[] IBM unveils Power8 and OpenPower pincer attack on Intel’s x86 server monopoly

[NTMBK]

"Pincer attack"? "Server monopoly"? Yeesh. This article reads like a PR piece...

 

[norseamd]

SPARC, POWER and Itanium is in rapid decline and have been for ages now. And I have proven POWER is in free fall as well from IBMs own numbers.

SPARC also went open license. And it kept its free fall to being completely irrelevant today.

I know the best thing by some would be the demise of Intel. But the fact is x86 keeps rapid expanding its lead over RISC based server uarchs. And x86 now sits on over 90% of all server revenue. Just 2 years ago it was just over 80%.

http://www.forbes.com/sites/greatsp...ed-since-its-acquisition-of-sun-microsystems/

POWER is only going open due to being dead, its a last resort attempt to keep the dead patient alive for a little longer.

The EX class x86 CPUs is simply slaughtering the RISC based CPUs in all important metrics.

thought you loved the itaniums

 

[ShintaiDK]

thought you loved the itaniums

It doesnt change the facts. Nice try tho.

 

[TrulyUncouth]

SPARC, POWER and Itanium is in rapid decline and have been for ages now. And I have proven POWER is in free fall as well from IBMs own numbers.

SPARC also went open license. And it kept its free fall to being completely irrelevant today.

I know the best thing by some would be the demise of Intel. But the fact is x86 keeps rapid expanding its lead over RISC based server uarchs. And x86 now sits on over 90% of all server revenue. Just 2 years ago it was just over 80%.

http://www.forbes.com/sites/greatsp...ed-since-its-acquisition-of-sun-microsystems/

POWER is only going open due to being dead, its a last resort attempt to keep the dead patient alive for a little longer.

The EX class x86 CPUs is simply slaughtering the RISC based CPUs in all important metrics.

Wouldn't this be a different situation though? Intel looks to be getting a full-on monopoly in servers. It is very much in the interest of server consumers that they have a second source for CPUs. Assuming the performance gap isn't too huge, wouldn't a big player be smarter to go with the more open alternative? Especially if they can integrate their own IP in like an ARM licensee?

These are different times than back in the SPARC days.

 

[Arachnotronic]

Wouldn't this be a different situation though? Intel looks to be getting a full-on monopoly in servers. It is very much in the interest of server consumers that they have a second source for CPUs. Assuming the performance gap isn't too huge, wouldn't a big player be smarter to go with the more open alternative? Especially if they can integrate their own IP in like an ARM licensee?

These are different times than back in the SPARC days.

AMD offers "alternatives" to Intel's server chips, but they're so laughably out of Intel's league that AMD's market share is next to nothing.

People who buy this kind of gear are worried about total cost of ownership...the upfront hardware costs are peanuts compared to how much it actually costs to run these data-centers. The company that can offer the best performance/watt/$ will win the orders, and that would be Intel in just about every server segment.

Nobody's going to buy 2nd tier product just because they want to join the "anybody but Intel" crusade that so many on these boards seem to blindly be a part of.

 

[blackened23]

^That pretty much sums the situation up nicely. Intel isn't a monopoly in servers, there have always been alternative products.

Unfortunately, intel offers the best product by the most important metric: Performance per watt. Imagine that. Intel's R+D across their entire product lines focusing on performance per watt, paying off, surprisingly enough. There are alternatives, but if one product is the clear winner then consumers aren't oblivious to this. Generally speaking, the anything but intel fandom mindset seems to be limited to certain forums while corporate buyers will just get the better product. That product certainly isn't a "monopoly" by any stretch of the imagination.

 

[Mark R]

What performance does it get from the ability to run so many threads on 12 cores? Hyperthreading only sees ~20% gains.

It allows slightly better efficiency at using the various core resources.

Modern CPU cores are oversized for their typical workload, so that they don't bottleneck under unusual cases. E.g. 80% of the time, code will run fine with access to a single ALU unit on a core, but sometimes, the ALU might be a bottleneck. So to ease this bottleneck, a 2nd ALU might be added to the core, which most of the time is idle, but releases the bottleneck on certain code.

However, bigger cores are more expensive and more power hungry, so its a fine balance deciding how much to oversize.

Hyperthreading was designed as a "free" performance upgrade. The core is the same design, but a 2nd thread has access to its resources - so in the above example, the 2nd ALU might occasionally be able to do some work on the 2nd thread - giving a small overall performance boost.

These server cores are redesigned for massive threading - so the cores are huge, with loads of maybe 10 or 12 ALUs, etc. with access to 8 threads per core. With a lot of resources available and a lot of threads to choose from, it is possible to get more efficient scheduling of work.

The disadvantage, is that the cores are huge, expensive, hot and power hungry - so you lack fine-grained power control and the deep idle power consumption will be much higher, even if at full load, the chip is more efficient. Fine for servers running at high load 24/7, not so good for a desktop/laptop/mobile device.

 

[Headfoot]

I'd love to see someone use the open licensed Power8 plus this interconnect fabric to make another console competitor. Not that the current ones are bad or anything, but even more competition would be cool. I'm sure the licensing is egregiously expensive though

 

[cytg111]

I'd love to see someone use the open licensed Power8 plus this interconnect fabric to make another console competitor. Not that the current ones are bad or anything, but even more competition would be cool. I'm sure the licensing is egregiously expensive though

Did you see the size of that thing? Put on some graphics and then imagine the cooling solution to run it.. Console rivaly in my book would be something like a haswell i5, 290 and mantle .. clock it low enough for passive cooling and you'd be in business.

 

[norseamd]

Fine for servers running at high load 24/7, not so good for a desktop/laptop/mobile device.

you speak lies

just kidding

honestly with how few mainstream consumers buy desktops these days i think the chip would be fine for enthusiast desktop use

 

[Enigmoid]

you speak lies

just kidding

honestly with how few mainstream consumers buy desktops these days i think the chip would be fine for enthusiast desktop use

With a 600+ mm^2 die I don't think anyone is going to really be able to afford this monstrosity.

 

[rainy]

With a 600+ mm^2 die I don't think anyone is going to really be able to afford this monstrosity.

In 2010 Intel has released chip called Tukwila (Itanium family) with die size of 698 mm2.
http://en.wikipedia.org/wiki/Tukwila_%28processor%29

 

[Bubbleawsome]

Intel monopoly here we come! :awe:

 

[Idontcare]

Intel monopoly here we come! :awe:

Doubtful that well-monied companies such as Google, Apple, Qualcomm, Samsung, etc are going to sit idly by and passively allow their economic future to become dependent on Intel's executive team and BoD.

Intel isn't going to disappear anytime soon, but I'm not seeing too many "AMD 2.0" companies out there that are going to snatch defeat from the jaws of victory and lie down so Intel can dominate them in the coming decade.

Now then, if we see Apple do something screwy, for example hire Ruiz to replace Cook, then we can rightfully expect history to repeat itself.

For now though, Intel's competition in the mobile space owe their present size and market presence to their ability to innovate and compete without stumbling and falling on their own petard (for instance, no one at Samsung is saying they are delaying node development because they see no need to keep pushing to stay ahead of the competition, 404 no hubris found)...so Intel most certainly faces an uphill battle in every direction at this time IMO.

 

[DrMrLordX]

Huh, interesting read, that article. There are a few gems in there, but this one takes the cake for me:

In terms of performance per watt, though, the Xeon (~150W TDP) is probably just ahead of the Power8 — but in general, when you’re talking servers, power consumption generally plays second fiddle to performance density (how many gigaflops you can squeeze out of a single server).

That's true if you don't have to pay the power bills. The last time anyone seriously challenged Intel in the server space (AMD 2004-2005), they did it by winning performance/watt consistently.

Overall, the article is a nice read (well, page one anyway), but are people really that excited about escaping the Intel server paradigm? It wasn't so long ago that Intel was not dominant in the server sector. Server buyers were more than happy to get away from old-school "big iron" when x86 proved to be a desirable alternative. What we need is more credible competition to Intel in the perf/watt category, not . . . more "big iron".

Power 8 may be great for some HPC applications tilted towards that particular architecture, but for any wide variety of tasks, Power 8 seems like a waste versus Intel's offerings. If anything, I would think Intel has more to fear from ARM-based microservers. Power 8 isn't going to be the "ARM of the server room". ARM is going to be the ARM of the server room.

And is CAPI going to be that big of a deal? AMD has (technically) been able to do the same thing with HyperTransport for ages, and some of us wanted to see AMD leverage that fact to produce multi-purpose sockets/slots (HTX or Torrenza, take your pick) for drop-in coprocessors/cards to add all kinds of functionality to servers or desktops. That was supposed to be a key part of Fusion at some point, but . . . that never really seemed to happen? At least not in the desktop space anyway (I guess HTX slots did show up in the server realm, but I'll be darned if I know what they did with them. High-speed clustering interconnects, check. Anything else?). I think Intel wanted Xeon Phi to use QPI interconnects, but it's PCIe instead . . .

But also, I have to look at this and wonder exactly how this all fits in with IBM dumping their fabs. It's like . . . hey, here's Power 8! It's an open standard now, so you guys make the chips for us, and we'll sell software services to run on your hardware. Especially the complicated CAPI-based coprocessor-laden hardware.

 

[Headfoot]

Did you see the size of that thing? Put on some graphics and then imagine the cooling solution to run it.. Console rivaly in my book would be something like a haswell i5, 290 and mantle .. clock it low enough for passive cooling and you'd be in business.

The licenseability of the tech probably means the customer could choose an arbitrary core count, likely a lot smaller than the current fist-sized CPU pictured above. Given that each core is highly multithreaded they could probably get away with 6 cores instead of 12. Maybe even 4 if the hyperthreading scales well enough

 

[Idontcare]

But also, I have to look at this and wonder exactly how this all fits in with IBM dumping their fabs. It's like . . . hey, here's Power 8! It's an open standard now, so you guys make the chips for us, and we'll sell software services to run on your hardware. Especially the complicated CAPI-based coprocessor-laden hardware.

In many ways, and arguably in all the relevant ways, it is no different than SUN doing OpenSPARC in Dec 2005 followed by migrating from TI to TSMC for foundry services at 45nm and beyond.

We are witnessing the individual steps of the inevitable slowly coming to pass, more steps are to follow but the procession itself is transpiring as expected and pretty much right on schedule.

 

[mrmt]

In many ways, and arguably in all the relevant ways, it is no different than SUN doing OpenSPARC in Dec 2005 followed by migrating from TI to TSMC for foundry services at 45nm and beyond.

We are witnessing the individual steps of the inevitable slowly coming to pass, more steps are to follow but the procession itself is transpiring as expected and pretty much right on schedule.

Much like HP is slowly calibrating Itanium death, IBM and SUN should also be doing the same, gauging at how much money they will put on POWER and SPARC until the time to pull the plug finally comes.

But I have to concede, IBM really knows how to put a press release. They built the failed foundry club to group the losers of the business and yet some people think that this was the group of winners that would take over the foundry market for themselves. Now they are opening up POWER, which is a lousy try to get more money for to invest in R&D for power, because the mothership wouldn't risk its money on it, and even in this position of extreme weakness, we see things in the press like "pincer attack on Intel".

 

[thunng8]

A lot of misconceptions here. A few points:

1. Whether or not POWER8 will reverse the revenue declines is yet to be seen, but at least IBM is trying something different. The core is licensable. At the presentation Tyan demonstrated a white box POWER8 box which will be available by the end of the year and there are some other big name licensees like Nvidia, Google and Samsung.

2. POWER8 technology wise is impressive. IBM is promising about double the performance of the best Intel Xeon boxes per socket. The only benchmarks (a Java enterprise benchmark) I've seen released so far show this to be true:

Power8 (note there is an error in the result .. the s824 is a 2 socket machine):
[SIZE=+2]22,543.34
[/SIZE]http://www.spec.org/jEnterprise2010/results/res2014q2/jEnterprise2010-20140402-00050.html

Xeon:
11,259.88
http://www.spec.org/jEnterprise2010/results/res2013q3/jEnterprise2010-20130904-00046.html

More benchmarks will be released today and I'll certainly check if performance in other benchmarks bear this 2X rule.

 

[thunng8]

$8000 for their low-end 1P server? You can get 12 Intel cores for $2k.

Not sure how you can compare a complete system to a single component (CPU).

BTW, the E5-2697 v2 CPU only is $2,650

 

[thunng8]

That is one big CPU.

That's undoubtely the future MCM version of the POWER8 which crams 48 cores on a die (4 chips). A single chip is "only" 650mm^2. As a comparison Xeon IVY Bridge-EP (12 cores) is 540mm^2 .. so not a huge difference.

BTW, the POWER8 TDP is 190W according to OPENPower datasheets. If it is really double the performance of Xeon, then performance per watt is higher for POWER8 as Xeon is 130W.

 

[DrMrLordX]

In many ways, and arguably in all the relevant ways, it is no different than SUN doing OpenSPARC in Dec 2005 followed by migrating from TI to TSMC for foundry services at 45nm and beyond.

We are witnessing the individual steps of the inevitable slowly coming to pass, more steps are to follow but the procession itself is transpiring as expected and pretty much right on schedule.

Well, right. Look what happened to OpenSPARC. It didn't exactly set the world on fire (quite the opposite). There seem to be a few derivative chips, like the FeiTeng and the MCST-R1000. They don't seem to have made much of a difference in the global server market, though.

The article author doesn't seem to want to acknowledge that history may be repeating itself.

A lot of misconceptions here. A few points:

1. Whether or not POWER8 will reverse the revenue declines is yet to be seen, but at least IBM is trying something different. The core is licensable. At the presentation Tyan demonstrated a white box POWER8 box which will be available by the end of the year and there are some other big name licensees like Nvidia, Google and Samsung.

Okay, how is this different from what Sun/Oracle did? CAPI is obviously a difference, so there is that. What else? Is the Power8 such a fantastic chip that it will beat Intel's pants off?

2. POWER8 technology wise is impressive. IBM is promising about double the performance of the best Intel Xeon boxes per socket. The only benchmarks (a Java enterprise benchmark) I've seen released so far show this to be true:

That would be great (at least according to the article's assertion that power density is king), except that Power8 will top out at 2P configurations. Or at least, that's what the article says?

More benchmarks will be released today and I'll certainly check if performance in other benchmarks bear this 2X rule.

Should be interesting to see.

That's undoubtely the future MCM version of the POWER8 which crams 48 cores on a die (4 chips). A single chip is "only" 650mm^2. As a comparison Xeon IVY Bridge-EP (12 cores) is 540mm^2 .. so not a huge difference.

BTW, the POWER8 TDP is 190W according to OPENPower datasheets. If it is really double the performance of Xeon, then performance per watt is higher for POWER8 as Xeon is 130W.

Hmm. The article says:

We expect the Power8 will eventually be capable of clock speeds around 4.5GHz, with a TDP in the region of 250 watts. At this speed, the Power8 CPU will be around 60% faster than the Power7+ in single-threaded applications, and more than two times faster in multithreaded tasks. In certain cases, IBM says the Power8 is capable of analyzing Big Data workloads between 50 and 1,000 times faster than comparable x86 systems (the same amount of RAM, the same number of cores).

They "expect" Power8 will "eventually be capable" of 4.5 ghz. So, production silicon isn't running at those speeds or at that power draw yet. Nice of the author to avoid the specs of the first run of chips.

Makes me wonder about the basis of the author's conclusion that Intel's Xeon (presumably Ivy Bridge EX) has better average performance per watt vs. Power 8.

 

[thunng8]

Okay, how is this different from what Sun/Oracle did? CAPI is obviously a difference, so there is that. What else? Is the Power8 such a fantastic chip that it will beat Intel's pants off?

Not 100% sure but ORACLE/SUN cores were not licensable. The specs were so anyone can make a SPARC compatible core- but can anyone license the M6 core design and build their own CPUs with it?

The POWER8 is licensable ..i.e. another company can take the POWER8 core design and incorporate it into their unique CPU. It is like how the ARM Cortex A15 core is licensable and Samsung made the Exynos (can remember exact code name) with the ARM A15 core(s).

BTW, IBM has always almost been ahead of Intel in performance.. but that wasn't enough to lose market-share. They are now trying something different and making it licensable. i.e POWER7 was released in 2010. It took until late 2013 until the first Xeon cpu managed to edge past the POWER7 in performance (Xeon IVY Bridge EP and then EX). BTW .. the latest Ivy Bridge EX/EP only beats POWER7 since it has more cores per chip .. POWER7 still has the advantage per core on server workloads.

That would be great (at least according to the article's assertion that power density is king), except that Power8 will top out at 2P configurations. Or at least, that's what the article says?

IBM have only announced the scale out servers. Enterprise servers will be coming later this year and scale up to 32 sockets.

They "expect" Power8 will "eventually be capable" of 4.5 ghz. So, production silicon isn't running at those speeds or at that power draw yet. Nice of the author to avoid the specs of the first run of chips.

Makes me wonder about the basis of the author's conclusion that Intel's Xeon (presumably Ivy Bridge EX) has better average performance per watt vs. Power 8.

The openPOWER datasheet says 190W for the current models. (12 core 3.52Ghz or 8 core 4.1Ghz). Presumably the socket can take 250W though .. so maybe future version will be 250W but at higher frequencies.

 

[henriok]

The POWER8 processors in these low end configurations mounted on an organic substrates.

IBM have always used ceramic substrates for their high end single chip and multi chip variants and that probably won't change for POWER8. Ceramic substrates can handle heat better, but are more expensive to manufacture and they reside in systems that are more complicated that these entry level Linux machines so I think IBM will take a little more time before they can launch POWER8 products using ceramic substrates.

I think 190 W for a chip mounted on a organic substrate is pretty amazing in its own right. I'm very much looking forward to read any paper on the new packaging technologies that IBM have pulled out to make this work.

 

[mavere]

SAP benchmark results are pretty good.

Power8, 24 cores @ 3.52GHz: 115,870 SAPS

IvyBridge, 24 cores @ 2.7GHz: 51,650 SAPS

Also like thunng8 mentioned, I think the Power8's system specs should only list 2 processors, as the S824 only has two socket and the 12-core Power8 is clocked at 3.52GHz (the 6-core variant is clocked higher)