EETimes: Microsoft calls for 16-core server SoCs

[cbn]

For general purpose desktops and laptops, ARM has an uphill battle ahead.

Sounds like we are going to see a good battle pretty soon. Will ARM be able to challenge for WebOS desktop/laptop Wins?

http://www.anandtech.com/show/4161/analysis-of-hps-event-its-raining-webos/4

Although HP's Todd Bradley didn't go too deep into it, he did mention that HP was definitely going to bring webOS to 'other new form factors', including printers and PC's. Yep, you heard that right. HP is planning on porting over webOS to its PC's sometime in the future. Now this announcement has multiple interesting implications. For one, this will give Intel the opportunity to have its x86 CPU's power webOS which in itself could lead to some interesting prospects in the tablet/netbook arena. Secondly, considering the number of PC's, laptops and netbooks HP sells world wide, having webOS run natively (or virtualized) on those machines would definitely help build-up a vast userbase of webOS 'devices'. And considering the fact that user install base is probably one of the biggest factors that decides how much developers would want to invest their resources in supporting a platform, HP would do itself and webOS very good by opening up this huge potential install-base to the devs. And finally, I can only imagine the prospects of technologies such as 'Touch-to-share' if it were to roll-out on to PC's as well. Hopefully, if HP were to continue onward with the same momentum they demonstrated today, what is listed above may actually happen a lot quicker than we can imagine.

 

[P4man]

Yep, a perfect storm is brewing. A *lot* of things are happening simultaneously that on their own, might not have much chance of unseating wintel, but combined Im pretty sure will cause a revolution over the next 10 years.

To name a few elements making up this storm, there is the single thread performance scaling brick wall, and the termal brick wall favoring ARM's ISA. There is a very strong trend of every day computing moving towards increasingly mobile devices, laptops, netbooks, tablets and smartphones. There is a huge competition in this space with a ton of exiting new operating systems with major companies backing them. None of these OS's are tied to x86, and most of them dont even run on x86. Software developers are flocking to these new platforms. There is a trend towards cloud computing and even remote rendering for the heavy lifting. OEMs are increasingly integrating vertically, wanting ownership over the entire product, from software to hardware. Apple, Samsung, Microsoft, HP, Google, RIM, even intel; they are all ensuring they have control over the product design, often even the CPU/SoC, the OS and increasingly parts of the applications that run on top of that.

Put all of these elements together, and its pretty tough to assume wintel will continue to dominate it has for the past decades. This is not like windows facing off against OS/2 or linux, nor it is like x86 vs PowerPC.

 

[Ajay]

I agree with P4man. But, just to fuel the flames, here's an article pointing out some trends: http://www.techrepublic.com/blog/hiner/the-next-wintel-android-plus-nvidia/7454?tag=nl.e101

 

[Mopetar]

I agree with P4man. But, just to fuel the flames, here's an article pointing out some trends: http://www.techrepublic.com/blog/hiner/the-next-wintel-android-plus-nvidia/7454?tag=nl.e101

I could agree that Android may become the next Windows, but I don't think it's Nvidia that's going to become the next Intel. The Tegra 2 honestly isn't all that great. The CPU cores are basic Cortex-A9s and the GPU portion of the chip isn't much better than an SGX.

Most of the reason that the Tegra 2 is seeing such wide use is because it was ready before most competitors had their offerings ready. After reading reading a few reviews of the Optimus 2X, dual cores might not be terribly useful for phones. It's also interesting that HP decided to go with a high clock speed, single core chip for the Pre 3.

Even if x86 ends up going away, you can be fairly certain that the best ARM chips are going to end up being made by Intel and AMD. Intel also has more advanced fabrication processes than anyone else so they'll have another leg up there as well.

 

[P4man]

I could agree that Android may become the next Windows, but I don't think it's Nvidia that's going to become the next Intel. The Tegra 2 honestly isn't all that great. The CPU cores are basic Cortex-A9s and the GPU portion of the chip isn't much better than an SGX.

Most of the reason that the Tegra 2 is seeing such wide use is because it was ready before most competitors had their offerings ready. After reading reading a few reviews of the Optimus 2X, dual cores might not be terribly useful for phones. It's also interesting that HP decided to go with a high clock speed, single core chip for the Pre 3.

Even if x86 ends up going away, you can be fairly certain that the best ARM chips are going to end up being made by Intel and AMD. Intel also has more advanced fabrication processes than anyone else so they'll have another leg up there as well.

I agree that future dominance by nvidia is probably a bit overly optimistic (or pessimistic if you dont like the company heh). Tegra is a fine SoC, but not earth shattering when compared to omap4 or samsung orion and others. However, it is remarkable that nvidia as a totally new player in this market, managed to get the new Cortex A9 in to the market pretty much a year before anyone else. I think that says something about nVidia, or more likely, its competition. If tegra 3 rumors are correct (quad core 1.5 GHz A9 with massively better GPU) and nVidia can actually deliver that this year, they will again be way ahead of all the rest. This could become a pattern.

However, I dont think nVidia's main strength is in designing better chips than Samsung, TI, qualcomm etc. Their real strength lies in software. No other company has their expertise when it comes to writing drivers, optimizing software (think games, but also stuff like adobe flash) and even GPGPU for the future. If tegra is now Android's reference design, a big part of that would be nVidia's ability to work with Google, with Adobe and the rest. None of the other SoC vendors can match nVidia there.

Another point we shouldnt forget is project denver. Almost nothing is known of it now, and it might be a complete dud, but few other companies are as well positioned as nVidia to make a full custom high performance ARM implementation and tightly integrate it with a powerful GPU. The only other company that might have been able to pull that off was AMD, but they have abandoned the ARM market and sold all their IP related to it (huge blunder IMHO).

Intel? Meh. They will only jump on the ARM wagon again when its way too late, if at all. Besides, last time they tried (Xscale), they hardly outclassed the competition. They were at best, an also-run. Add to that their terrible trackrecord in GPU and GPU drivers, and I wouldnt hold my breath.

So while I agree nVidia isnt going to become the new intel just because how awesome tegra is(nt), I am putting my money on them becoming the market leader here over time, because of their software, gaming and GPGPU clout.

 

[cbn]

To name a few elements making up this storm, there is the single thread performance scaling brick wall, and the termal brick wall favoring ARM's ISA.

That is really interesting. How does having legacy support reduce IPC/clockspeed potential?

OEMs are increasingly integrating vertically, wanting ownership over the entire product, from software to hardware. Apple, Samsung, Microsoft, HP, Google, RIM, even intel; they are all ensuring they have control over the product design, often even the CPU/SoC, the OS and increasingly parts of the applications that run on top of that.

This makes a person wonder how Microsoft could use that ARM architectural license they bought last year?

How would a partially vertically integrated MS view x86?

 

[cbn]

Another point we shouldnt forget is project denver. Almost nothing is known of it now, and it might be a complete dud, but few other companies are as well positioned as nVidia to make a full custom high performance ARM implementation and tightly integrate it with a powerful GPU. The only other company that might have been able to pull that off was AMD, but they have abandoned the ARM market and sold all their IP related to it (huge blunder IMHO).

I thought AMD once owned a MIPS based mobile CPU company?

Or maybe you are talking about the mobile GPU company they sold off (now owned by Qualcomm)?

 

[P4man]

That is really interesting. How does having legacy support reduce IPC/clockspeed potential?

It doesnt. But x86 complex instruction decoding makes for bigger, leakier and hotter chips, and thats no good when you are power limited or when the only real way to scale performance is high core counts .

This makes a person wonder how Microsoft could use that ARM architectural license they bought last year?

How would a partially vertically integrated MS view x86?

Pragmatically Im sure. MS makes like 95% of its profits on x86 and thats not about to change dramatically, even if they port windows to ARM (which I suspect, wont be very succesful against android/chrome/webos etc). I dont know why they got that license, but I think its for xbox, which already is a non x86 architecture.

I thought AMD once owned a MIPS based mobile CPU company?

Or maybe you are talking about the mobile GPU company they sold off (now owned by Qualcomm)?

I mean both really. their MIPS business wasnt really about mips, it was about digital tv image processors, which could have been an interesting part of the puzzle in the raging battle for "smart tvs" between google and intel and nvidia. AMD could have been a player too. That and their mobile phone GPU business which they indeed sold to qualcomm.

Anyway, this is getting pretty far OT.

 

[cbn]

Pragmatically Im sure. MS makes like 95% of its profits on x86 and thats not about to change dramatically, even if they port windows to ARM (which I suspect, wont be very succesful against android/chrome/webos etc). I dont know why they got that license, but I think its for xbox, which already is a non x86 architecture.

I was thinking all this mobile OS morphing into Tablet/Desktop OS competition might force MS to lower prices on Windows 7 Home Premium.

If that happened, MS might try to vertically integrate to compensate for reduced software volume and/or profits. As far as Xbox goes I think you are probably right on that being the area of vertical integration (with ARM). Windows 7 Phone and Big Windows (whether we see Cloud versions or not) have too many partners to make vertical integration worthwhile.

 

[Mopetar]

I agree that future dominance by nvidia is probably a bit overly optimistic (or pessimistic if you dont like the company heh). Tegra is a fine SoC, but not earth shattering when compared to omap4 or samsung orion and others. However, it is remarkable that nvidia as a totally new player in this market, managed to get the new Cortex A9 in to the market pretty much a year before anyone else.

Part of me thinks they shipped ahead of everyone else because they really needed to get into this market space. Intel basically cut them out of the chipset business and is working to make discrete cards useless. AMD bought ATI and isn't going to do them any favors. AMD/ATI's last two graphics card cycles have been better or on par with Nvidia's offerings so they can't count on dominating the market to reap massive profits. They really had nowhere else to go.

Denver probably won't pan out. Projects like that almost never do. It'll end up being another Larrabee or Itanium. The Tegra 3 might be useful for netbooks running ARM Windows, but for phones or tablets it's probably going to be overkill. Phones aren't really taking much advantage of the second core in the Tegra 2 and most of the heavy lifting on phones is done by dedicated hardware rather than the ARM cores. Tablets at least have a chance of becoming productivity devices that will use multiple cores, but for phones it may be a lot less useful.

I can conceivably see Intel wanting to buy Nvidia in the future depending on how well AMD is able to integrate ATI graphics with their CPU. Intel is much further behind AMD in the graphics side of the equation than AMD is behind Intel on the CPU side. Any other long-term future seems unlikely by comparison.

 

[cbn]

Anyway, this is getting pretty far OT.

Okay lets get back to talking about servers. (something I am trying to learn about.)

My guess is that MS made this very polite request for more efficient x86 server parts because ARM is trying to sell them on new ideas?

Speaking of new ideas I'll throw a crazy one out there.

This talk of HP/WebOS/ARM/Vertical integration (we did in this thread) makes things interesting because last year HP Labs announced research for methane (from anaerobic digestion) powered server farms. More information here. IMHO, if HP pulled this off it would be vertical integration++.

But how are they planning to deploy these in the future? Is ARM (or an ARM partner) trying to sell HP on a Low power 3D 64 bit Server chip? This with water cooling channels built in to enhance chip reliability and improve transfer heat to the anaerobic digester?

EDIT: More info on hot water and anaerobic digestion here.

Inside the digester, suspended heating pipes allow hot water to circulate. The hot water heats the digester to keep the slurry at 25°C to 40°C (77°F to 104°F), a temperature range suitable for methane-producing bacteria.

With regard to 3D stacking this ARM Press release mentions:

“With scaling in semiconductors and 3D stacking technology, future server machines will be more prone to faults and errors. Thus, reliability will be as an important factor in designing future servers as performance and power,”

 

[Ajay]

It doesnt. But x86 complex instruction decoding makes for bigger, leakier and hotter chips, and thats no good when you are power limited or when the only real way to scale performance is high core counts .

Nah, just bigger, buggier, slower compilers that have to play tricks to use x86 instructions even though at their core modern x86 CPUs are RISC processors.

The only way to get large increases in modern CPUs is to go multi-core - that's because higher IPC is too expensive (thermally, IIRC) not because of a flaw in x86 per se.

 

[P4man]

Nah, just bigger, buggier, slower compilers that have to play tricks to use x86 instructions even though at their core modern x86 CPUs are RISC processors.

The only way to get large increases in modern CPUs is to go multi-core - that's because higher IPC is too expensive (thermally, IIRC) not because of a flaw in x86 per se.

I agree with the latter; x86 doesnt have an IPC handicap, if anything its the contrary. But the fact everyone needs to go wide multicore makes the handicaps x86 does have, hurt a lot more, because any ISA becomes either (or both) power and die size limited when performance scaling becomes a matter of throwing ever more cores at it.

As for the former, I disagree. x86 might be "risc at heart', but they still need all the complex decode logic, trace caches and what not. Those transistors arent free, not in size, not in power consumption and not in static leakage. If someday intel decides to expose the internal risc ISA to compilers and apps, and apps are recompiled directly for that ISA rather than x86, then perhaps, one day, they might shed their legacy baggage and compete on more equal footing with ARM. But that would probably also break all binary compatibility between intel and amd, and most likely even between various intel chip generations.

 

[Voo]

Nah, just bigger, buggier, slower compilers that have to play tricks to use x86 instructions even though at their core modern x86 CPUs are RISC processors.

Not really. Most compilers use only a subset of the ISA and the things that make compilers more complicated (vectorstuff, vliw,..) are there for a performance reason and may be supported in other architectures as well.
Also x86 decoding IS complicated, slow and takes lots of energy. Sure the backend isn't worried about it, but that doesn't mean you don't have to support the thousands of x86 instructions. Variable length, fixed registers, duplicate instructions, prefixes, half a dozen operation modes - there's so much really unpleasant stuff in there it's not funny.

If it was easy to implement the decoding efficient, Intel wouldn't have needed to change that with SB to make it more efficient (and try to use it as much as possible for energy reasons; see their instruction cache).

In the larger picture it may not matter that much, but if you can save energy and die size without any consequences (which is at least true in the mobile space where x86 backwards comp is minor at most), it's a nice bonus.

PS:

[..]that's because higher IPC is too expensive (thermally, IIRC) not because of a flaw in x86 per se.

IPC means instructions per cycle and describe the average number of instructions you can execute in one cycle. The problem you mean, is that you can't increase the frequency (i.e. how many cycles per second) without running into energy problems. But IPC is actually the best way to increase performance throughout the board.. but obviously you need some good ideas how to do that and it's not as easy as it sounds

 

[Ajay]

PS:
IPC means instructions per cycle and describe the average number of instructions you can execute in one cycle. The problem you mean, is that you can't increase the frequency (i.e. how many cycles per second) without running into energy problems. But IPC is actually the best way to increase performance throughout the board.. but obviously you need some good ideas how to do that and it's not as easy as it sounds

Yes, I know. I should have said for single threaded performance increases, increasing IPC is expensive (power wise, if we don't consider multiple cores) and all the low hanging fruit has been picked.

 

[Voo]

I should have said for single threaded performance increases, increasing IPC is expensive (power wise, if we don't consider multiple cores)

Don't agree with that. Increasing IPC usually increases power efficiency and most of the time isn't especially power intense. At most we're talking about a constant increase, contrary to quadratic or linear with frequency/voltage.. at least I can't think of any examples (well OoO execution is probably the most expensive example.. and even that is getting implemented in most CPUs, even those targeted at the mobile space [see comming ARM cores]). The problem is just the R&D cost involved in increasing IPC.

 

[cbn]

Well MS could develop Windows Phone 7 into some type of Tablet/Desktop OS (similar to Android) for ARM processors.

So with this strategy MS has all bases covered:

1. Windows 7 Phone/Tablet for ARM non-cloud
2. Windows 8/Next for Cloud
3. Windows 7 for x86 non-cloud

P.S. Maybe the following quote supports this idea also-->http://www.xtremesystems.org/forums/showpost.php?p=4733090&postcount=34

Windows 7 Phone Just got a major boost today. http://www.anandtech.com/show/4162/nokia-and-microsoft-announce-strategic-partnership

Could Windows Phone based OS and a Windows Cloud based OS become the major focus of MS efforts? (possibly leaving Windows 7 Home Premium as a long lived product that receives slower development)

 

[CTho9305]

I thought AMD once owned a MIPS based mobile CPU company?

Alchemy. The product line was sold to RMI.

I mean both really. their MIPS business wasnt really about mips, it was about digital tv image processors, which could have been an interesting part of the puzzle in the raging battle for "smart tvs" between google and intel and nvidia. AMD could have been a player too. That and their mobile phone GPU business which they indeed sold to qualcomm.

I think you're confusing the Xilleon TV stuff and the handheld stuff that ATI used to do with the completely-unrelated Alchemy stuff (though I have no idea what CPU instruction set, if any, those products used).

I agree with the latter; x86 doesnt have an IPC handicap, if anything its the contrary. But the fact everyone needs to go wide multicore makes the handicaps x86 does have, hurt a lot more, because any ISA becomes either (or both) power and die size limited when performance scaling becomes a matter of throwing ever more cores at it.

As for the former, I disagree. x86 might be "risc at heart', but they still need all the complex decode logic, trace caches and what not. Those transistors arent free, not in size, not in power consumption and not in static leakage. If someday intel decides to expose the internal risc ISA to compilers and apps, and apps are recompiled directly for that ISA rather than x86, then perhaps, one day, they might shed their legacy baggage and compete on more equal footing with ARM. But that would probably also break all binary compatibility between intel and amd, and most likely even between various intel chip generations.

It's not just fetch and decode. The whole backend is full of extra ugliness to handle x86 (memory segmentation, misaligned memory access, self-modifying code, 80-bit floating point, etc etc etc). If someone selling x86 tells you "the decoder is only X%"...keep in mind that they're selling x86 and pay close attention to the exact wording they use.

 

[P4man]

Alchemy. The product line was sold to RMI.
I think you're confusing the Xilleon TV stuff and the handheld stuff that ATI used to do with the completely-unrelated Alchemy stuff (though I have no idea what CPU instruction set, if any, those products used).

xilleon was what I meant, and it was mips based (well, not that mattered a whole lot). I had forgotten about alchemy.

 

[Ajay]

The problem is just the R&D cost involved in increasing IPC.

I agree with that and it's why I mentioned that all the low hanging fruit has already been picked. As to the rest - I suppose it's not clear (at least to me). CPU designers are using additional die space to add (or increase) features outside the pipeline itself to get higher IPC. I suppose the most notable increase is larger total on chip cache - but given the TDP on these designs, the impact doesn't seem to be all that significant.

 

[Idontcare]

The problem is just the R&D cost involved in increasing IPC.

This, same is true of process tech too.

We had HK/MG at 90nm but it wasn't worth the cost to implement into production at the time.

Just like carbon nanotube-based xtors today. We can do it, but no one wants to spend $3m per CPU regardless its performance.

We are never operating at the bleeding edge of capability, we are always operating at the bleeding edge of affordability. Be it design or manufacturing.

It takes government to do things that push the limits of capability in the absence of cost concerns...landing on the moon had nothing to do with affordability, which is partly the reason why we haven't gone back in 40yrs.

 

[Modelworks]

Software compatibility is mentioned a lot when talking about x86 and ARM and that has its good and bad sides. X86 is tied to the software, when Intel releases a new cpu they have no choice but to keep legacy support for that software. They can't radically change the design. ARM has been able to allow changes to whatever works best for the design regardless of compatibility with other software. The ARM cpu was made for a specific product and used with specific software so it didn't matter if the next chip version wasn't compatible. This is the one are where i have concerns. The flexibility has been one of ARM best strengths and trying to make it a universal platform is going to come at a cost of having to maintain compatibility with each revision.

If my software requires a cpu with java jit hardware and I want a faster desktop I will have to have a cpu with that same hardware. It really is a problem because the current way ARM works is if the hardware feature is not supported then the software doesn't work, emulation really isn't supported. It is like trying to run directx11 on a directx9 card , you can't emulate it, you either have the hardware or you don't.

That is a concept I think most people coming from x86 will need to understand. The ARM way of doing things is to use something dedicated in hardware for the task where the x86 way is to do functions in software. If you want an x86 application to run faster you brute force it with a faster Mhz cpu. ARM would add another DSP or dedicated processor to make the task go faster. ARM has never really been about raw clock speed and using software when a hardware module can do it better.


One area where ARM has a huge advantage over x86 is the open nature of the platform. Anyone can design their own ARM core. You don't have to be a corporation. Just go to the ARM site , register and download all the programs. All the programs are free of charge. Then when you have a design you like , you find a fab to make it for you.

http://www.arm.com/products/processors/designstart-processor-ip/index.php

Things like mali graphics are totally open to the public, unlike the closed nature of nvidia or ati cards.
http://www.malideveloper.com/developer-resources/tools/index.php

 

[Mopetar]

I agree with P4man. But, just to fuel the flames, here's an article pointing out some trends: http://www.techrepublic.com/blog/hiner/the-next-wintel-android-plus-nvidia/7454?tag=nl.e101

After seeing results for the new TI OMAP 4430 and Samsung Exynos, Tegra 2 doesn't look particularly amazing. The OMAP4's SGX540 outperforms it, although the one in the OMAP may run at higher clock speeds than the SGX540 normally does so the numbers may not be as great as they seem.

Sony's NGP is supposed to launch later this year and will use the SGX543 and there have been rumors that Apple will use the SGX543 in the iPad 2 and iPhone 5. Sony has said they're going to ship in time for the holiday season and depending on how true these rumors are regarding Apple products, we may see the SGX543 in products as early as April - July.

Nvidia's probably not gaining dominance in this market area any time soon. If the Tegra platform were an undisputed king for about a year they might have had a shot if they were to get Google to optimize for their GPU, but it doesn't look like they're going to have a very large window. Maybe Tegra 3 will put them ahead again, but there's going to be a wave of chips with similar or better performance before the next generation Tegra parts come out as I don't believe the Tegra 2 3D improves the graphics over the Tegra 2.

 

[cbn]

Nvidia's probably not gaining dominance in this market area any time soon. If the Tegra platform were an undisputed king for about a year they might have had a shot if they were to get Google to optimize for their GPU, but it doesn't look like they're going to have a very large window. Maybe Tegra 3 will put them ahead again, but there's going to be a wave of chips with similar or better performance before the next generation Tegra parts come out as I don't believe the Tegra 2 3D improves the graphics over the Tegra 2.

Maybe the current hardware is not that different performance wise, but Nvidia appears to be at the beginning stages of using some kind of vertically integrated app store.

http://www.nvidia.com/object/tegra-zone.html

If it pans out, it might allow them to successfully launch future Tegra hardware that is more differentiated from the competition (rather than forcing them to use "lowest common denominator" strategy.)

 

[Idontcare]

Nvidia's probably not gaining dominance in this market area any time soon. If the Tegra platform were an undisputed king for about a year they might have had a shot if they were to get Google to optimize for their GPU, but it doesn't look like they're going to have a very large window. Maybe Tegra 3 will put them ahead again, but there's going to be a wave of chips with similar or better performance before the next generation Tegra parts come out as I don't believe the Tegra 2 3D improves the graphics over the Tegra 2.

Their situation is sort of similiar to Intel's in that when they try and step outside of their core area of expertise they find themselves on the low-end of the learning curve while competing with companies that have years more experience in the business segment.

TI is no slouch in ARM, but TI is surrounded by very adept competitors in the field.

Nvidia showing up is kinda like Cyrix or IDT or Transmeta showing up to the x86 party some 15 yrs after it got started. They hit with a splash, lots of hype, and dissipated nearly as quickly.

Will Nvidia be different? If they are then it will be outside the norm.