I use POWER8 hardware at Red Hat (supplied by IBM[1]) and it certainly does rock. It helps that I don't have to pay the electricity bills.
I just wish the hardware was more available. There is precisely one, barely affordable, non-IBM POWER8 system available to buy[2] (if you have $3-4k burning a hole in your pocket). Where are the development boards?
What are some of the easy benefits of using the POWER series of processors/architecture?
My company runs an IBM POWER7 series system as our ERP/WMS, but that's really only because our vendor sold it as part of a package.
From a user standpoint, POWER architecture just means it will cost me more (on all fronts, from acquisition, to maintenance, to power consumption, etc...). Come upgrade time, it makes it very unlikely I'll be able to unload this system on someone else and recoup anything but a few hundred bucks, sold to someone as a testbench.
When compared to a run-of-the-mill x86 system, POWER doesn't seem to have any real tangible benefits.
I agree with all your points. The benefit is per-core performance is very very fast indeed. If you need that, and price is no object, then you should use POWER.
That's the problem: price and availability. It only challenges Xeon in raw numbers, but $/performance matters a lot especially at scale. If I were building a data center I'd choose more slightly slower CPUs if the total power per $ were higher. Energy use maters a lot too.
The only niche I can see for a platform that offers maximum performance per-machine at a huge price premium is hard-to-parallelize and legacy applications that need reliable maximum performance in a single box. That's not really that big of a niche these days.
If IBM wants to launch this they're going to have to take a loss to prime the pump: sell the chip at cost or even slightly below cost to get it out there into the market and get people using it. Otherwise it's going nowhere.
You can definitely assume that :-) The IBM kit linked above is on loan from IBM and is publicly accessible - you just have to fill in the form for an account. It runs Linux (Fedora or RHEL IIRC).
I remember, a long time ago, the installation charge for MVS hitting $1 million, for the biggest IBM 3090 mainframes. I thought those days had gone....
That's unfortunate. RunAbove was very nice and gave us one last year for Haskell.org (so we could port to PPC64le), and our instance is still running well (176 cores/48GB of RAM or so), but I can confirm I can't spin up any new instances.
That's a shame, they're excellent machines... I guess we'll have to find another sponsor soon.
I think it is telling that Tyan is the manufacturer for these boards. My experience with them is that they're definitely a runner-up company compared with Supermicro, let alone the the big OEMs.
Benchmarks sound like their fighting hardware frequency throttling more then actually benchmarking the chips.
One big issue to point out:
>Floating Point: NAMD
GCC sucks at automatic hardware vectorization. So does the LLVM. Really the only time you can count on getting automatic hardware vectorization is if you shell out for the ICC AND write your code in Fortran. I'm gonna bet they didn't vectorized a goddamn thing, but we can't inspect anand's binary so we'll never know. The results are still likely correct, but IBM should have lost by a smaller margin.
TL;DR
POWER8 is fun but costs 5k more then Xeon per rack mount and uses about 2x the power usage for 10% less performance on generalized work loads. But can pull off 10-15% more performance on _some specialized_ workloads. So meh?
My experience has been that icc (the C compiler) does a very good job of auto-vectorization. xlc does as well, provided you avoid the horribly misdesigned POWER6 architecture.
I agree that gcc sucks for this; we didn't have LLVM last time I was doing this kind of work, so I can't comment on that.
If you want to really use SIMD units (on x86) check out ISPC. I recommend writing small functions in it that work over large chunks of memory so that SIMD can run at full speed with good cache locality.
LLVM is about on par with the GCC. You just keep moving variables around it decides the moons of Jupiter are properly positioned to vectorize your code.
System intrinsics are always your best bet, just really annoying.
We ended up doing some hand-rolled vectorization that way on the POWER7, because SIMD only really worked well if the data was aligned with the SIMD register width. Intel fixed that problem with a relatively low-cost unaligned load instruction.
It's ok guys. It's just our high-power server CPU architecture. Totally cool to slap a yellow star sticker with misaligned text on it. I'm not sure if that's comic sans or a knockoff. Also I just love the teal bevel and big bank bag of cash. And the stock icons for checkmark, arrow, and question mark. And the complete and total misalignment of all the text.
The design language used in those slides has been iteratively refined over decades and exhaustively A/B tested. It has many years of psychometric analysis applied to make it one of the most powerful sales tools in the world.
To win over enterprise PHBs, you need to make them think that you are one of them. That you understand and empathise with them deeply. That you speak their language, and that you are part of their culture.
What better way to do that than to give them a powerpoint presentation that looks just like the ones they give each other. That shows that you are already part of their world, and that you are worthy of the key to the executive washroom.
I'd be a lot more excited about this if they were more accessible. I can barely justify an upgrade to whitebox Intel Xeon servers and there's just no way I could rationalize the prices that IBM demands.
> POWER, OS/2, and The Thinkpad has convinced me that IBM just doesn't make the most of what they create.
They are all kind of different.
I used to use Thinkpads and I hated IBM for selling it. But from their point of view it was the absolute right thing to do. High end consumer laptops are dominated by Apple (heck IBM is buying Apple and giving them to all their workers, something like 300K Mac Book Pro's). Everything else is racing to the bottom with thinner and thinner margins. Windows running laptops have to compete with Asus and friends and the money just wasn't there.
Not sure about OS/2 much, don't remember the history. But with POWER, IBM has kind of started to turn around in the last few years.
At some point in the past they have made an explicit choice to not play in the consumer market. Heck, there used to be IBM stores, you'd walk in and buy IBM products like you go to an Apple store now. But they decided they don't want to play in that market (or better or for worse). We'll still see how it ends up working out.
So far it still stays in business after hundreds of years, maybe it just luck or there is something to its business approach. (Fun fact, it used to sell cheese slicers and time tracking devices as well at some point).
OS/2: decent OS, ahead of Windows for awhile, but IBM was never going to go to the mattresses enough to get it the install base it would need to compete with Windows. IBM's dysfunctional PC division didn't help. It would have been nice if IBM kept OS/2 up as a specialty operating system, but I think the overhead costs were too much and Gerstner was, by 1996-1997, very much in a kill-anything-that-isn't profitable mode.
Prior to Gerstner, IBM had a variety of esoteric products which were solely designed as loss leaders, never earned a profit, and relied on subsidies by other parts of the company to stay alive. Post 1993, really starting in 1994-1995 these got killed off or sold off, rapidly.
It wasn't enough to break even, one number I recall being thrown around was that we had to get to a 12% Expense-to-Revenue ratio, ignoring SG&A which was a corporate-wide number. Growth products, products in new markets were exempted entirely or given better targets, but old-line products were held to this magical 12% ratio (I was in the mainframe division at the time, which was grotesquely profitable and even today subsidizes much of the rest of IBM).
I don't think the physical stores lasted long, they eventually were sold off to ComputerLand, possibly as part of a consent decree (before my time). A lot of bad business decisions at IBM start with the various consent decrees it operated under as a result of various antitrust cases, and the utter fear of yet another antitrust case developing.
Yeah it didn't seem like they were really big and would have lasted, but it was kind of a contrast with today where a young person on the street would probably have not idea what IBM does.
IBM was big, really big. At one time, it had more than 70% of the computer business, so it was twice as big as every other computer company added together.
In fact, IBM was really big before computers even arrived. It dominated data processing based on punch cards and was sued for monopoly abuse in the 1930s. It was the original Evil Empire.
Microsoft -- founded in 1975 -- has been spectacularly successful, and IBM's performance has been mediocre for the past 15 years. Even so, Microsoft has only just overtaken IBM in revenues. (Both are now around $93 billion. On any reasonable growth path, IBM would be well over $200 billion.)
Indeed, Microsoft would probably be nowhere without IBM. It was IBM that set the PC standard with the IBM Personal Computer in 1981, and Microsoft was lucky to be part of it. In IBM terms, Microsoft stole a small part of IBM's rightful monopoly. IBM's response was to try to kill it with OS/2 EE and the MCA bus in PS/2 computers, as part of SAA.
IBM got out of selling ThinkPads long before Apple had any kind of dominance in the high-end laptop market.
Arguably they got out of it because there was no such thing as high-end consumer PC hardware, the margins were low, competition fierce, and they didn't want to play in the low-margin discount market.
Arguably the ThinkPad could have been a success as a high end business laptop. I think under Lenovo's stewardship it has languished and become a typical cheap Asian-manufacturer product of 500 SKUs, all slightly different, with quality on a downward curve.
On the other hand the fate of RIM/Blackberry shows there isn't much of a market for a 'business' device that is separate from a high-end 'consumer' device. The latter wins over the former.
The thing is they didn't at all continue on their "business device" strategy, they got suckered into trying to match Apple - they tried going touchscreen crazy and it cost them their core market.
In hindsight, not many people will understand this, but their greatest blunder was trying to charge / keep BES as a revenue product.
They should've given it away far and wide, made it open, encouraged people to customize it.
Imagine if there was a robust, easy to use, free version of BES available? How many business would be using it (>50%)
That would've kept RIM in the game for far longer, but they were short sighted and stupid.
I've wondered something: MIPS and ARM have both made inroads in the low-power market, supposedly because RISC supports power-efficiency (x86 microcodes notwithstanding), but POWER has always been described as less power-efficient than x86 or other architectures. What makes POWER need so much... uh, power?
My uninformed take on the issue is that IBM's fabrication technology was never able to deliver on the performance vs. power projections so they traded off power in order to maintain the performance, and that is why POWER takes so much power.
IBM went down the Silicon On Insulator (SOI) path and bet the farm that they could beat Intel's fab technology with that. IBM promised SOI would allow them to make high speed, low power chips. They were unable to deliver on the promise and IBM lost to Intel, along with the rest of the "IBM Alliance" (AMD, Freescale, Chartered Semi, and others).
With respect to the IBM Alliance, it was not kind to any of the members. Many of the members ended up rolled up into Global Foundaries. Global Foundaries was the AMD foundary that was spun out of AMD, Chartered Semi was bought by Global Foundaries, and IBM ended up paying Global Foundaries $1.5 billion (plus took a pretax charge of $4.7 billion) to take their fab plants. Freescale lost the Apple CPU socket because of power and performance problems and has struggled.
At the high-end, the power-costs of ISA are negligible. Even at laptop levels of power consumption it's barely noticeable. Once an Intel chip has decoded the instructions to uOps, it is mostly on an even playing-field with RISC chips.
Also, to be pedantic, POWER is very much not RISC; the only thing it really has in common with RISC is fixed-length instruction encoding (though simpler instruction decoding is an advantage at really low powers).
>At the high-end, the power-costs of ISA are negligible. Even at laptop levels of power consumption it's barely noticeable. Once an Intel chip has decoded the instructions to uOps, it is mostly on an even playing-field with RISC chips.
This doesn't actually answer my question at all, it's just a canned response to something I alluded to in my question (CISC vs. RISC) and a debate which I don't fully understand and would rather avoid. The question was why does POWER consume so much power, not why doesn't POWER being RISC matter? The only reason I compared it to MIPS and ARM was that I thought it was similar to those architectures and they're considered to be low-power.
The impetus is that POWER is actually worse than other architectures, including x86, for unexplained reasons, w.r.t. power consumption.
POWER consumes more power because IBM has not optimized for that to the degree that Intel has. The PA6T was extremely power efficient (and was designed by a team headed by the same person who designed StrongARM (became X-scale) and SiByte's (now purchaesed by Broadcom) MIPS chips).
[edit]
To clarify, the difference between ISAs for power efficiency is negligible at all but the lowest powers, which is what I was trying to say in my original response. The Pentium 4 was far less power efficient than the contemporary G4 PowerPC, as a counterexample. If the power difference between two chips is more than 1 watt per instruction decode unit, you can probably safely assume it's not due to the ISA.
IBM's power consumption is probably mostly because they've never put the processors into energy sensitive applications. More importantly, they are sold because of some benchmarks and extreme reliability.
It's a superscalar CPU running Tomasulo’s dynamic scheduling algorithm to extract instruction level parallelism. Once you are beyond two-issue in-order processors, the CISC vs. RISC distinction becomes a minor point.
Note that most ARM processors are two-issue in-order, so RISC is still relevant at the low end. Compare power consumption of ARM vs. equivalent Atom processor. Intel has a process advantage with Atom, but ARM is still lower power.
Intel had a huge process advantage over Freescale when Apple ditched PowerPC. PA-semi proved it was possible to make power-efficient cores (i.e. the PA6T, which everyone assumed Apple was going to use in their next line before the Intel transition was announced).
also as a completely unrelated side-note, the number of processor lines started by Dobberpuhl is crazy: Dec Alpha StrongARM (Became X-Scale), what is now the Broadcom MIPS SoCs. Apple picked him up via an acqui-hire, but he retired the following year.
Why are they using MIPS as a metric for comparing integer CPU performance? When you're looking at two different compiler backends targeting two (very) different ISAs, it seems pretty thoroughly uninformative.
In benchmarks MIPS usually means "we arbitrarily say that reference machine X runs this benchmark at Y MIPS". Then, for a different machine Z, Z_MIPS = Y * (X_TIME / Z_TIME).
For example, the 7zip benchmark used in this article is "normalized with results of Intel Core 2 CPU" [1], or the famous Dhrystone integer benchmark's "VAX MIPS" [2].
Still I agree it's weird terminology. It would make more sense to use the true units of the benchmark, or just say "performance relative to reference machine".
Yep, with you 100%. Anandtech (especially Anand himself) usually does pretty rigorous benchmarks but this is either a poorly executed comparison or intentionally biased towards POWER8 (and I say this as POWER fan and former IBMer). I'm really let down by this review.
Another HNer went into detail as to why the comparison is flawed but basically when you compile with a compiler made by the chip designer, especially for numerical computation, the performance is almost definitely going to slant towards that direction. GCC and Clang have to focus on dozens and dozens of backends. Anandtech has more than enough money to spend 1k for an ICC license.
POWER/PowerPC is the same architecture that Apple once used in the Macintosh, so if you want to play with Linux on a PowerPC machine they can be had for almost nothing on the used market.
There are really only two models worth having, at opposite ends of the spectrum: The PowerMac G5 (64-bit, based on POWER4) was the fastest available but they're enormous and power hungry. At the other end, the G4 mini is much smaller and more efficient but 32-bit and not as fast.
32 bit is not worth having, nothing modern supports it.
The G5 was supported with Fedora until recently but alas support was dropped. Also the newer Power distros are the new ppc64le flavour, which is a slightly different ABI and little endian but the G5 is not alas dual endian so does not support it. But they are cheap and quiet and mine has been running reliably, although I might install FreeBSD on it next as that has ongoing support.
I have a 2.3GHz dual core PowerMac11,2 running FreeBSD 10.2 PPC64 on a ZFS root. My machine has 8GB of RAM and works brilliantly.
I tried all flavours of PPC linux and none of them worked particularly well. FreeBSD 10.x had the advantage of 64 bit support and the nvidia 6600LE works acceptably with the open source drivers.
I highly recommend FreeBSD 10.x PPC64 to anyone with a G5 PowerMac who wants to run a modern OS.
Ubuntu Mate has a PPC version which seems to be pretty popular.
Unfortunately the G5 power consumption and fan noise (unless you have the liquid-cooled version, maybe) are high, so trying to reanimate a G5 is not really a compelling proposition for me (besides the fact that I ended up stuck somewhere in openfirmware-land last time I tried).
This is true, but honestly no one is really interested in playing with the POWER ISA, which for most purposes is Just Another 90's RISC Box. The thing drawing clicks here is the question of whether or not the specific POWER8 architecture can beat contemporary Intel (which means Haswell right now, as Broadwell doesn't have a high performance Xeon variant) architecture in some regimes. And... it kinda doesn't.
I think that apple figured out a long time ago that the price performance would never be as good, and that pure performance would be on par - at best.
Since then it has all been about making it look good for enterprise goons. Consolidating and vertical scaling and all that.
And poor buggers that end up with high cpu steal and poor performance because a new box is just unimaginably expensive when you have filled the existing system.
I got a PowerMac G5 recently from a friend who no longer needed it, because I always loved those big beasts and have fond memories of the time I got a prerelease model so I could port a codegenerator backend to it. Alas, I haven't found any practical application for the machine this time around, but I'm reluctant to pass it along as I'm afraid it'll just be scrapped.
I don't think IBM wants customers that would buy Power 8 were it not for the price premium over roughly equivalent Xeons. The customers they want will be writing much, much larger checks for services and support.
I just wish the hardware was more available. There is precisely one, barely affordable, non-IBM POWER8 system available to buy[2] (if you have $3-4k burning a hole in your pocket). Where are the development boards?
[1] http://research.redhat.com/powerlinux-openpower-development-...
[2] http://www.tyan.com/campaign/openpower/