EETimes: Microsoft calls for 16-core server SoCs

[CTho9305]

Floating point exist as a module right now as fast as any of the current application need, it is just there hasn't been really any demand for FPU performance in ARM applications. Most of the ARM chips have a DSP core to handle task that are normally used for FPU like signal processing or video decoding.

That is the coolest part of ARM. If I want a chip with just basic cpu processing power, I can get that. If I want one with the basics and Java in hardware , I can get that too. It is like a set of lego blocks where you can design the hardware to have the features you want.

That's one of the interesting things about the ARM ecosystem, particularly compared to x86. When you need AES to go fast on ARM system, you slap a little DMA-capable state machine next to your core. When you need AES to go fast on x86, you add AES instructions and then run AES on your branch-predicting, out-of-order core with a complex load-store unit which shuffles data through multiple levels of caches. I can't imagine x86 gets anywhere close to an external accelerator when it comes to energy efficiency. Plus, you're tying up the core doing crypto work when it could be used for something else. Does Intel put accelerators on Atom SOCs to improve energy efficiency even when they compete with features of x86? I could imagine Intel preferring to keep as much as possible in the instruction set of the "main" core to make it harder to switch away from x86.

 

[P4man]

Hmmm.

Bobcat has 2 cores @ 9W TDP or 4.5W per core.

A Bulldozer HE has 16 cores @85W TDP or 5.3W per core.

Bobcat is the name of the CPU rather than the SoC, and it isnt anywhere near 4.5W per core. There is a 5W dualcore announced and that includes the same beefy GPU that probably takes up the majority of those 5W. Even if its just half, you are looking at closer to 1.25W per CPU core which is inline with AMD's claims of bobcat being sub 1W capable. Are you still confident BD will be faster per W on all workloads? I'm not.

As for ARM; it isnt too surprising MS would prefer an x86 solution at this point, considering they dont have any ARM software stack (besides phone software) - yet. The situation is rather different for someone running cloud servers on linux. Many of those machines facing the cloud run a very limited software stack, usually fully customized and/or apache/php based. x86 compatibility isn't a big deal there.

A bigger problem for ARM is memory bandwidth per core, as webserving tends to be bandwidth sensitive and current ARM SoCs are terrible in that regard as they are designed for mobile phones and the like, and are crippled with narrow and slow LPDDR memory busses. I wouldnt expect them to perform well here until someone fits a proper memory interface to them.

 

[IntelUser2000]

Are you still confident BD will be faster per W on all workloads? I'm not.

You need to include that the 85W figure for Bulldozer is not per core either. And even assuming Bulldozer will perform identically to the K10.5 core, and even in single thread, it'll be much faster than Bobcat. That's just per clock performance, include MHz and maybe the differences are far closer.

 

[cbn]

Floating point exist as a module right now as fast as any of the current application need, it is just there hasn't been really any demand for FPU performance in ARM applications. Most of the ARM chips have a DSP core to handle task that are normally used for FPU like signal processing or video decoding.

That is the coolest part of ARM. If I want a chip with just basic cpu processing power, I can get that. If I want one with the basics and Java in hardware , I can get that too. It is like a set of lego blocks where you can design the hardware to have the features you want.

http://eda360insider.wordpress.com/...sor-ip-core-need-a-new-category…superstar-ip/

A very interesting article I found on the upcoming Cortex A15. According to these marketing materials floating point has been increased 7x. (but integer has only ~ doubled compared to 1 Ghz Cortex A8). Is am presuming this is to allow ARM Cortex A15 processors to more easily handle programs originally written for x86. Maybe this extra floating point the beginning step toward making a design more compatible with server programs?

Memory, Floating Point, Gaming and other comparisons are also listed.

 

[P4man]

You need to include that the 85W figure for Bulldozer is not per core either. And even assuming Bulldozer will perform identically to the K10.5 core, and even in single thread, it'll be much faster than Bobcat. That's just per clock performance, include MHz and maybe the differences are far closer.

The issue here is performance per watt. Will a low power bulldozer be 4x faster than a 1 GHz bobcat ? Of course it will depend on workloads, but I have my doubts it will be true for across the board. Moreover, TDP is the wrong metric here, the more relevant metric would be closer to ACP, though even idle power matters quite a bit as well.

Its one thing to use a completely ignorant (/strawman) customer that doesnt know the difference between x86 and ARM ISA to "prove" a point, its something else to assume MS doesnt know what they are talking about. MS runs huge datacenters, some of the biggest in the world; If MS is pushing for 16 core Atom-like solutions, it is because they tested their workloads on xeons and opterons and found atom to deliver better performance/watt for their workloads. MS may or may not have tested bulldozer yet, but Im not expecting a quantum leap there especially not for network facing workloads, and from what Ive seen, bobcat is (far) better than Atom as well.

 

[LiuKangBakinPie]

Blue Gene Q 17 Cores
http://www.top500.org/blog/2010/12/06/about_16_and_17_core_processors
MS are full of crap they call for nothing which wont benefit their pockets

 

[P4man]

Bluegene is for HPC, thats an entirely different market.
As for MS calling for something that benefits their pockets.. yeah DUH.
But if it benefits MS, it probably benefits a lot of other datacenter operators as well. AMD would be wise to listen carefully.

 

[Mopetar]

http://eda360insider.wordpress.com/...sor-ip-core-need-a-new-category…superstar-ip/

A very interesting article I found on the upcoming Cortex A15. According to these marketing materials floating point has been increased 7x. (but integer has only ~ doubled compared to 1 Ghz Cortex A8). Is am presuming this is to allow ARM Cortex A15 processors to more easily handle programs originally written for x86. Maybe this extra floating point the beginning step toward making a design more compatible with server programs?

I believe the performance increase for FP calculations comes from adding better dedicated FP hardware. A huge improvement like that isn't hard if your previous performance levels weren't very good.

It's always been my understanding that for most cases, the vast majority of server load is integer-based. Better FP performance is nice, but isn't a huge part of most workloads.

I don't know if they are targeting anything in particular with those improvements. It may have just been a case of having extra space available so why not add the hardware to get a seven-fold performance increase.

 

[cbn]

It's always been my understanding that for most cases, the vast majority of server load is integer-based. Better FP performance is nice, but isn't a huge part of most workloads.

I don't know if they are targeting anything in particular with those improvements. It may have just been a case of having extra space available so why not add the hardware to get a seven-fold performance increase.

Yep, that is what I have read also. (eg, the rationale behind the AMD bulldozer server design.)

A huge improvement like that isn't hard if your previous performance levels weren't very good.

Yep, that totally makes sense. (eg, take a very small amount of floating point and multiply by x7. It would still only be a small amount of floating point.).

With that being said it would be interesting to compare atom and bobcat's floating point performance to ARM Cortex A15 FP.

 

[tweakboy]

Hmmm.

Bobcat has 2 cores @ 9W TDP or 4.5W per core.

A Bulldozer HE has 16 cores @85W TDP or 5.3W per core.

And I am pretty sure you would get better performance out of a bulldozer core than a bobcat core.

I am betting a 16-core Bulldozer HE would give you better performance per watt than a 16-core bobcat would. And you can run 2 or 4P, vs. 1P for an SOC.

Oh, and I agree on the ARM statement. If you think about it, there will be a windows for ARM but you also have to port apps, so while it seems interesting, you have far fewer apps supported.

Oh wow bulldozer has 16 cores wow plut HT for 32 logical cores. Oh boy What will Intel do to compete with that. AMD also won the hexacore race as their hexacore is at least 150 to 200 less then Intels hexacore. I feel bad for peep that got the SandyBridge B2 with this bug it has SATA bug hard drive slowing down something like that I read. You can however return your mobo to place you got it for replacement the B3 stepping Sandy.

 

[Ajay]

With that being said it would be interesting to compare atom and bobcat's floating point performance to ARM Cortex A15 FP.

Well, I think it will still be low with TI's OMAP 5. It's an SOC, so there allot more functionality to cram into the chip (including 2 Cortex cores and to M4s). ARM isn't near being a desktop competitor yet - but it's a different story with tablets.

 

[Ajay]

Oh wow bulldozer has 16 cores wow plut HT for 32 logical cores. Oh boy What will Intel do to compete with that. AMD also won the hexacore race as their hexacore is at least 150 to 200 less then Intels hexacore. I feel bad for peep that got the SandyBridge B2 with this bug it has SATA bug hard drive slowing down something like that I read. You can however return your mobo to place you got it for replacement the B3 stepping Sandy.

WTH?

 

[cbn]

ARM isn't near being a desktop competitor yet - but it's a different story with tablets.

http://news.cnet.com/8301-1001_3-20030497-92.html

February 3, 2011 6:38 AM PST
ARM's 2015 plan: Grab PC, server share
by Larry Dignan

ARM Holdings owns the mobile market when it comes to licensing chip architecture, but by 2015 the company expects to have a foothold in the PC and server market.

That's the primary takeaway from ARM Holdings' earnings conference call earlier this week. ARM is basically an intellectual property licensing company. As a result, it's a dominant processing company without actually manufacturing a processor. Instead, companies like Nvidia do the heavy lifting. Nevertheless, the message is clear: ARM Holdings is deadly serious about being a PC and server player, and at CES 2011 some of the pieces fell into place.

And why not? Microsoft is supporting ARM. Nvidia's Tegra chipset is landing server design wins.

It looks like ARM is planning desktop/server for 2015.

In the meantime. It does look like this Cortex A15 chip does fairly well in Integer. If we take a look at the chart below it looks like atom and cortex A9 (normalized at equivalent clocks) would have the same integer performance. Cortex A15 would be 50% better than both (at the same clock speed).

http://en.wikipedia.org/wiki/SPECint

SPECint benchmark is usually run on only a single CPU, even if the system has many CPUs. If a single CPU has multiple cores, only a single core is used; hyper-threading is also typically disabled,

A more complete system-level benchmark that allows all CPUs to be used is known as SPECint_rate2006, also called "CINT2006 Rate".

 

[Idontcare]

Personally I'm reluctant to dismiss any of these ARM prospects when it comes to the possibility of them upending x86, given the history of x86 itself.

Remember Cyrix and NexGen before they were bought out? These were two companies that came from nowhere, literally started out with a blank slate and a couple dollars, and developed CPU architectures that rivaled the best of the best in the business at the time.

Sure there is a barrier to entry in this market, and that barrier to entry has now passed the multi-billion dollar threshold, but look at ATIC and what motivated people who do have billions of dollars can do when they decide they are going to do it.

At this time, this is ARM's defining moment. This is circa 1998 with Linux versus Microsoft all over again, only it will be 2012 and ARM vs x86. Back in 1998 many of us had an impassioned view of the changing world order and we were quite confident Linux would unseat and upend Microsoft. That never happened, and in hindsight it is easy to understand why.

Will ARM succeed against Intel and AMD in ways that Linux could not in its marketshare battle with Microsoft?

Once again we have an impassioned audience who thinks so, but has ARM really done much to change the game or are they just talking about the game itself a lot more nowadays and we are too excited to realize that it is mostly hype (and hope) and not a lot of reality or practicality in the mix?

Nexgen enabled an AMD to not die at the hands of its floundering K5 design, will an ARM upstart project (denver?) or the absorption of a dying ember (DEC Alpha and the Athlon K7) be the phoenix that ensures a swan song is sung for the x86 industry?

 

[JFAMD]

This entire argument is worthless. Everyone is focused on the chips, but the real question is x86 vs. non x86.

ARM only stands a chance if people are willing to give up on running every application that they want.

As a compromise device, ARM may be competitive. The real quesiton to ask is does the world get to a point where only a browser and applets matter.

 

[Mopetar]

Really depends on how much of a pain it is to recompile the applications for ARM. It's probably a pain for some ancient legacy apps, but Apple has had their OS running on PPC, x86, and ARM, albeit in a stripped down fashion. If it becomes financially beneficial for other companies to support multiple architectures, they will start to do so.

The other possibility is that tablets and smart phones become the computers of the future and just overtake the current PC market. New software written for these devices starts to replace incumbents and recompiling doesn't even enter the equation. If that happens x86 is is in the same position where even if they get similar power characteristics as the ARM chips, if anyone wanted to use one they would have to give up running every application that they want.

The professional market would still be safe and servers would still run x86 chips, but the mass market could very well be taken over by ARM. Past architectures tried to compete head on with x86 and found a heavily entrenched market. ARM went off and started its own market that's grown at an enormous pace over the last decade. For the most part, ARM hasn't had to directly compete with x86 yet and it may well be a few more years before they really start to butt heads over the same market space.

 

[Ajay]

This entire argument is worthless. Everyone is focused on the chips, but the real question is x86 vs. non x86.

ARM only stands a chance if people are willing to give up on running every application that they want.

As a compromise device, ARM may be competitive. The real quesiton to ask is does the world get to a point where only a browser and applets matter.

The "Golden Handcuffs". x86 CPU designs can only improve incrementally because of the vast amount of software that runs on it and software vendors are often stuck with x86 because they have a large base of installed applications.

 

[cbn]

As a compromise device, ARM may be competitive. The real quesiton to ask is does the world get to a point where only a browser and applets matter.

Let me ask you another question: Why do most people buy high performance x86 desktop processors? Is there any reason to suspect this market will begin to decrease? (I know you don't work in AMD's Client devision, so this question will probably be really tough to answer.)

The reason I am asking is because at the moment it sounds like MS thinks low voltage server chips originally based on high power desktop designs are less optimal than dedicated bobcat or atom server designs.

But what keeps both Intel and AMD building server chips based on High Power x86 desktop? HPC server/High Power x86 desktop demand vs other types of servers needs requiring less power/lower cost? Who will win the tug of war and what will the tipping point be?

 

[P4man]

This entire argument is worthless. Everyone is focused on the chips, but the real question is x86 vs. non x86.

ARM only stands a chance if people are willing to give up on running every application that they want.

As a compromise device, ARM may be competitive. The real quesiton to ask is does the world get to a point where only a browser and applets matter.

The "world" isnt exactly a uniform place.

When it comes to mobile phones, tablets and increasingly, crossover products (think Atrix 4G which is a phone/laptop/"HTPC" hybrid) its x86 that has a large software handicap, rather than ARM. If these mobile OSs keep growing up, as iOS and Android/ChromeOS seem to be doing, you may have to eat those words in a few years.

In the server market no one cares about "browsers and applets". ARM is at a clear disadvantage here, but most servers are single purpose devices and either the software you need is linux based and available for ARM in which case there is no problem, or its not. Even ignoring the upcoming Windows port of ARM, and the possible server apps MS will port, I think the market share where ARM will be able to compete over the next years is significantly bigger than AMDs total server market share.

For general purpose desktops and laptops, ARM has an uphill battle ahead. But when you have 0% marketshare in a segment, an opportunity to grab just a few % basically upselling existing stuff running iOS/ChromeOS/Meego/etc there isnt much to loose. For ARM. AMD and Intel will have to decide if they want to compete on price, or allow ARM a foothold there. Neither would seem very a compelling prospect to me if I worked for an x86 only company.

 

[tweakboy]

Man thats crazy 16 cores Bulldozer. I saw in a graph tho it owuld be 10 to 12 cores.

I guess its the server line that has the Bulldozer HE with that many cores. Lets see how Intel combats this. From what I read their Ivy would be 10 to 12 core,, Im shure they will have server line thats 16 core or maybe Im wrong, well see..

 

[Ajay]

Man thats crazy 16 cores Bulldozer. I saw in a graph tho it owuld be 10 to 12 cores.

I guess its the server line that has the Bulldozer HE with that many cores. Lets see how Intel combats this. From what I read their Ivy would be 10 to 12 core,, Im shure they will have server line thats 16 core or maybe Im wrong, well see..

Last I read Intel's next server chip was going to be 10 cores (that was the 'sweet' spot). I don't know if it'll be SB (32nm) or IB (28nm).

 

[cbn]

For general purpose desktops and laptops, ARM has an uphill battle ahead. But when you have 0% marketshare in a segment, an opportunity to grab just a few % basically upselling existing stuff running iOS/ChromeOS/Meego/etc there isnt much to loose. For ARM. AMD and Intel will have to decide if they want to compete on price, or allow ARM a foothold there. Neither would seem very a compelling prospect to me if I worked for an x86 only company.

I agree with you about price. But I also wonder about Power?

With Ivy bridge rumored to have on package stacked DRAM, I can only imagine the next step after that might be die stacked DRAM. Maybe die stacked DRAM on that same type of high IPC design, but this using Intel's special low leakage silicon (to further control heat)

That leaves me wondering what AMD will do? Bulldozer is being marketed as very efficient processor for the die size. Bulldozer on special low leakage silicon vs bobcat on special low leakage silicon? Which way will AMD go?

 

[Mopetar]

If I were AMD I would focus on taking the APU to the next level. Don't just put a GPU on the die, but design a chip that melds the CPU and GPU together into one entity.

I don't know if AMD can afford the investment it would take to move into the market space currently dominated by ARM. I don't know if it would be worth it either. I don't expect that Intel will find too much success there, but they can afford to investment on the off chance that things work out in their favor.

 

[cbn]

If I were AMD I would focus on taking the APU to the next level. Don't just put a GPU on the die, but design a chip that melds the CPU and GPU together into one entity.

I've been looking forward to reading about floating point being moved to the GPU, but I don't know how close AMD is to achieving this or how much money it would take to develop?

Maybe this is one of the questions AMD is asking itself: "Do we spend a greater proportion of resources on x86 to keep ourselves competitive with Intel or should we make a shift towards advancing our Radeon GPU and not worry about which CPU it uses?"

 

[cbn]

I don't know if AMD can afford the investment it would take to move into the market space currently dominated by ARM. I don't know if it would be worth it either. I don't expect that Intel will find too much success there, but they can afford to investment on the off chance that things work out in their favor.

I don't understand the economics of an ARM license, but it appears to me that Google Android is progressing at an absolutely rapid rate.

In fact, recently one of the lead engineers for MS kinect moved to Google. More information here

So maybe this is a developing ARM ecosystem AMD cannot afford to pass up? (I'm under the impression, right or wrong, that Denver may contain a fairly large GPU. To me this implies a focus on graphics/GPGPU heavy development if it ever came to desktop).