Intel Confirms Skylake Xeon D in Early 2018

[amdfan111]

EPYC has more bandwidth, also in real-world workloads. It also has capability for more memory, which is a concern for some niche workloads.

EPYC has only 42 GB/s bandwidth per domain. Having "more" bandwidth through 8-way NUMA is meaningless if that bandwidth is unavailable when it is needed. The most important server workload, the database, performs exceedingly poorly on EPYC.

In sparse matrix FP workloads (where you can't pack stuff neatly to use AVX512 units) EPYC outperforms Xeons significantly.

The Anandtech site benchmarks are a full of obsolete code and basically a joke. Furthermore, AVX-512 offers great performance gains for sparse matrices, as it features improved load/store instructions like embedded broadcast, compress/expand, wide permute etc. Sparse matrix operations are bound by instruction latency, not memory bandwidth.

But yeah, of course DELL will go through all of this trouble, validating and announcing their EPYC line just to produce servers that are slower in every possible real-world workload

Everybody and their mother can get their chip announced by the big-names and some PR stuff published. What matters is actual shipments and actual datacenter usage. You may notice how all the Cloud announcements of Intel "alternatives" are for non-performance sectors like ultracheap or storage.

Basically nobody is investing seriously in first-generation EPYC processors. All the major players are waiting for Zen2 with 16-core dies and possibly reworked interconnect. Given Icelake delays and the expected lower core count, the first opportunity for AMD to move real volume in the datacenter will be later this year on Zen2.

 

[amdfan111]

In any case, Xeon D Skylake is likely to destroy all of AMD's server hopes, as there isn't actually anywhere that AMD can fit their chips in the ecosystem when Intel is willing to sell so much (superior) silicon for so cheap.

I do not see Skylake-DE making much difference either way in terms of EPYC competitiveness. The new generation of Xeon D is not exactly a revolutionary product the way the original Xeon D was. Given that the new processors are just Skylake-SP with on-package PCH, the only real difference is form factor (physical size). Anybody could have already been deploying Xeon Silver or Gold setups, which are available at low TDPs and have the same connectivity capabilities (Lewisburg).

 

[24601]

I do not see Skylake-DE making much difference either way in terms of EPYC competitiveness. The new generation of Xeon D is not exactly a revolutionary prod.uct the way the original Xeon D was. Given that the new processors are just Skylake-SP with on-package PCH, the only real difference is form factor (physical size). Anybody could have already been deploying Xeon Silver or Gold setups, which are available at low TDPs and have the same connectivity capabilities (Lewisburg).

You are forgetting the fact that that PCH isn't free outside of the Xeon-D Skylake.

It's cheaper on both the CPU side (due to not increasing cost while including PCH) as well as the integration side (as board has more space for anything they want which lowers cost of motherboards in various ways)




And the whole "wait for Ryzen 2 in 2019/2020" is silly.
As long as AMD is using glued together 4 core CPUs, their CPUs are going to be dead on arrival for actual workloads.

 

[LTC8K6]

And the whole "wait for Ryzen 2 in 2019/2020" is silly.
As long as AMD is using glued together 4 core CPUs, their CPUs are going to be dead on arrival for actual workloads.

As an Intel fan and AMD user, I say that's BS.

 

[amdfan111]

You are forgetting the fact that that PCH isn't free outside of the Xeon-D Skylake.

It's cheaper on both the CPU side (due to not increasing cost while including PCH) as well as the integration side (as board has more space for anything they want which lowers cost of motherboards in various ways)

This is true, but we are talking about tens of dollars of savings here for any significant customer. Given the expanded TDP of the D-2100, the main advantage remaining is physical size. This does matter on the "edge," but it is not going to be a dealbreaker if EPYC were already competitive. Even when compared with Xeon Silver 4000 and Xeon Gold 5000, EPYC was already a hard sell due to the high baseline power of the MCM design.

And the whole "wait for Ryzen 2 in 2019/2020" is silly.
As long as AMD is using glued together 4 core CPUs, their CPUs are going to be dead on arrival for actual workloads.

Zen2 could significantly rework the interconnect and reduce NUMA penalties. Even if it were the exact same design, 16-core dies will go a long ways in mitigating scheduling (job packing) issues present on first-generation EPYC. Remember that AMD does not have to win everywhere, since they are the "underdog." They only need to win in a few non-token places (i.e. not "storage") to enter Zen3 on more advantageous footing.

 

[24601]

The problem is that that time-frame also has Intel getting massive core inflation due to the transition from 14nm to 10nm.

We will essentially be where we are today but with 50% to 100% more cores for each respectively.

And that's before we factor in the fact that the Icelake core is expected to be a fairly large increase in performance over the Skylake core we have today.

 

[amdfan111]

That is actually not clear at all. Leaks have shown that Zen2 will reach 16 cores/die and 64 cores/package, but the Icelake leaks only reveal up to 38 cores. It may well be that second-generation EPYCs consume much more power than the current SKUs, but increased density has a value of its own. It could take Intel up to Sapphire Rapids to adjust their manufacturing processes for more cores (e.g. EMIB MCM).

 

[itsmydamnation]

I work for a cloud company and have spent a decade working in large enterprise doing design and deployment, the agenda being driven in here is so far off base its funny. Simple fact is no matter what the workload the vast major are running on hypervisors and the majority are 4 and under vcpu's almost never see any over 8. Then you can look at vcpu oversubscription of anywhere from about 2-5 for a typical org ( depending on function of vm's run on that server).

As a hypervisor EPYC is very attractive because high density over subscribed hypervisors are very rarely cpu limited and are almost always either memory bandwidth, total memory or I/O limited. In all 3 of those situations EPYC has a strong play.

Even big DB servers are running lots of DB's and guess what , their NUMA aware, if you look at total number of DB's any organisation has you can run the vast majority of them on EPYC without any issue. This is also ignoring the general push ( rightly or wrongly) to DB's that aren't relational because hyperscale.

2018 is the year of EPYC being deployed onto the mainstream, ifs it dead on arrival then AMD will see 0% increase in server market share. Do you want to put your money where your mouth is? Talk is cheap, put something of value on the line, im more then willing to bet your wrong!

 

[amdfan111]

It sounds like you are talking about the "ultracheap" segment. The fact of the matter is, as soon as you need a 5 or more core VM, you have serious packing issues on EPYC which reduce the value proposition. This is on top of the fact that Skylake-SP is dirt cheap for "cloud" companies, but I am sure you already knew that.

 

[itsmydamnation]

It sounds like you are talking about the "ultracheap" segment. The fact of the matter is, as soon as you need a 5 or more core VM, you have serious packing issues on EPYC which reduce the value proposition. This is on top of the fact that Skylake-SP is dirt cheap for "cloud" companies, but I am sure you already knew that.

No I am talking big enterprise with 10s of millions of customers and 10's of thousands of employees.

It's also 9 or more vcpu's and an app that's not NUMA aware that needs to be NUMA aware to scale. So now your really scraping the bottom of the barrel in terms of justification.

 

[jpiniero]

Yeah, remember the target use of processors like Xeon D is more like web servers and application servers. So Epyc should be fine there, but I imagine they would rather want Snowy Owl which isn't available yet I don't think it has some of the features which could be desirable (like quad 10gbe).

 

[scannall]

You are forgetting the fact that that PCH isn't free outside of the Xeon-D Skylake.

It's cheaper on both the CPU side (due to not increasing cost while including PCH) as well as the integration side (as board has more space for anything they want which lowers cost of motherboards in various ways)




And the whole "wait for Ryzen 2 in 2019/2020" is silly.
As long as AMD is using glued together 4 core CPUs, their CPUs are going to be dead on arrival for actual workloads.

So, just buy Intel then? Pay more, get less and have a lot of fun dodging security problems? What a great deal!

 

[amdfan111]

It's also 9 or more vcpu's and an app that's not NUMA aware that needs to be NUMA aware to scale. So now your really scraping the bottom of the barrel in terms of justification.

No? If you have 4x 8 core nodes, you can only fit 4x 5 core VMs. Sure, you can also pack 4x 3 core VMs to make the difference, but at the margins your efficiency will go down, because you can not always find the right mix of jobs. Before you try to hand-wave this away by claiming most VMs are oversubscribed, if you were a "cloud" company, you would surely know that large/dedicated VM shapes earn far more than oversubscribed "container" VMs. From your low-effort argumentation, it sounds very much like you do not have "millions of customers."

P.S. Being "NUMA aware" does not magically make NUMA penalties go away.

Yeah, remember the target use of processors like Xeon D is more like web servers and application servers. So Epyc should be fine there, but I imagine they would rather want Snowy Owl which isn't available yet I don't think it has some of the features which could be desirable (like quad 10gbe).

EPYC is not competitive with Xeon D (and Bronze/Silver) mainly because all those dies do not scale well to low power usage. Snowy Owl can make a lot of improvements here, if AMD can get the connectivity performance to-par. Leaked slides claim it has 64 PCIe lanes and 8x 10 Gbit connections, which sounds good on paper. On the other hand, none of the available EPYC solutions feature networking from the CPU package, so there is still doubt.

 

[UsandThem]

Keep the discussion focused on the Intel Skylake Xeon D.
If you want to discuss EPYC or Zen2 in great detail, discuss
it in one of those existing threads.

AT Mod Usandthem

 

[itsmydamnation]

No? If you have 4x 8 core nodes, you can only fit 4x 5 core VMs. Sure, you can also pack 4x 3 core VMs to make the difference, but at the margins your efficiency will go down, because you can not always find the right mix of jobs. Before you try to hand-wave this away by claiming most VMs are oversubscribed, if you were a "cloud" company, you would surely know that large/dedicated VM shapes earn far more than oversubscribed "container" VMs. From your low-effort argumentation, it sounds very much like you do not have "millions of customers."

And if your Memory or I/O limited to begin with what does it matter?
The next point is where have you been the last decade? Its been all about sideways scale out software. So no , in reality you find server farms with real world workloads scale out with small numbers of symmetric cores. an example , one customer i worked with had aprox 104x 4 vcpu servers for one piece of application infrastructure. If they needed more throughput its more vm's not more vcpu its more vm, this is because of the above i/o + memory.

Your inventing situations that dont exist in large scale to justify the position. You make statements of if you are at a cloud company, well you would know regardless of sku you pick if you dont spin up and spin down capacity as needed you generally end up paying a lot more for you capacity vs on prem. So if your doing this effective and can scale out you pick the cost effective configurations, what does that look like.....

Also in enterprise you are almost always running a vcpu over-subscription and last tiem a checked enterprise still buy massive volumes over servers.......

P.S. Being "NUMA aware" does not magically make NUMA penalties go away.

And hand waving look at the NUMA issue doesn't make all applications or workloads will having scaling problems, its all about data access.



edit: sorry i was typing this and didn't see your post.

 

[amdfan111]

Snip.

Let us look at the leading "cloud" provider, Amazon. Who knows, maybe you work for Amazon, since you claim to have so much "cloud" knowledge. Every month, they announce a new, bigger, faster, etc. machine shape. If nobody wants anything but 1-core oversubscribed commodity hosting, why is Amazon buying so much hardware (multiple TB of RAM per machine, exotic NICs, etc.)? The most valuable VM shapes are dedicated, high-capacity instances. Oversubscribed VM shapes are mainly beneficial from a business perspective to smooth-out unused capacity on larger machines.

Your(sic) inventing situations that dont exist in large scale to justify the position. And hand waving look at the NUMA issue doesn't make all applications or workloads will having scaling problems, its all about data access.

If you are really a "cloud" business, you would know that EPYC and Skylake-SP have the same price. This means the value comes to power-efficiency (AMD behind) and utilization effectiveness. NUMA issues are real and affect job packing in obvious ways. Consider your "memory limited" application that requires high bandwidth. On Skylake-SP, that job can be co-located with applications that are not memory-intensive to increase the effective utilization of the server. With NUMA boundaries everywhere, the efficiency of that job mix is reduced. This job mix is also very real, since one customer could be running a NoSQL server, while another is running video transcoding, JAVA application hosting, etc.

The same reasoning applies for memory capacity usage. Any job that requires more memory than can fit on one die incurs a penalty on EPYC. On Skylake-SP, that job can be co-located with lower memory usage jobs, again increasing the machine utilization. Anything that reduces machine utilization reduces cost effectiveness, and uniform architectures have the highest utilization efficiency.

 

[DrMrLordX]

The original Xeon D was an interesting product: it showed us Intel's 14nm process and Broadwell uarch in a low-power enterprise package that provided compelling perf/watt.

Skylake Xeon D is just undervolted, underclocked enterprise Skylake with the chipset layout rejiggered. No 10nm process, no Cannonlake uarch.

Take that along with the inevitable performance hits from Meltdown/Spectre patches that are already wrecking havoc on Intel-based enterprise systems, and you have a recipe for disappointment.

 

[amdfan111]

The original Xeon D was an interesting product: it showed us Intel's 14nm process and Broadwell uarch in a low-power enterprise package that provided compelling perf/watt.

Skylake Xeon D is just undervolted, underclocked enterprise Skylake with the chipset layout rejiggered. No 10nm process, no Cannonlake uarch..

How would Cannonlake-DE be any more interesting than Skylake-DE? Broadwell-DE was interesting because it integrated a higher level of connectivity than was available in Grantley in a compact and low-cost platform. Skylake-DE is disappointing because its connectivity capabilities were already present in Purley/Lewisburg, and because it no longer targets ultra-low power ranges of 25-45 W. It does feel like there was a missed opportunity to bring in a higher level of connectivity, such as 25 or 40 Gbit networking, to match the increased compute throughput. That does not necessarily mean it is poorly positioned; in fact, it likely meets the requirements of its main customer, Facebook, which was demanding increasingly larger Broadwell-DE chips.

 

[Burpo]

Intel Sharpens The Edge With Skylake Xeon D

https://www.nextplatform.com/2018/02/07/intel-sharpens-edge-skylake-xeon-d/

 

[DrMrLordX]

How would Cannonlake-DE be any more interesting than Skylake-DE? Broadwell-DE was interesting because it integrated a higher level of connectivity than was available in Grantley in a compact and low-cost platform. Skylake-DE is disappointing because its connectivity capabilities were already present in Purley/Lewisburg, and because it no longer targets ultra-low power ranges of 25-45 W. It does feel like there was a missed opportunity to bring in a higher level of connectivity, such as 25 or 40 Gbit networking, to match the increased compute throughput. That does not necessarily mean it is poorly positioned; in fact, it likely meets the requirements of its main customer, Facebook, which was demanding increasingly larger Broadwell-DE chips.

Skylake-DE is disappointing simply for the reason you brought up above: it no longer targets the power ranges of 25-45W. For its time, Broadwell-D provided excellent performance in that power envelope (making Intel's Atom-based server solutions look a big ungainly in comparison). Cannonlake-D/DE would have upped the ante on what Intel could deliver in that power envelope. Skylake-DE, not so much.

 

[Dayman1225]

Some Numbers on Skylake Xeon D if anyone is interested:
Platform Power Consumption and First Benchmarks of the Intel Xeon D-2100 Series
Supermicro X11SDV-4C-TLN2F Review with Intel Xeon D-2123IT

 

[IntelUser2000]

Cannonlake-D/DE would have upped the ante on what Intel could deliver in that power envelope. Skylake-DE, not so much.

What about Denverton? They probably scaled Xeon D up because the Atom C3000 was encroaching upon its territory. The high end chip like the C3955 does extremely well: https://www.servethehome.com/intel-...d-embedded-qat-linux-benchmarks-and-review/2/

Cannonlake would have been nice, but its practically dead! Assuming 10nm went as Intel originally planned, we should have seen Icelake by now, and Knights Hill would have been released last year.

 

[Dayman1225]

OCPUS18 –Technical Review of Twin Lakes Intel Architecture Server for Yosemite V2

Uses Skylake Xeon D, thought it may interest some!

 

[Dayman1225]

ServeTheHome explores the evolution of Xeon D from Gen to Gen:

Article Here

 

[jpiniero]

What about Denverton? They probably scaled Xeon D up because the Atom C3000 was encroaching upon its territory.

Nah, because remember the 8/12/16 Broadwell-D products are faster/better than the server Atoms. I imagine it was just the classic needing to release something just to have something new; and lets face it at the middle tier core counts Skylake-SP isn't any better than Broadwell if not worse because of the hit due to the gimped L3. Hence the need to jack up the TDP/power for Skylake-D to give something different enough to get some sales.