Naples CPU Xeon competitor!

[juanrga]

It's the same chip as Summit Ridge.

Naples is four chips on on package. Each one is an 8-core Zen, with 2 memory channels and 32 PCIe lanes. That's how they get 8 memory channels and 128 PCIe lanes, using NUMA.

Desktop Zen has 20 PCIe lanes visible because that's all your average motherboard needs, WITHOUT BEING TOO EXPENSIVE TO ALSO HANDLE 8-LANE Bristol Ridge, and not having to switch to LGA to bring-up pin density.

But you can be sure those extra 12 leads are there on the die, just not connected to anything

Seriously, why would they do a separate core revision just to add more PCIe lanes, when NOT CONNECTING EXCESS I/O IN THE PACKAGE is one of the easiest things you can do with an existing CPU?

Of course desktop Ryzen are just dies don't used for server. A minor note, those 128 PCIe lanes are only available on single-socket platform. Dual socket Naples reduces available PCIe lanes to 64.

 

[Riccochet]

I would love to get my hands on one of these when they release.

Now, VMWare actually incorporating compatibility in a timely fashion is another story. As it sits ESXi 6.5 does not support Ryzen. PSOD on boot.

AMD has the potential here for a real winner in the server market.

 

[SighTurtle]

Project Olympus, Microsoft's attempt to standardize server hardware (i think that's what it means) gains AMD, Quadcomm support and partnership.

 

[Nathan_P]

Of course desktop Ryzen are just dies don't used for server. A minor note, those 128 PCIe lanes are only available on single-socket platform. Dual socket Naples reduces available PCIe lanes to 64.

It reduces the lane count to 64 per cpu as 64 are used to talk to the other cpu, this still gives 128 lanes when running 2 cpu's.

 

[Shintai]

First off, don't ever assume that the people making purchasing decision are well informed. That's not always the case. In my experience, it seldom is the case. In the IT world there is a huge bias against AMD. It stems from the days when AMD made shitty knock offs of Intel's 386 and 486 CPUs at bargain basement prices. Old school IT guys with 20 and 30 years in the business are now the guys in the management positions who make decisions. They may not even be current on technology so they'll hold to any old thinking or habits they have. Intel has a solid reputation for reliability in its CPUs, chipsets, network solutions, SSD's you name it. People will buy other shit that's cheaper. That's how Broadcrap manages to survive in the server world. That said, Intel has a foot hold in the CPU world and part of that isn't as much brand loyalty as recognition. People know what it is.

AMD does have to fight against brand bias. There are organizations like Match.com who had a good sized HP / AMD Opteron based infrastructure but knowing the guys who made those decisions it was usually about cost and bang for the buck when the company was new. This changed later on but at one time that's how those decisions were made in the earlier days. Other companies I've worked at won't ever entertain the idea of buying non-Intel systems anytime soon. One reason for this is some companies with complex infrastructure have a long validation period before standardizing on new hardware. Some companies will continue to buy servers that have been supplanted by newer models simply because it is there standard. I've seen hundreds of thousands of dollars spent on older models for this reason.

I've worked in datacenters and large infrastructures. They are slow to change and the people in charge of the final purchasing decisions are less technically inclined than you might imagine.

Cloud companies are not buying with bias for example. I can straight out guarantee you that and we as a cloud company test servers for AMD each time they make something new, despite how big of a time waste it may be when the PR slides comes up short. AMD had 1 single good server product in its entire lifespan and that was K8. And it they sold everything they could make. You can see the same with OEMs like Lenovo and Dell with a lineup that is all Skylake-EP/Skyake-W in servers and (desktop)workstations.

 

[Shintai]

You know people are saying Naples is based on zen. At the core yes, but to me it seems like they started with a monster server chip and cut it down to make desktop chip. If naples has all those lanes and 8 channel memory, you can bet your end that slowly amd will be upping the desktop side with more lanes and 4 channel memory may be. This thing looks like a beast..

Naples 32 core CPU is 4 Zeppelin dies in MCM. The Zeppelin die is what Ryzen is.

 

[Dan_D]

Cloud companies are not buying with bias for example. I can straight out guarantee you that and we as a cloud company test servers for AMD each time they make something new, despite how big of a time waste it may be when the PR slides comes up short. AMD had 1 single good server product in its entire lifespan and that was K8. And it they sold everything they could make. You can see the same with OEMs like Lenovo and Dell with a lineup that is all Skylake-EP/Skyake-W in servers and (desktop)workstations.

Cloud stuff is a little bit different in their design goals aren't quite the same. Even so, you can't speak for all such companies. As I said, some guys making purchase decisions in IT are set in their ways and will not change. I've worked at a ton of different companies in my 20 years in the business. I've met people of all kinds with varying skill levels, education and biases. Trust me, there are people who aren't smart enough to be doing their job who make purchase decisions for large corporations. Sometimes they have no biases and don't fully understand their companies needs. I had a director of technology at one place that bought whatever he wanted without talking to the people who'd have to deploy it or actually utilize it. That behavior eventually caught up to him but this went on for years. He wasted 10's of thousands of dollars, perhaps more doing this.

I worked for a company that had a very large infrastructure about a year and a half ago. They had a mixture of hardware but their hardware manager would falsify build requests to match hardware he had in stock so he wouldn't have to buy new stuff. I never put anything past anyone.

 

[Shintai]

Project Olympus, Microsoft's attempt to standardize server hardware (i think that's what it means) gains AMD, Quadcomm support and partnership.

Its nothing but a teaser to make other companies commit some money while they keep buying Xeons to put it blunt. Note that MS isn't committing any money.

 

[Shintai]

Cloud stuff is a little bit different in their design goals aren't quite the same. Even so, you can't speak for all such companies. As I said, some guys making purchase decisions in IT are set in their ways and will not change. I've worked at a ton of different companies in my 20 years in the business. I've met people of all kinds with varying skill levels, education and biases. Trust me, there are people who aren't smart enough to be doing their job who make purchase decisions for large corporations. Sometimes they have no biases and don't fully understand their companies needs. I had a director of technology at one place that bought whatever he wanted without talking to the people who'd have to deploy it or actually utilize it. That behavior eventually caught up to him but this went on for years. He wasted 10's of thousands of dollars, perhaps more doing this.

I worked for a company that had a very large infrastructure about a year and a half ago. They had a mixture of hardware but their hardware manager would falsify build requests to match hardware he had in stock so he wouldn't have to buy new stuff. I never put anything past anyone.

But still, cloud is so big that if AMD had anything relevant, it would show as big numbers. And they are not. So you can make the conclusion that some enterprises may be Intel biased as a theory. But you pretty much cant claim that close to everyone is since nobody as such buy AMD for servers. Its simply based on performance metrics. The same applies for ARM.

 

[Gideon]

Shintai arguing their is no bias against AMD, argument ended by his own obvious bias posts, that was a good laugh. Sad part is I still get to see his worthless posts when someone quotes him, at least this quoted post made me laugh tho.

 

[SighTurtle]

https://www.nextplatform.com/2017/03/08/arm-amd-x86-server-chips-get-mainstream-lift-microsoft/

A Naples motherboard AMD made to show off for Olympus.

Summary (You really should take the time to read the entire article through):

These ThunderX2 and Centriq 2400 ARM servers and the AMD Naples server are not deployed in production within Microsoft’s Azure cloud, but are running in the labs at the moment using early samples of the chips from their respective vendors. Microsoft is not making a formal commitment to use any of them in production, but does say that ARM servers (and we think the Naples X86 chips, too) are particularly well suited for “high-throughput computing” such as search engine indexing and serving, storage, databases, data analytics, and machine learning. The demos of the two ARM servers running at OCP Summit were doing a subset of the Bing search engine workload on top of Windows Server 2016, in fact.

Edit: Actually, I realize its not really a summary. oh well.

 

[juanrga]

Wow! Just checked AMD slides/demos about Naples and they are running 1866MHz on the Xeon E5-2699 v4 despite it supports 2400MHz.

 

[Pieter3dnow]

Wow! Just checked AMD slides/demos about Naples and they are running 1866MHz on the Xeon E5-2699 v4 despite it supports 2400MHz.

Women and children first ?

 

[SighTurtle]

Wow! Just checked AMD slides/demos about Naples and they are running 1866MHz on the Xeon E5-2699 v4 despite it supports 2400MHz.

Women and children first ?

Not quite. I think.
Anandtech seems to have the entire press deck of slides from Ryzen Tech Day on Naples, which I assume juanrga is referencing to, the 3 demos done by AMD comparing a Naples system to the E5-2699A v4.

Slide 1: http://images.anandtech.com/galleries/5506/Forrest_Naples_Embargoed Until 3_7_17-page-017_575px.jpg
Slide 2: http://images.anandtech.com/galleries/5506/Forrest_Naples_Embargoed Until 3_7_17-page-018_575px.jpg
Slide 3: http://images.anandtech.com/galleries/5506/Forrest_Naples_Embargoed Until 3_7_17-page-019_575px.jpg

First off, the first slide indicates that AMD demoed the Naples system at the same speed and cores as Intel, so I don't know why hes ignoring that. Second, as the quoted stuff below states, AMD's Naples system was handicapped in the first test to show parity with Intel on a core to core, memory speed equality. The second test is to show Naples's full potential. The third demo/slide, was done to show that Naples has some type of advantage with bigger workloads? i dunno.

I want to state, I know absolutely nothing about what I'm saying, so I may be wrong, but thats why we have journalists, and why I'm just copy and pasting their words. Of course they might be wrong, in which case wtf are they doing being tech journalists.

TheNextPlatform wrote up a lengthy play by play of the demos and systems specs which I'm quoting wholly. I have no clue the relation between the memory channels, bandwidth, and how that affects the memory frequency of Naples vs Intel. Someone else should answer that, if they care enough to.

As you can see in the chart above, the Naples processor has a total of eight memory channels (one per Zen CCX unit, we think), and each channel support two DDR4 memory sticks. So that is up to 16 memory sticks for a single socket machine and up to 32 sticks for a two-socket box. Here is how the Naples system will stack up to a current Broadwell Xeon system using the top-bin 22-core E5-2699A v4 processor:

As you can see, AMD will have the advantage in terms of cores and threads, and with twice as many memory channels it will also have a memory bandwidth advantage. On Naples machines it has tested, AMD can push 170.7 GB/sec of memory bandwidth, compared to Intel’s 76.8 GB/sec on its own tests using a Broadwell system with the top bin parts. Intel Broadwells can have up to three memory sticks per channel, compared to two for Naples, so the Naples chip tops out at 512 GB using very cheap 16 GB sticks compared to 384 GB for the Broadwells. So that is a 122 percent memory bandwidth and a 33 percent memory capacity advantage to AMD. (The Naples chip can also run the memory at 2.4 GHz compared to 1.87 GHz for the Broadwells, and that 29 percent speed bump is why it has higher bandwidth than the memory subsystem on the Xeon chip.) If you need to maximize memory in either a Naples or Broadwell system, the former tops out at 2 TB while the latter maxes at 1.5 TB in two-socket machines.

In terms of performance, AMD gave us a taste of how the Naples server will compare to a current Broadwell server using a seismic analysis workload from the oil and gas industry that does a bunch of Laplacian transforms of 3D wave equations to try to figure out where oil might be lurking in the ground using echoes bouncing off the rocks. This particular workload, says Bounds, stresses the cores, the memory, and the I/O subsystem alike and is representative of many workloads that are running in enterprises. “We wanted to create a scenario where cores matter, but what really matters is the memory, both the bandwidth and the capacity,” explains Bounds.

In the first test that AMD did, it had a 1 billion cell sample grid of data and ran ten iterations of the simulation. It only activated 44 cores on the Naples machine and geared back the main memory to run at 1.87 GHz, just like the two-socket Broadwell Xeon system. It took the Intel box 35 seconds to run the simulation at a rate of 286 computations per second, but the AMD machine could do the simulation in 18 seconds, or just about twice as fast, at a rate of 527 computations per second. In a second run, using the same sample size of 1 billion data points and ten iterations, but with all 64 cores turned on and with the memory boosted up to the full 2.4 GHz speed, the Naples machine was able to do 713 transactions per second on the seismic simulation, and that mean it finished in 14 seconds and was 2.5X faster than the Broadwell machine. Just for fun, AMD did a third test, quadrupling the dataset size to a 4 billion sample grid, and it could run that in 54 seconds on the Naples box, but it would not even load the data on the Broadwell box.

PCPer has a paragraph on the same demos:

AMD claims that its Naples platform offers up to 45% more cores, 122% more memory bandwidth, and 60% more I/O than its competition. For its internal comparison, AMD chose the Intel Xeon E5-2699A V4 which is the processor with highest core count that is intended for dual socket systems (there are E7s with more cores but those are in 4P systems). The Intel Xeon E5-2699A V4 system is a 14nm 22 core (44 thread) processor clocked at 2.4 GHz base to 3.6 GHz turbo with 55MB cache. It supports four channels of DDR4-2400 for a maximum bandwidth of 76.8 GB/s (19.2 GB/s per channel) as well as 40 PCI-E 3.0 lanes. A dual socket system with two of those Xeons features 44 cores, 88 threads, and a theoretical maximum of 1.54 TB of ECC RAM.

AMD's reference platform with two 32 core Naples SoCs and 512 GB DDR4 2400 MHz was purportedly 2.5x faster at the seismic analysis workload than the dual Xeon E5-2699A V4 OEM system with 1866 MHz DDR4. Curiously, when AMD compared a Naples reference platform with 44 cores enabled and running 1866 MHz memory to a similarly configured Intel system the Naples platform was twice as fast. It seems that the increased number of memory channels and memory bandwidth are really helping the Naples platform pull ahead in this workload.

 

[Zedicus]

as the slides state, in dual socket format there will still be 128 lanes available to devices, but the other 128 (64 per cpu) will go to the interconnect. this is actually a good thing.

AMD NEEDS to get this vetted on VM platforms, they NEED to get it working on correctly on XEN and KVM, and they need to get VMWare on board. VMWare is crap, but is the number 1 hypervisor in small to medium enterprise.

also to note, ever seen AMD dove into the server market, they have ALWAYS built the server CPU and then modified it for desktop use. heck the 939 had ECC support, is was up to the board manufacturer to allow it in BIOS. That said, they have finally figured out that you have got to have a good desktop platform if you want to make server sales. it is a geeks industry and if you bomb the desktop, geeks see that and go to work and talk crap and soon bosses are buying the more expensive option. even though bulldozer was actually a good competitor in the server market.

 

[Factum]

First off, don't ever assume that the people making purchasing decision are well informed. That's not always the case. In my experience, it seldom is the case. In the IT world there is a huge bias against AMD. It stems from the days when AMD made shitty knock offs of Intel's 386 and 486 CPUs at bargain basement prices. Old school IT guys with 20 and 30 years in the business are now the guys in the management positions who make decisions. They may not even be current on technology so they'll hold to any old thinking or habits they have. Intel has a solid reputation for reliability in its CPUs, chipsets, network solutions, SSD's you name it. People will buy other shit that's cheaper. That's how Broadcrap manages to survive in the server world. That said, Intel has a foot hold in the CPU world and part of that isn't as much brand loyalty as recognition. People know what it is.

AMD does have to fight against brand bias. There are organizations like Match.com who had a good sized HP / AMD Opteron based infrastructure but knowing the guys who made those decisions it was usually about cost and bang for the buck when the company was new. This changed later on but at one time that's how those decisions were made in the earlier days. Other companies I've worked at won't ever entertain the idea of buying non-Intel systems anytime soon. One reason for this is some companies with complex infrastructure have a long validation period before standardizing on new hardware. Some companies will continue to buy servers that have been supplanted by newer models simply because it is there standard. I've seen hundreds of thousands of dollars spent on older models for this reason.

I've worked in datacenters and large infrastructures. They are slow to change and the people in charge of the final purchasing decisions are less technically inclined than you might imagine.

I have worked in IT in +20 years, I am very up-to-date about current tech...so far I have seen nothing from our vendors that in any way, shape or form seems like a threat to Intel's servermarket.

So far, no AMD products are slated for our datacenters.

PR, forums and reality are seldom eye-toyey

 

[juanrga]

Not quite. I think.
Anandtech seems to have the entire press deck of slides from Ryzen Tech Day on Naples, which I assume juanrga is referencing to, the 3 demos done by AMD comparing a Naples system to the E5-2699A v4.

Slide 1: http://images.anandtech.com/galleries/5506/Forrest_Naples_Embargoed Until 3_7_17-page-017_575px.jpg
Slide 2: http://images.anandtech.com/galleries/5506/Forrest_Naples_Embargoed Until 3_7_17-page-018_575px.jpg
Slide 3: http://images.anandtech.com/galleries/5506/Forrest_Naples_Embargoed Until 3_7_17-page-019_575px.jpg

First off, the first slide indicates that AMD demoed the Naples system at the same speed and cores as Intel, so I don't know why hes ignoring that. Second, as the quoted stuff below states, AMD's Naples system was handicapped in the first test to show parity with Intel on a core to core, memory speed equality. The second test is to show Naples's full potential. The third demo/slide, was done to show that Naples has some type of advantage with bigger workloads? i dunno.

I want to state, I know absolutely nothing about what I'm saying, so I may be wrong, but thats why we have journalists, and why I'm just copy and pasting their words. Of course they might be wrong, in which case wtf are they doing being tech journalists.

TheNextPlatform wrote up a lengthy play by play of the demos and systems specs which I'm quoting wholly. I have no clue the relation between the memory channels, bandwidth, and how that affects the memory frequency of Naples vs Intel. Someone else should answer that, if they care enough to.

PCPer has a paragraph on the same demos:

The first test compares the Naples system with sixteen channels whereas the Xeon has eight both at 1866MHz. AMD carefully chose a memory bound workload to get advantage of that; twice the number of channels explains twice the performance in that workload.

Running 2400MHz on both would probably improve performance much more on the Xeon than in 44C Naples and then kill the "~2x faster" slogan, And surely increasing core count to full 64C Naples from a 2400MHz baseline would reduce the scaling.

I agree that the second test is showing Naples in its "full potential", but the Xeon system is not running at "full potential".

I find those slides a bit ridiculous, luckily sites are taking them with "a grain of salt". From Tomshardware:

AMD was light on the details of its custom seismic workload, and although we do know that it employs AVX instructions, it is impossible to compare the results to standardized workloads or provide a detailed analysis of the tests. AMD indicated that the workload is a computationally intensive analysis involving iterations of 3D wave equations that stresses the CPU, memory, and I/O subsystem. We also weren't provided with more detailed system specifications or settings, so take the results with a grain of salt.

The first workload consisted of 10 iterations of a 1 billion sample grid. AMD restricted its core count and memory speed to match the Intel system, yet still managed to complete the workload in roughly half the time.

For the second test, AMD conducted the same test but brought all 64 cores to bear and bumped its memory speed up to 2,400MHz while the Intel system remained at 1,866MHz. Once again, AMD's carefully selected workload completed faster on the Naples system, yielding a 2.5X advantage. It's impossible to derive any useful scalability comparisons between the workload completion time of the 44-core Naples configuration and the 64-core native configuration due to a lack of information on the workload.

Finally, AMD provided a demo specifically designed to highlight its memory capacity advantage. The company increased the dataset to 10 iterations of a 4 billion sample grid, which simply couldn't run on the Intel system due to its memory capacity disadvantage.

 

[Pieter3dnow]

The first test compares the Naples system with sixteen channels whereas the Xeon has eight both at 1866MHz. AMD carefully chose a memory bound workload to get advantage of that; twice the number of channels explains twice the performance in that workload.

Running 2400MHz on both would probably improve performance much more on the Xeon than in 44C Naples and then kill the "~2x faster" slogan, And surely increasing core count to full 64C Naples from a 2400MHz baseline would reduce the scaling.

What is this nonsense carefully choose memory bound workloads , if that is suited for testing it is nothing but blatantly obvious they have a far better solution then the competition.
Your 2400mhz on both statement holds as much water as "when pigs can fly" and they can't.

 

[gigaxtreme1]

juan, sell that intel stock already.

 

[{NG}Fidel]

No matter how good Naples is, there is a lot of brand recognition for Intel and the Xeon to overcome. Additionally, AMD has to overcome its own negative stereo type in the industry as "that other CPU maker" or a "knock off" manufacturer of CPUs. In the industry, Intel has a reputation that isn't as much about performance as it is reliability. If you want performance from Intel you simply buy a bigger server. If you want something reliable, virtually any server with "Intel" inside will do. Over the last 20 years these thought processes haven't always been rooted in reality but this is how the company is still perceived. I think word of mouth and the regurgitation of information without fact checking is why these stereo types about AMD hold true even when it has a good product on their hands.

It's the same type of brand bias that GM and Ford have to overcome in a lot of markets with their products. In fairness they did it to themselves but that's another topic.

Eh reliability isnt just a CPU or just a platform. Its alot more than that. CPU is no better than the motherboard its inserted on. And even at Major DCs I still see AMD deployed.

 

[Dan_D]

Eh reliability isnt just a CPU or just a platform. Its alot more than that. CPU is no better than the motherboard its inserted on. And even at Major DCs I still see AMD deployed.

You bring up a good point. The platform is one area where AMD has often and usually lagged behind Intel. Either in regard to features or driver quality. It's one of the things that's often worked against AMD and one reason why some IT folks swear by Intel based servers. It isn't as though I never see AMD servers in big data centers. It's just rare compared to Intel based servers.

 

[{NG}Fidel]

You bring up a good point. The platform is one area where AMD has often and usually lagged behind Intel. Either in regard to features or driver quality. It's one of the things that's often worked against AMD and one reason why some IT folks swear by Intel based servers. It isn't as though I never see AMD servers in big data centers. It's just rare compared to Intel based servers.

Problem is that its not just drivers but the damn hardware itself. I have worked on plent of Foxconn, Quanta, and Supermicro intel boards that make you wanna throw them against the wall for how poorly built they are.

Alot of people in this business want cheap boards, cheap cpus, and as many cores as they can get. AMD has a huge in with this market.

 

[Dan_D]

Problem is that its not just drivers but the damn hardware itself. I have worked on plent of Foxconn, Quanta, and Supermicro intel boards that make you wanna throw them against the wall for how poorly built they are.

Alot of people in this business want cheap boards, cheap cpus, and as many cores as they can get. AMD has a huge in with this market.

You'll get no argument from me on these points. There are plenty of Intel based systems that are fucking junk. There is no doubt about that. I'm talking about perception, not necessarily facts. There is an old saying that "Perception is reality." For some of the people cutting checks or making hardware purchase decisions that perception is all they need to make their choices. That said, I've seen plenty of deployments of servers where the objective was to get the best specs on paper for the least amount of money. You are right in that AMD has a nice shot to make sales to people with this mentality.

I have no doubt that Naples will be successful for AMD. How successful remains to be seen.

 

[juanrga]

What is this nonsense carefully choose memory bound workloads , if that is suited for testing it is nothing but blatantly obvious they have a far better solution then the competition.
Your 2400mhz on both statement holds as much water as "when pigs can fly" and they can't.

It is a better solution when the workloads are limited to a custom CPU-based seismic analysis, the competition is older BDW Xeon with memory crippled at 1866MHz, and we don't know anything about power consumption or pricing.

The press is correct on taking those demos with a "grain of salt".

 

[Shintai]

https://www.techpowerup.com/231585/amd-ryzen-infinity-fabric-ticks-at-memory-speed

So in short, Zeppelin dies are really a SMP setup in a package. At 2133Mhz its 34GB/sec. Now imagine 4 dies in a socket and 2 sockets. That's 8 dies and 16 CCXes that's linked with 2 GMI links per die with a link capacity around 34GB/sec in reality depending on 1866, 2133 or 2400Mhz memory.