AMD Announcement: ARMv8 Opterons In 2014

[NTMBK]

The problem that keeps returning is that, like with Atom. Slow gets expensive when you try to scale up the cores and extra platform costs needed.

The extra platform costs are precisely what Seamicro tries to eliminate- a ridiculous amount of platform is shared between multiple independent processors.

 

[Cerb]

Well no - but it's the same argument with PHI MC.

Not really. Some problems just don't scale out enough for something like a Tesla, but still waste time doing nothing on normal CPUs (thus SMT). And, HPC is both still growing, fairly profitable for companies in on it, and is a space where x86 has a major presence for years. It's not easy for Intel to scale a good many-core cache-coherent x86 setup down to size, but I think it's a safe bet that the Phi will have ready and waiting customers, and be moderately successful, even though Teslas will have higher theoretical performance. It's also the kind of thing that, once they get their foot in, they can scale upwards generationally, improving the CPUs, not merely the count of them.

I see it more likely that atleast half the "potential" players in this space dropping out after finding out the initial ROI and volume needed to get good margins can't support this many players.

Outside of shared web hosting, which is about the only sure thing, and which is definitely a big enough market for custom hardware packages, that wouldn't surprise me a bit. Even where successful, though, CPUs that buck ARM's trend of low performance per core will be needed just to compete with Atoms, much less big fat x86 cores.

 

[MisterMac]

Not really. Some problems just don't scale out enough for something like a Tesla, but still waste time doing nothing on normal CPUs (thus SMT). And, HPC is both still growing, fairly profitable for companies in on it, and is a space where x86 has a major presence for years. It's not easy for Intel to scale a good many-core cache-coherent x86 setup down to size, but I think it's a safe bet that the Phi will have ready and waiting customers, and be moderately successful, even though Teslas will have higher theoretical performance. It's also the kind of thing that, once they get their foot in, they can scale upwards generationally, improving the CPUs, not merely the count of them.

Outside of shared web hosting, which is about the only sure thing, and which is definitely a big enough market for custom hardware packages, that wouldn't surprise me a bit. Even where successful, though, CPUs that buck ARM's trend of low performance per core will be needed just to compete with Atoms, much less big fat x86 cores.

So in other words - a interconnect is all that ARM has riding for it for 2014?


That's not really a great bet.
Not that i see other bets - i think rory is finding plenty of skeleton's hence his weird CC and general "make it til you fake it - smile" self esteem letters.


Intel seems to be playing Shark in most of these new markets.
Let someone else create the foundation - and then buy your R&D way in-to market with the execution machine that is Chipzilla.

One could argue they're late to the mobile party - but i'm starting to wonder if it's a calculated move to make sure they can enter with enough margins to satisfy themselves and the board.


Even Otellini's comments on ARM being a Transmeta...well that's extremely arrogant i think most of us can agree.
And Otellini generally doesn't go hype-machine - so he's clearly extremely confident.

 

[ShintaiDK]

The extra platform costs are precisely what Seamicro tries to eliminate- a ridiculous amount of platform is shared between multiple independent processors.

If Seamicro cant sell this, how are they gonna sell ARM?

 

[Cerb]

So in other words - a interconnect is all that ARM has riding for it for 2014?

No. In other words, the Xeon Phi should be an answer to Tesla (AFAIK, there's very little else out there cheaper than IBM's BlueGene), and that all this large-scale ARM and Atom business is not even on its radar.

It's an unknown at any rate, but until larger-scale 64-bit systems can be made, nobody will even know how large the scale is, and everyone is trying to get it on it. I'm sure several companies involved are either preparing for a battle of attrition, or are spending little enough that they can gracefully bow out if it turns out to be a bad niche.

 

[pelov]

Yes, but it is far easier to develop a custom interconnect than to develop a custom ARM core. Even AMD could afford to buy an interconnect maker, and AMD financial position is subpar.

AMD Seamicro will make a difference on highly complex or specific designs. For your everyday servers the chip itself will matter a lot more than the interconnect.

So they are switching to cheap, vanilla ARM servers, most likely inferior to its competitors, which means that whatever gains comes from manufacturing the boxes themselves. I doubt that this business will be able to muster enough margins to sustain the company in the long run.

I don't think you're understanding this.

Right now, there are no major ARM vendors in server other than Calxeda and AMD. This has mainly to due with interconnect fabric. If you provide a high density cluster of servers, you need a way to cut down power and costs and a viable way to provide a net to connect these things together. There was a reason why Intel had close ties to SeaMicro prior to AMD's acquisition and why that acquisition was a big deal for AMD. SeaMicro is able to cut down PCB and power costs pretty significantly by using a PCIE interface to tie these high density clusters together.

Calxeda is also using the vanilla ARM core design just like AMD, though you'd have to figure that AMD has the personnel and experience (Jim Keller helped design the Apple A6) to dabble in their own custom ARM architectures. The reason they're using vanilla cores for the first iteration has to do with bringing these solutions to market first. It's sort of what Apple, and many other ARM licensees did with their early products.

https://docs.google.com/viewer?a=v&...NXVae_&sig=AHIEtbTwBYVxojojR-wLe98Uty9hFZtM1g

 

[ViRGE]

Right now, there are no major ARM vendors in server other than Calxeda and AMD.

Is Calxeda's ARM stuff even shipping? I thought no one was shipping anything ARM server until ARMv8?

 

[pelov]

Is Calxeda's ARM stuff even shipping? I thought no one was shipping anything ARM server until ARMv8?

There's ARM server stuff, currently ARM is at roughly AMD-like market share.

The big struggle was the 64-bit hurdle, as we saw in x86 as well. Calxeda will be shipping 64-bit A15s in 2014 just like AMD. Currently it's just Calxeda, SeaMicro and Marvell.

 

[Ferzerp]

There's ARM server stuff, currently ARM is at roughly AMD-like market share.

You're making things up. If you're going to make outlandish claims, please provide proof. There is pretty much 0% ARM server marketshare. I'm not even sure that anyone is shipping product. This is just more ARM pixie dust belief.

 

[mrmt]

Calxeda is also using the vanilla ARM core design just like AMD, though you'd have to figure that AMD has the personnel and experience (Jim Keller helped design the Apple A6) to dabble in their own custom ARM architectures. The reason they're using vanilla cores for the first iteration has to do with bringing these solutions to market first. It's sort of what Apple, and many other ARM licensees did with their early productshttps://docs.google.com/viewer?a=v&...NXVae_&sig=AHIEtbTwBYVxojojR-wLe98Uty9hFZtM1g

First Calxeda is using a vanilla ARM core in a custom SoC. They went great lenghts on power gating, tune the memory controler and other uncore features. AMD is going with a Vanilla ARM SoC, they have only the fabric to show, a fabric that will help on only the most complex designs, on most of the times they will be competing on the cut-throat environment of ARM micro-servers.

But I don't really think you get where things are going here.

AMD and Seamicro are very different things. One is a multibillion company that designs and sells entire chips. The other just designs the interconnect and assemble servers. You can see the former leveraging on the later but as sure as hell the later isn't going to save the former if the former is in trouble.

To give you an idea, Seamicro costed AMD 355 million, which puts AMD current and future cash flow expectations around 30-50 million >>per year<< depending on the rates you applied in the valuation. This should amount to a couple of hundred million dollars per year in revenues.

AMD in Q313 must deal with R&D and SGA expenses amounting to 450 million >>per quarter<<, which means that the Seamicro business probably isn't even in the top five business factors in AMD planning. As AMD is now its future depends on engineering new products and being able to sell it at competitive margins. And by selling ARM AMD is simply bringing lower margin products that will compete with their higher margin product.

ARM servers will eat margins, it will generate internal competition within the sales force, it will sooner or later divert resources from their x86 R&D. Unless AMD aims to ditch x86 in 2-3 years down the road, ARM simply does not make sense.

 

[pelov]

First Calxeda is using a vanilla ARM core in a custom SoC

AMD's design is a custom SoC as well. They're using the A15 core design but they're also integrating the fabric on-die.

But I don't really think you get where things are going here.

AMD and Seamicro are very different things. One is a multibillion company that designs and sells entire chips. The other just designs the interconnect and assemble servers. You can see the former leveraging on the later but as sure as hell the later isn't going to save the former if the former is in trouble.

You seem to think I'm an AMD fanboy who's lost sense of reality. I've already mentioned a few times on this board that I don't think AMD will survive 2013. This isn't going to change that, but that doesn't mean it isn't a smart move

 

[Idontcare]

Is Calxeda's ARM stuff even shipping? I thought no one was shipping anything ARM server until ARMv8?

According to DELL, HP, and Calxeda; DELL and HP have been shipping Calxeda-based servers for a few months now.

Now I haven't seen much fanfare made over this to date so I have assumed these businesses employ the term "shipping" in the same way AMD does when they say "we are shipping bulldozer in Q2" and what that really means is that it is 6 months away from being seen by an actual paying customer/end-user.

 

[NTMBK]

AMD is going with a Vanilla ARM SoC,

I don't really know where you're getting that from. The quote from Anandtech's article is this:

Today comes a much bigger announcement: AMD will be building Opteron processors based on a 64-bit ARM architecture. There are no product announcements today, but the 64-bit ARM Opterons will go into production in 2014. Today's announcement is about a processor license, not an ARM architecture license - in other words, AMD will integrate an ARM designed 64-bit core for this new Opteron.

(Emphasis mine) So in other words, doing what Calxeda do.

AMD and Seamicro are very different things. One is a multibillion company that designs and sells entire chips. The other just designs the interconnect and assemble servers. You can see the former leveraging on the later but as sure as hell the later isn't going to save the former if the former is in trouble.

To give you an idea, Seamicro costed AMD 355 million, which puts AMD current and future cash flow expectations around 30-50 million >>per year<< depending on the rates you applied in the valuation. This should amount to a couple of hundred million dollars per year in revenues.

AMD in Q313 must deal with R&D and SGA expenses amounting to 450 million >>per quarter<<, which means that the Seamicro business probably isn't even in the top five business factors in AMD planning. As AMD is now its future depends on engineering new products and being able to sell it at competitive margins. And by selling ARM AMD is simply bringing lower margin products that will compete with their higher margin product.

You're pulling those cash flow expectations out of a hat. Values of companies and acquisitions can change a lot- especially if they're pioneers in a new market, if that market takes off. (Whether or not it will is yet to be seen.)

The ARM chips won't be competing with their high performance Bulldozer based Opterons, it will be competing with low power designs like Atom and Jaguar. And don't forget about HSA- that should in theory make replacing Jaguar cores with ARM cores on any given chip design pretty straightforward, so the design costs can be lowered by sharing with Jaguar chips. And those Jaguar low power chips will still be coming, as indicated by this slide:

Note that the left hand column refers to both ARM and x86 CPU. By offering both ARM and x86 chips they can give customers precisely what they want in their Seamicro box.

 

[ViRGE]

The big struggle was the 64-bit hurdle, as we saw in x86 as well. Calxeda will be shipping 64-bit A15s in 2014 just like AMD. Currently it's just Calxeda, SeaMicro and Marvell.

That's what I thought. In which case it seems a little premature to say that somehow it's only going to be those 3-ish companies. NVIDIA will obviously have Denver, and anyone else can license the same cores AMD is using (I imagine we'll find out if anyone else has licensed them once ARM officially announces them).

Oh, and technically speaking A15 isn't 64bit (since it's ARMv7). Though the Atlas design AMD uses could end up being close to A15 in design if that's how ARM made it.

 

[mrmt]

You're pulling those cash flow expectations out of a hat. Values of companies and acquisitions can change a lot- especially if they're pioneers in a new market, if that market takes off. (Whether or not it will is yet to be seen.)

Not out of a hat, but out of a valuation model.

 

[USER8000]

AMD is using ARM A50 based cores:

http://www.theinquirer.net/inquirer/news/2221030/arm-announces-64bit-cortex-a50-series-processors

According to ARM these will be seen in products during 2014 and 2015.

 

[Ajay]

I can't believe that AMD is going with a generic core (especially since Papermaster has supposedly been working on this strategy for two years). Why they didn't become an ARM Architecture Partner and design some custom SOC's is something I don't get. AMD remains short sighted, or maybe just too cash strapped to do any significant engineering

 

[Idontcare]

I can't believe that AMD is going with a generic core (especially since Papermaster has supposedly been working on this strategy for two years). Why they didn't become an ARM Architecture Partner and design some custom SOC's is something I don't get. AMD remains short sighted, or maybe just too cash strapped to do any significant engineering

I know you know the answer already - $$$

Look at the kinds of R&D dollars the fabless ARM-users find themselves throwing towards developing their ARM-based products.

Just because the ARM license might be inexpensive, relatively, doesn't mean it is any less expensive to hire buildings full of design/layout/validation engineers versus filling the same buildings with x86 engineers.

AMD needs money. If they had money then they'd be in a better position in x86, or a better position in APU, or a better position in GPU, or a better position in HSA.

"Ambidextrous" means AMD is attempting to do all of the above and more by adding ARM-based products into the mix but doing it all with even less R&D (and 15% less people).

It is easy for you and me to monday-morning-armchair Papermaster's job by identifying the ideal strategy, but Papermaster is obviously trying to make the best of a rather dire situation. I'm sure he'd like to have the resources to go Apple on everyone, buyout Calxeda like Apple bought out PA Semi, and bring a custom ARM chip (ala A6) to the market.

But where is going to get that kind of money?

AMD has to cut corners somewhere for the ambidextrous strategy to work out financially, and using off-the-shelf ARM cores is one corner they can cut for now.

 

[CTho9305]

I know you know the answer already - $$$

Look at the kinds of R&D dollars the fabless ARM-users find themselves throwing towards developing their ARM-based products.

The problem I have with that assessment is that I don't think it justifies Bulldozer being a downgrade from the K7-derived family of cores for so many workloads. I'd think it would have been much cheaper, and lower risk, to continue incrementally improving K7/K8/Family 10h/Husky rather than trying a new design as complex as Bulldozer. The people who set out to do that design would have been well aware that they didn't have Intel-style resources available, and, even if they had planned for 2003 to continue forever, by 2006-2007 I think it was clear that the picture was changing. They had to keep tweaking and shrinking the old core to 45nm and 32nm... so at any point they could have accepted that the sunk costs on BD were lost, and redirected the company. Looking back at Anand's FX-8150 review, a 6-Husky-core chip with AVX, AES, and the updated implementation of Turbo Boost would've made a much better part. Keep in mind that the Phenom II X6 1100T is on 45nm (346mm^2), and holds its own against the 32nm Bulldozer (315mm^2). Unless the 45nm->32nm scaling was worse than about 0.7X area, 6 Husky cores would likely have been faster, and even an 8 core version would have been cheaper and lower power than Bulldozer.

I guess what I'm saying is, Bulldozer was too complex in too many dimensions (microarchitecture, custom circuit design, semicustom implementation, etc), and AMD could likely have done better with fewer resources by simply biting off less.

Just because the ARM license might be inexpensive, relatively, doesn't mean it is any less expensive to hire buildings full of design/layout/validation engineers versus filling the same buildings with x86 engineers.

I think licensing an ARM core saves piles of money. Verifying an x86 core takes huge amounts of manpower and compute infrastructure (they have thousands of machines running tests 24/7 through the whole design process). Buying a known-good core eliminates all of that.

 

[anongineer]

Is this not a war on two fronts? One of these days, Samsung will ship their own ARMv8 uarch implementation and with their ownership of a fab, become to ARM what Intel is to x86.

And for all the talk of fabrics and interconnects, the Facebook rep at the press event doesn't even care:

But ask Jay Parikh, Facebook's vice president of infrastructure, what he thinks of these hot fabrics and he’ll tell you he’s cold on them. “Proprietary fabrics are not something we are interested in,” he said in conversation after a panel session at AMD’s ARM server event.

Indeed, I have talked to Facebook engineers who said they want to use nothing but standard interconnects like PCI Express in their servers. They don’t even want Ethernet on server motherboards because it only makes it harder to upgrade rapidly changing CPUs.

Strangely enough, Intel's DMI is very similar to PCI-E.

 

[youshotwhointhe]

One of the most interesting possibilities is for AMD to allow customers to mix and match architectures in a single system. Allow people to run a VM for memcached on top of some low power ARM cores, and a VM for your database server on top of some high performance x86 cores, and have them all connected over the high speed fabric. Of course eventually you will be able to plug in APUs and GPUs (for things like image processing, encryption, etc).

 

[MisterMac]

Is this not a war on two fronts? One of these days, Samsung will ship their own ARMv8 uarch implementation and with their ownership of a fab, become to ARM what Intel is to x86.

And for all the talk of fabrics and interconnects, the Facebook rep at the press event doesn't even care:

Strangely enough, Intel's DMI is very similar to PCI-E.

You'd think that - but honestly i don't think whomever is in charge at samsung for SemiConductor has enough power to do it.


Samsung is one of the few players that could help shape ARM into something (Or shape Samsung into a ARM player capable of atleast attacking intel).

But i somehow don't see Samsung being agile enough to do that.
Especially with the whole Apple debacle.

They need to get a real move on - before Intels pure bruteforce breaks the barriers of ARMs defence.

 

[mrmt]

The problem I have with that assessment is that I don't think it justifies Bulldozer being a downgrade from the K7-derived family of cores for so many workloads. I'd think it would have been much cheaper, and lower risk, to continue incrementally improving K7/K8/Family 10h/Husky rather than trying a new design as complex as Bulldozer

It is clear that there were alternatives to Bulldozer inside AMD but somehow the Bulldozer crew won the argument inside the company and got the funds and resources needed to develop their project. This is when they lost it.

With 20/20 hindsight the best alternative would have been to kill Bulldozer in the design stage and come with a more conventional uarch based around phenom or something along Intel lines, beefier cores with SMT.

One of the reasons that I think they didn't go with a phenom die shrink is that they would have lost a lot of work they did for Trinity.

That said, a new die shrink of phenom processors may have been cheaper for AMD, and even improved their competitive position a bit, but it would not solve AMD fundamental problem, which is that Intel is opening a HUGE gap in both IPC and performance per watt on top of the process gap.

I guess what I'm saying is, Bulldozer was too complex in too many dimensions (microarchitecture, custom circuit design, semicustom implementation, etc), and AMD could likely have done better with fewer resources by simply biting off less.

I don't buy this argument that AMD could not correctly predict the workloads that servers and desktops would be running in 2012. Bulldozer isn't really competitive in performance per area and performance per watt with Core in both server and desktop workloads. In servers, AMD cannot even match Intel EP line, let alone the EX line.

Simply put, Bulldozer does not work even in their best case scenario. This does not point out to bad execution, but for a fundamentally flawed concept, much like Netburst was. No matter how Intel tuned Netburst, they could not consistently beat AMD, and Intel engineers had nothing to blame, no bug in the chip, no failure in the process, just the concept.

I think licensing an ARM core saves piles of money. Verifying an x86 core takes huge amounts of manpower and compute infrastructure (they have thousands of machines running tests 24/7 through the whole design process). Buying a known-good core eliminates all of that.

It does, but it hurts your market position.

If I am to buy an ARM server, I can ask prices to AMD and, let's say, generic manufacturer XYZ that also sells vanilla A5x based SoCs. If my design isn't too complex to the point of Freedom Fabric not being the deciding factor but a nice plus, as there isn't any differentiation between the two chips, I can just make the two race to the bottom.

AMD may win the order but at razor-thin margins. This model could work for Seamicro alone, but it does not work for a company that designs its own chips.

 

[boxleitnerb]

Is AMD in trouble, when even semiaccurate doesn't see this as a good thing?

http://semiaccurate.com/2012/10/30/amd-announces-an-arm64-server-soc-for-seamicro/

Fabric not flexible enough, cores too slow, ARM v8 available to the competition as well, too much competition, too little experience with ARM custom design.

 

[KompuKare]

Nothing too surprising there since SA Charlie hasn't had anything good to say about AMD's management for a long time. I think people misunderstand Charlie though: it's well known that he hates Nvidia and cannot stand Intel or Microsoft but that doesn't mean he really likes AMD (just maybe some of their products). The biggest problem for SA as a site (aside from Charlie's rants) is that if you have nothing good to say about some of the biggest IT players, you're not going to get much access and review samples.

Still, compared to some of the softly softly 'reviews' going around (*cough* *cough* Windows 8), SA is good fun.

One thing with the Bulldozer approach is that AMD seem to be very keen to have modular automated designs they can quickly tweak to build new chips. And it seems that approach is probably very popular with some technical managers there. Their pet project and even if the results were poor they won't abandon it now. So they've ended up in an even worse position: from having to sell a large Phenom die for less than Intel they now have to sell an even larger die for less and have lost a lot of perf/watt as well.