Intel Introduces New Mesh Architecture For Xeon And Skylake-X Processors

[InvalidError]

Wow! So many Intel apologists, typical Tom's 'In either case, the mesh architecture has a |tremendous| scalability advantage over its predecessor' can we wait to see real numbers?

You don't need to see numbers to determine that a mesh has better flexibility than a ring bus... a ring bus is one of the worst possible topologies for any latency-sensitive algorithm and that is made further complicated when that latency is widely variable depending on where the two endpoints of a transaction are while a mesh is just about as close to the ideal case of every core being connected to everything as you can practically get.

The main reason meshes weren't used much sooner is the added complexity of managing that mesh on a clock-to-clock basis. AMD might be a little closer to the ideal balance here by clustering cores to reduce the number of stops on the mesh/fabric.

 

[bit_user]

I don't get your point.

Like you, I was just trying to set some facts straight.

I guess if you want to boil my post down to a single point, it would be what I said at the end: none of this stuff is really novel. All of these ideas were pioneered by the minicomputer & supercomputer companies of prior decades. So, I just see arguments over who's copying whom as utterly pointless and misguided.

So, I'll take that to mean you agree with my point. I appreciate the support.

Well, if my post seemed a bit sharp, it's because I thought you were unjustly harsh on AMD for x86-64. Plus, you completely ignored what I consider to be a worse sin: Intel's PAE.

And I didn't fact-check anything but the MIPS point, but I'd suggest double-checking your facts, when giving history lessons.

 

[bit_user]

Wow! So many Intel apologists, typical Tom's 'In either case, the mesh architecture has a |tremendous| scalability advantage over its predecessor' can we wait to see real numbers?

You don't need to see numbers to determine that a mesh has better flexibility than a ring bus...

Yeah, we should be able to compare and critique architectures without taking sides.

Look, AMD is the underdog, so it naturally gets more sympathy. And, being the big dog, Intel has managed to implement a decent number of objectionable business practices, in its long history. But let's not forget that neither of these companies is a charity. In the end, they're seeking to maximize shareholder value. Anything less could earn the senior management teams pink slips. In fact, a couple years ago, Qualcomm was sued by a major shareholder who thought they weren't acting in the best interest of shareholders. And that's not so uncommon.

AMD might be a little closer to the ideal balance here by clustering cores to reduce the number of stops on the mesh/fabric.

AMD's topology seems to be more tree-structured. The worst-case is worse than a mesh, but the typical case might be a bit better. Software that's well-mapped to it should scale more linearly, since the memory controllers are local to each die. A Skylake with the same number of cores could be less efficient, in the same cases, as some memory accesses will always have to traverse parts of the mesh.

 

[ZolaIII]

Looks like ARM's DynamIQ solution is the most advanced & most scalable one currently available + it's licensable IP.

 

[dusty13]

fascinating how much innovation intel can show, when they get their ass handed to them by amd ...

still they will have to do a lot better than what they currently show. all announcements made lately are out at least half a year to a year, while amd is launching their products inside the next few months.

i can only hope it will be enough to win amd some serious market share so intel is forced to cut the crap its played for more than 5 years now and finally get of its greedy and lazy ass to do some real investment into innovation again.

 

[Abhishek_112]

Interesting, hope to see some reviews on Monday. I'm pretty curious how the 7820x will turn out.

This is really good piece of information. I also write articles on technology trends and would seek you guys feedback to improve.
http://www.biztechpost.com

 

[InvalidError]

fascinating how much innovation intel can show, when they get their ass handed to them by amd ...

Intel cannot overhaul a CPU design overnight. SKL-X's architecture was decided well over a year before Ryzen was announced. The only change is that ThreadRipper forced Intel to bring its 12-18 cores die to HEDT instead of remaining a server/workstation exclusive in the E5 series (whatever metal that maps to in the new naming scheme) at twice the price.

 

[Bubblebob]

The die shot seems to be from an HCC die and not an XCC die. The XCC die is supposed to have 28 cores.

 

[JamesSneed]

fascinating how much innovation intel can show, when they get their ass handed to them by amd ...

Intel cannot overhaul a CPU design overnight. SKL-X's architecture was decided well over a year before Ryzen was announced. The only change is that ThreadRipper forced Intel to bring its 12-18 cores die to HEDT instead of remaining a server/workstation exclusive in the E5 series (whatever metal that maps to in the new naming scheme) at twice the price.

Exactly! Lets not forget the most important thing AMD caused Intel to do, lower prices. The 8 core from last gen was $1,000 now its $600. If threadripper carries on with the great multi-threaded performance of Ryzen we should see more price cuts on all the higher core count CPU's from Intel.

 

[PaulAlcorn]

The die shot seems to be from an HCC die and not an XCC die. The XCC die is supposed to have 28 cores.

Good catch, fixed!

 

[kinggremlin]

Exactly! Lets not forget the most important thing AMD caused Intel to do, lower prices. The 8 core from last gen was $1,000 now its $600. If threadripper carries on with the great multi-threaded performance of Ryzen we should see more price cuts on all the higher core count CPU's from Intel.

Based on the history of Intel's pricing for Xtreme/HEDT cpu's, AMD's Ryzen had no effect on Intel's pricing. They've used the same 3 tier sub $400/$600/$1000 pricing for years with higher core counts dropping to lower pricing tiers.

It appears the only response Intel had to Ryzen was specifically to the threadripper release which was to announce 12-18 core CPU's. And let's be realistic to even enthusiast users, those are useless cpu's. Due to power limitations, the 8 and 10 core CPU's have higher base and boost clock rates which will make them faster in games and any applications desktop users will actually be running. Threadripper and Intel's 12+ core CPU's are just marketing dick wagging material for the companies as far as home users are concerned.

 

[InvalidError]

Based on the history of Intel's pricing for Xtreme/HEDT cpu's, AMD's Ryzen had no effect on Intel's pricing. They've used the same 3 tier sub $400/$600/$1000 pricing for years with higher core counts dropping to lower pricing tiers.

More cores for a lower price is an effect. Had Ryzen and ThreadRipper not come along, I doubt Intel would have bothered to offer more than a slight discount over the previous generation for a given number of cores and the new top-end part would have set a new high watermark for MSRP with minor benefits like previous models since AMD disappeared from the map have.

As far as a direct response to Ryzen is concerned, it takes at least two years to design a chip, so don't expect a true response for at least another year. Intel does have its hex-core Coffee Lake in the near-future and if Intel prices it right, it should be more than enough to retain most would-be deserters who haven't gone AMD already.

I've said before in other threads that Intel needs AMD to survive its potential bankruptcy in 2019-2020 to avoid falling under special anti-trust regulations as a monopoly provider of essential infrastructure. Shaking regulators off of its back must be the main reason why Intel appears in no real hurry to assault AMD with new products and price drops. Also, since Intel has claimed that its rising CPU prices are due to increasing costs, it would look bad in an anti-trust investigation if Intel suddenly slashed most of its prices by as much as 50% to catch up with AMD's.

Intel's response to Ryzen/TR/EPYC will span a few years, assuming AMD can keep the pressure up that long.

 

[JamesSneed]

Based on the history of Intel's pricing for Xtreme/HEDT cpu's, AMD's Ryzen had no effect on Intel's pricing. They've used the same 3 tier sub $400/$600/$1000 pricing for years with higher core counts dropping to lower pricing tiers.

More cores for a lower price is an effect. Had Ryzen and ThreadRipper not come along, I doubt Intel would have bothered to offer more than a slight discount over the previous generation for a given number of cores and the new top-end part would have set a new high watermark for MSRP with minor benefits like previous models since AMD disappeared from the map have.

As far as a direct response to Ryzen is concerned, it takes at least two years to design a chip, so don't expect a true response for at least another year. Intel does have its hex-core Coffee Lake in the near-future and if Intel prices it right, it should be more than enough to retain most would-be deserters who haven't gone AMD already.

I've said before in other threads that Intel needs AMD to survive its potential bankruptcy in 2019-2020 to avoid falling under special anti-trust regulations as a monopoly provider of essential infrastructure. Shaking regulators off of its back must be the main reason why Intel appears in no real hurry to assault AMD with new products and price drops. Also, since Intel has claimed that its rising CPU prices are due to increasing costs, it would look bad in an anti-trust investigation if Intel suddenly slashed most of its prices by as much as 50% to catch up with AMD's.

Intel's response to Ryzen/TR/EPYC will span a few years, assuming AMD can keep the pressure up that long.

Exactly. How can we prove it? When Skylake-X hits the shelves in a couple weeks will they have a 12 core product, nope they wont. My point is those three tiers would have been the same as Broadwell-E(6-cores $617, 8-cores $1089, 10-cores $1723) on release date then Ryzen came along. The fact Intel doesn't have 12-18 core models means they are adjusting to Ryzen/Threadripper along with there pricing structure. This is a very good thing for us.

 

[kinggremlin]

Based on the history of Intel's pricing for Xtreme/HEDT cpu's, AMD's Ryzen had no effect on Intel's pricing. They've used the same 3 tier sub $400/$600/$1000 pricing for years with higher core counts dropping to lower pricing tiers.

More cores for a lower price is an effect. Had Ryzen and ThreadRipper not come along, I doubt Intel would have bothered to offer more than a slight discount over the previous generation for a given number of cores and the new top-end part would have set a new high watermark for MSRP with minor benefits like previous models since AMD disappeared from the map have.

Again, you are completely ignoring the previous 5 generations of Intel HEDT releases if you think Skylake X low core lineup was in any way modified in response to Ryzen.

Sandy Bridge E
6 core - $990
6 core - $555
4 core - $294

Ivy Bridge E
6 core - $990
6 core - $555
4 core - $310

2nd generation, no changes to core counts.

Haswell E
8 core - $999
6 core - $583
6 core - $389

Third generation, 6 core drops to bottom tier, higher core count added to top. What did AMD release to force Intel into doing this?

Broadwell E
8 core - $1089
6 core - $617
6 core - $434

2nd generation with 8 core, again no changes to core count just like Ivy Bridge. We don't expect a higher core count for the $1000 bracket, so instead Intel adds a tier above this for the 10 core.

Skylake X is the third generation of 8 core, so we expect core counts to drop a tier with a higher core count at the $1000 tier. So what actually happened?

10 core - $999
8 core - $599
6 core - $389

How is that not exactly what we should have expected for both core counts and price range based on history? Intel doesn't seem to care about Ryzen in the slightest as they did not make any unexpected changes to the low core count release for Sky Lake X.

The only unexpected releases Intel announced were the high core count CPU which are obviously not intended to counter the 1700x and 1800x. Threadripper CPU's appear to be the only AMD CPU's that Intel is paying attention to.

 

[InvalidError]

Again, you are completely ignoring the previous 5 generations of Intel HEDT releases if you think Skylake X low core lineup was in any way modified in response to Ryzen.

Where the heck did you get that idea? I've said multiple times that Skylake-X's architecture was decided long before Ryzen was announced, Ryzen has absolutely nothing to do with how Skylake-X turned out. The only thing that changed is pricing and mid-core-count coming to HEDT to fill the upper-end of the range.

 

[bit_user]

Again, you are completely ignoring the previous 5 generations of Intel HEDT releases if you think Skylake X low core lineup was in any way modified in response to Ryzen.

Where the heck did you get that idea? I've said multiple times that Skylake-X's architecture was decided long before Ryzen was announced, Ryzen has absolutely nothing to do with how Skylake-X turned out. The only thing that changed is pricing and mid-core-count coming to HEDT to fill the upper-end of the range.

In case you didn't notice, his whole post was about pricing structure.

 

[InvalidError]

In case you didn't notice, his whole post was about pricing structure.

Would Intel have bumped 10 cores to $1000 without Ryzen? Every previous HEDT core count bump is associated with an increase in the LCC die's core count. SKL-X LCC still has 10 cores, which means that in all likelihood, Intel had no intention of bumping price points until ThreadRipper came along, telling Intel that it is unlikely to get away with it this time around. As a result, Intel got forced to rush-announce 12-18 core parts based on the HCC die instead.

Had Intel actually planned to bump HEDT core counts with SKL-X, it would have increased LCC to 12-14 cores.

 

[JamesSneed]

You left the 10 core off for Broadwell-e. Intel obviously did price adjust due to Ryzen and Threadripper.

Broadwell E
10 core - $1699
8 core - $1089
6 core - $617
6 core - $434

Skylake-x
10 core - $999
8 core - $599
6 core - $389

 

[jimmysmitty]

Competition is good, but intel haven't just plucked this design out of the air in response to AMD releasing Ryzen. The truth is if all intel wanted to do was compete with Ryzen they already have products that match or better them and all they'd need to do is drop the prices - admittedly hugely. I'm impressed by Ryzen, but it only wins on price.

I have to disagree a bit. If you compare Ryzen to the fastest non HEDT CPU, the 7700k, there are lot workstation and scientific tasks Ryzen is faster than Intel. Threadriper is coming to compete in the HEDT line so I think we can have more apples to apples discussions instead of comparing Ryzen to Intel's HEDT line which is a bit disingenuous. Will your statement hold true once all the chips on the table, maybe, but I think before we jump to conclusions we see Intels and AMD's HEDT chips then call it.

Ryzens 6 and 8 core will still have to be compared on a per core basis to Skylake-X. Price is always subjective, some people don't mind pending $120K on an Audi A8 that drops like a rock in resale value. Others hate spending more than $30K on a mid sized car. The other day I found a guy who had a Ryzen system with Quad GTX 1080Tis because you know, ePeen.

The only thing I can say for sure is that if Intel keeps up their normal I think their mesh will have a faster bus than AMDs fabric, much like QPI vs HT. That is to be seen though based on a lot of factors.

MIPS went 64 bit in 1999

This is definitely wrong. According to this, MIPS 64-bit R4000 was launched in 1991.

http://processortimeline.info/proc1990.htm

I personally remember the 64-bit race among RISC CPUs of the early 1990's, with MIPS among them. In fact, the Nintendo 64 launched in 1996 and had a legitimately 64-bit CPU.

AMD was dead last in the race to 64 bit among major architectures which is why they could only come up with the band aid 64bit extensions for x86 which ended up screwing us all and handicapping us to all the outdated garbage that was in x86.

This definitely makes you sound like an AMD-hater. Let's not forget that Intel had plenty of time to extend x86 to 64-bit however they wanted. Instead, they chose to use it as an opportunity to force their new IA64 micro-architecture on the mainstream computing world.

And if you want to talk about band-aid solutions, how can you overlook the hack that Intel bestowed upon the x86 ecosystem that is PAE - extending 32-bit x86 to address 36-bits of physical memory? This was purely to breathe a bit more life into x86 until their Itanium CPU could take hold (but IA64 was short on more than just time). Had AMD not launched x86-64, maybe Intel would've just updated this to support 40 bits, etc.

You might nit-pick a few decisions AMD made in x86-64, but I didn't hear much criticism at the time. IMO, it was eminently pragmatic, and included a number of worthwhile improvements, beyond just 64-bit addressing.

As for the rest of your CPU history lesson, I'd suggest anyone truly interested in the timeline of microprocessor developments would be much better served by reading this:

http://processortimeline.info

From my perspective, discovering it was the best thing to come of your post.

I'm not sure what it is about AMD fans that they think AMD invented everything when they've not really innovated much of anything. You don't see that with fanboys of other brands.

I dunno which is worse: fanboys or haters.

A lot of the innovators of the computing world are no longer with us, or are shadows of their former selves. It turns out that market timing & execution are things which matter just as much as an idea's genius and originality. The fact is that virtually all of the ideas underpinning today's computing products had their origins decades earlier. I knew a computer engineer who worked at Data General & Stratus, in the 1980's and 1990's. He said he thought they were innovating, only later to discover pretty much all of the original ideas they thought they had were already done by IBM and others, in the 1960's and 1970's.

Credit should go where it's due, but I don't really worry too much about who invented what. The things that really matter are bringing good products to market, pricing them reasonably, and supporting them properly. How much of their IP is truly original is fairly moot, IMO.

If you want to see real innovation in computing hardware, I think you'll need to look beyond semiconductors. For example, I find quantum computers fascinating - like cutting-edge physics experiments, except you can use them to run algorithms!

Intel was pushing IA64 because it was a pure 64bit platform which in most ways for 64bit computing is superior than x86-64. It was a logical idea for Intel to attempt to wipe out x86 in favor of 64bit without holding it back. Look at your system. It may be 64bit but the majority of processes running still run in 32bit and the majority of programs that install still install in 32bit as well.

The only true downside to Itanium was the x86 emulation. I am sure they could do it now without the major drop in performance.

As for innovation, plenty of the big companies are looking beyond semiconductors, Intel included. However until Quantum computing becomes viable semiconductors will be the headline.

fascinating how much innovation intel can show, when they get their ass handed to them by amd ...

Intel cannot overhaul a CPU design overnight. SKL-X's architecture was decided well over a year before Ryzen was announced. The only change is that ThreadRipper forced Intel to bring its 12-18 cores die to HEDT instead of remaining a server/workstation exclusive in the E5 series (whatever metal that maps to in the new naming scheme) at twice the price.

I think ignoring these people is all we can do. Pricing is the major change by Intel in response but for some reason there are a lot of people who assume that Intel can pull an entire socket and CPU lineup out of their behind in less than 3 months when all of this was probably in the works for a very long time. Actually thanks to their delays due to some difficulties with process tech they probably were planning on Skylake to launch in 2014 but delays happen and cause changes. Even so that means I would say more than two years ago Intel was working on this if not longer.

 

[InvalidError]

Actually thanks to their delays due to some difficulties with process tech they probably were planning on Skylake to launch in 2014 but delays happen and cause changes.

Broadwell was always supposed to be the 2014 architecture but slipped into 2015 for most SKUs and the desktop version didn't launch until mere months before Skylake launched practically on-schedule.

AFAIK, Skylake was always meant to be the 2015 model and if its schedule slipped, it didn't slip by much. That's why the desktop variant got phased out nearly instantly. (Not surprising given Intel's original intention of making Broadwell a BGA-only generation.)

 

[bit_user]

Intel was pushing IA64 because it was a pure 64bit platform which in most ways for 64bit computing is superior than x86-64. It was a logical idea for Intel to attempt to wipe out x86 in favor of 64bit without holding it back.

Oh, for sure. All the legacy of x86 means more complex instruction decoders, and that's probably one of the reasons they lost the mobile market to ARM. It would've been nice to wipe the slate clean and go with a new architecture. I followed the Itanium developments, at the time, and was excited for it to succeed.

But Intel made a big gamble, in IA64 - on static scheduling. The only places where static instruction scheduling has succeeded is in power-optimized embedded applications (ARM's in-order cores) and extremely parallel chips (i.e. GPU) that can use massive amounts of SMT (like 64 CPU-type threads per core) to cover stalls. Even performance-optimized embedded CPUs pretty quickly had to adopt out-of-order execution.

Of course, there were other mistakes that also fed into IA64's demise. But Intel's "EPIC" approach is the main technical disadvantage it had vs. x86-64. If you go back and look at the benchmarks, IA64 CPUs only ever excelled at linear algebra and DSP type applications, where the control flow and data access patterns were very regular and predictable. On things like database queries, web serving, Java execution, etc., I think they never had a lead over x86 and competing RISC architectures. At least, not ones with similarly sized caches.

 

[Shumok]

Unfortunately late to this discussion....but what I am curious about is, does the mesh increase scaleability in the die/yield sense? or strictly in the maintaining performance as core count goes up sense?

The impression I have is that AMD's infinity fabric allows AMD to scale up cores without having to worry about the increased yield problems with larger dies. Does the mesh do this too in some way, or is it strictly a means to allow higher core dies to communicate between cores more effectively?

If it's only better scaleability from a communications standpoint, it seems to me that AMD has a very important advantage with its infinity fabric that will allow them to do very high core count cpu's much more cheaply than Intel.

I understand Intel did have some kind of core cpu with dual dies or something several years back, and wonder what kind of communication the dies had and if it can be scaled this way too though. I am sure Intel will come up with something with their massive resources.

I do have to hand it to AMD though for seeming to do so much with so little. It sure does seem like the whole x299 cpu lineup is some kind of reaction to AMD, what with certain cpu's having limited functionality and the 18 core that isn't out for months and all.

 

[InvalidError]

The impression I have is that AMD's infinity fabric allows AMD to scale up cores without having to worry about the increased yield problems with larger dies. Does the mesh do this too in some way, or is it strictly a means to allow higher core dies to communicate between cores more effectively?

A larger die means fewer dies per wafer and a higher risk of fatals defect ruining each die. AMD's fabric allows it to tie up to eight dies together (four per MCM using dedicated fabric interconnect and two sockets by re-purposing half of the PCIe lanes) to avoid the yield issues that come with larger dies. It also enables AMD to dump lower quality dies that fail to make the R7/TR/EPYC grade into the mainstream R3/R5 lines.

Intel's mesh doesn't change the fact that larger dies have a higher chance of having a fatal defect than smaller ones and a lower die count per wafer simply because much fewer larger whole rectangles can be cut from the same size circle. If you could fit 200 of 'die A' on a wafer, you may only get 90 of 'die B' which is twice as large. To make it worse, if 'A' had 20 fatal defects on average, 'B' could have as many as 20 dead dies per wafer on average. 180 out of 200 is a much better yield than 70 out of 90.

As for Intel's first dual die CPUs, these communicated over the FSB which was the shared link to the chipset. Scaling is horrible as both CPUs end up competing for RAM and IO access on top of cache snoops between CPUs over that shared link, which is why the FSB was eliminated by integrating memory controller and main complement of PCIe lanes in the CPU.

Intel's mesh might be able to scale more evenly to higher core count per die, but larger dies are horrible for yield per wafer, which means much higher production cost per usable die. As for Intel's x299 lineup, ThreadRipper came up way too late in the development process to have any effect on the chipset or CPUs. The only reactionary changes are in the marketing and pricing with Intel adding HCC models to repopulate the top of the range.

 

[bit_user]

If it's only better scaleability from a communications standpoint, it seems to me that AMD has a very important advantage with its infinity fabric that will allow them to do very high core count cpu's much more cheaply than Intel.

AMD's approach is not without its tradeoffs, however. You might examine the tables toward the bottom of Page 2 of the i9-7900X review:

http://www.tomshardware.com/reviews/intel-core-i9-7900x-skylake-x,5092-2.html

The first column of the Ryzen table is surely a typo, and meant to be "Intra-Core latency". Their Intra-CCX latency is good - between that of Skylake and Skylake-X. However, once you leave the CCX (which is only 4 cores), then it shoots up. Once they go above 8 cores, we'll see another big jump. So, for workloads requiring a lot of inter-thread communication, we should see much better scaling vs. core count on Intel's mesh than with AMD's tree.

Now, how many workloads will that affect? Remains to be seen. But that's their gamble and I'd bet they did enough market research to know that most server workloads will run on it just fine.