completely". Often times resulting in happier people.

There's winning and then there is true winning. AMD IMHO, is close to missing the latter. And Intel, well.... lots of work to do there, but for now, right or wrong, they still "own" (they just own less, but still the volume champ).

 

[DukenukemX]

They already did?

People think empires crash gloriously with a crescendo and fireworks and an explosion and ewoks dancing to music. But shit RARELY works like that. Cray, IBM, Intel, All the big "Only way to go" names in tech just slowly fizzle out and never truly go away. There is no one pinpoint moment where IBM "died": it was a slow, gradual fade into irrelevancy. Their HQ didn't explode, their CEO didn't crash-out... They just... stopped mattering.

Except that Intel has 50% of the CPU market which is more like healthy competition than death, which is what we consumers want. The death of Intel are just shareholders upset that their stonks went down instead of up. The only way to "fix" Intel is to make them damn near a monopoly, which isn't going to happen.

 

[erek]

"Intel Drops 8-Channel "Diamond Rapids" Xeon 7 in Favour of 16-Channel SKUs​

by AleksandarK Today, 00:54 Discuss (7 Comments)
Intel is simplifying its product stack in the upcoming "Diamond Rapids" Xeon 7 family of processors, and the focus is on memory bandwidth. An Intel spokesperson confirmed for ServeTheHome that the company will be exclusively focusing on 16-channel Xeon 7 CPUs, and that the eight-channel designs have been cancelled. "We have removed Diamond Rapids 8CH from our roadmap. We're simplifying the Diamond Rapids platform with a focus on 16-channel processors and extending its benefits down the stack to support a range of unique customers and their use cases." This is a potential massive advantage for Intel customers across the stack as the added memory bandwidth will increase performance across most workloads."

 

[LukeTbk]

Except that Intel has 50% of the CPU market which is more like healthy competition than death, which is what we consumers want. The death of Intel are just shareholders upset that their stonks went down instead of up. The only way to "fix" Intel is to make them damn near a monopoly, which isn't going to happen.

Really depend on how you define cpu market and how much access to that data exist.

Impossible to know maybe, but estimate are about what 3-4 billions of CPU sold in 2024, Intel did not sold more than 200 millions of them, ~5-7%, Intel market share in the CPU world is probably closer than 50% and declining fast.

Giant share of the x86 cpu subset, which is a really important and more profitable if you keep cost down than average, sure (70% of that one maybe ?).

But that cover up a giant fraction (almost all of it) of the market failure of Intel on their foundry-design business since 2005, they are not much in phones, cars, robots, drones,etc... not much in a large part of AI training or inference, more and more of AWS and Google does not run on them.

All this does is "make us all WAIT".

That seem a bit gloomy considering how explosive the CPU market has been versus anything that people can buy in the world the last 7 years....

Spoiler: intel amd evolution

With PCI 5.0/6.0, needing large amount of lane could become extremely niche if the industry start offering product for it, a single PCI 5.0x1 can support 25Gbps networking with a lot of room left, PCi 5.0 x1 is more than fast enough for an NVME drive (that 3.0x4 after all, not many drives are able to do significantly more in a sustained way and few user would have use for it).

Imo pci-express tech advanced way faster than regular user need, 8x for a gpu pci 5/6 is more than enough, 12x to use freely for the rest would be a lot if the environment adjust, 1x25gbps network, 2x for a super fast main NVME, 2x for an other super fast main nvme, single line 3 time for other NVME, and you still have enough for a secondary GPU

And here AMD success is not just on the gaming PC side of things:
https://www.tomshardware.com/pc-com...-and-powers-33-percent-of-all-desktop-systems

30% of x86 cpu units are from them now apparently and more in dollars in the datacenter space, they are the premium option not the cheaper one.

 

[Lakados]

Here's how that reads to me.
1. Intel knows it has too many SKUs, and trying to supply all of them leads to higher costs, because it can't bring anything to a good "economies of scale" production point. This is something AMD has focused heavily on.
2. Intel knows the 8CH variant was going to lose to a combination of Epic and Threadripper parts
3. Intel knows it needs to simplify. AMD recently did the same when it bought ZT systems to bring server design in-house, then quickly spun it off Intel should do the same; they can no longer meet the first-party support demands placed on them.

TL;DR
Intel needs to clean up its offerings; it has way too much overlap, and it spends too many resources catering to every imaginable price point in seemingly $50 increments. Intel needs to reduce waste so it can bring down costs and improve product presentation.

 

[Randall Stephens]

What does this really mean, cutting through the corporate cotton-in-mouth speak? Is Intel properly focusing or (unwillingly) yielding part of the market to AMD?

By the way how many pins on a 16channel CPU socket?

They’re leveraging core competencies to maximize growth across key sectors.

 

[Aurelius]

Here's how that reads to me.
1. Intel knows it has too many SKUs, and trying to supply all of them leads to higher costs, because it can't bring anything to a good "economies of scale" production point. This is something AMD has focused heavily on.
2. Intel knows the 8CH variant was going to lose to a combination of Epic and Threadripper parts
3. Intel knows it needs to simplify. AMD recently did the same when it bought ZT systems to bring server design in-house, then quickly spun it off Intel should do the same; they can no longer meet the first-party support demands placed on them.

TL;DR
Intel needs to clean up its offerings; it has way too much overlap, and it spends too many resources catering to every imaginable price point in seemingly $50 increments. Intel needs to reduce waste so it can bring down costs and improve product presentation.

That's the core of it to me. I'd add hat the insatiable thirst for AI-related computing power makes 'mainstream' Xeon CPUs awkward as they neither deliver raw speed (yes, you still need CPUs to feed those GPUs) nor are economical enough to make sense for small shops over AMD's offerings.

 

[Lakados]

That's the core of it to me. I'd add hat the insatiable thirst for AI-related computing power makes 'mainstream' Xeon CPUs awkward as they neither deliver raw speed (yes, you still need CPUs to feed those GPUs) nor are economical enough to make sense for small shops over AMD's offerings.

The Intel 16 Channel configurations work incredibly well for the AI applications, and the latest Compute in Memory (CIM) models do very well on Intel platforms. The Intel In-Memory Analytics Accelerator (IAA) does a really good job at accelerating computing work done in memory, so good infact that we are now seeing a huge run on all memory types to take advantage of it... None of us wanted to add RAM or storage to our systems anyway.....

But yeah, advancements in CIM, when paired with Intel IAA, make AI training more feasible than ever; it's supported on Xeon Scalable 4th Gen and newer.

 

[uOpt]

What does this really mean, cutting through the corporate cotton-in-mouth speak? Is Intel properly focusing or (unwillingly) yielding part of the market to AMD?

By the way how many pins on a 16channel CPU socket?

Intel decided to make one socket less. Good, they made more than enough.

7529 pins.

 

[DanNeely]

Does SMT make subletting out VPS slots more complicated or something? I know under full load SMT didn’t add much, but under partial load isn’t SMT beneficial? Why get rid of it? Or is that because of the switch to power and efficiency cores? But I thought they talked about backing out of power and efficiency cores…well getting away from the efficiency core at least.

SMT's potential value has been declining for years. Deeper instruction reorder buffers allow single thread work loads to use higher shares of CPU instruction ports inside each core meaning there's less slack capacity that could be used by a second thread. From the other direction, hardening against speculative instruction attacks has increased the overhead that SMT adds to a system. In addition, removing it entirely eliminates a large attack surface from play. Doing so in hardware (vs disabling it at the BIOS/OS level) also frees up a small amount of die area.

Then - especially on the server side - it runs into the problem that workloads that don't care about single threaded performance can get more throughput running on a cluster of 4 small single threaded cores than a single bit one with SMT.

While similar logic applies on the consumer side the arguments for keeping it are a bit stronger. Low end CPUs (especially low power laptop models) still have fairly modest core counts and are more likely to show benefits from SMT (although with performance binning often removing SMT from the lowest tier parts to save a few dollars the buyers at the bottom end clearly don't care about the performance impact). While that's less of an issue for mid/high end consumer PCs it still can be an issue; and unlike in the enterprise market where just adding more servers is always an option we don't have the ability to split our ex gaming workloads across two PCs for more performance if needed.

Despite Intel seeming to hesitate now, I still give it even odds that SMT will be gone from new server designs by 2030; and wouldn't be surprised to see it trickle down to consumer PCs as well. With much more severe hardware constraints I suspect mobile (and mid-high end embedded) will remain the last stronghold for SMT.

 

[Lakados]

SMT's potential value has been declining for years. Deeper instruction reorder buffers allow single thread work loads to use higher shares of CPU instruction ports inside each core meaning there's less slack capacity that could be used by a second thread. From the other direction, hardening against speculative instruction attacks has increased the overhead that SMT adds to a system. In addition, removing it entirely eliminates a large attack surface from play. Doing so in hardware (vs disabling it at the BIOS/OS level) also frees up a small amount of die area.

Then - especially on the server side - it runs into the problem that workloads that don't care about single threaded performance can get more throughput running on a cluster of 4 small single threaded cores than a single bit one with SMT.

While similar logic applies on the consumer side the arguments for keeping it are a bit stronger. Low end CPUs (especially low power laptop models) still have fairly modest core counts and are more likely to show benefits from SMT (although with performance binning often removing SMT from the lowest tier parts to save a few dollars the buyers at the bottom end clearly don't care about the performance impact). While that's less of an issue for mid/high end consumer PCs it still can be an issue; and unlike in the enterprise market where just adding more servers is always an option we don't have the ability to split our ex gaming workloads across two PCs for more performance if needed.

Despite Intel seeming to hesitate now, I still give it even odds that SMT will be gone from new server designs by 2030; and wouldn't be surprised to see it trickle down to consumer PCs as well. With much more severe hardware constraints I suspect mobile (and mid-high end embedded) will remain the last stronghold for SMT.

Not to mention that when dealing with virtual hosts, disabling HT/SMT often yields better results.
Disabling it doesn't change your max vCPU assignments; it remains 4:1 based on physical cores. But it makes scheduling those vCPUs easier for the Host OS, especially when a single VM suddenly spikes in usage.

Yes, there are lots of cases where having it enabled for a single application can yield performance benefits, but VM Hosts assign workloads wherever it can fit them.
When HT/SMT is enabled, it will very often fit them where you don't want them, because the HT/SMT threads behave differently than physical ones, and can do unpredictable things when the physical thread is already under a heavy workload.

Most data centers turn it off because the drawbacks far outweigh the benefits, and ARM adoption feeds off Intel and AMD needing those cores turned off.
Because if your vCPU assignment is limited to 4:1, and your VMs are mostly running services, websites, payment gateways, databases, etc, having a crapload of small cores is better than having a few big ones.