Google and Rackspace develop open server architecture on IBM Power9 hardware | Hacker News Reader

Google and Rackspace develop open server architecture on IBM Power9 hardware

cloudplatform.googleblog.com

50 pointsprotomyth12 comments

Superficially, I want to say "this is great news". BUT... I feel burned out on POWER. In fact, I'm kind of tired of hearing about it.

WAIT before you down-vote... I'm not saying it's bad, and if anything, it's just the opposite. I WANT a POWER based machine. And I've been patiently waiting for something like a decade (or more? Hell, I've forgotten by now) to be able to call up Tyan or Asus and order a $200.00 motherboard, order a POWER chip from NewEgg, slap the processor and some DIMM's onto the motherboard and jam it into an ATX case and go. But no...

OK, I understand that IBM are trying to appeal to the upper end of the market. It was never their choice to try and win over the hobbyists and experimenters and the low-end of the market. But now, with OpenPOWER we should, in theory, finally get some of that. But still, announcement after announcement and rumor after rumor appear, and I still can't order a cheap ATX motherboard, stick it in a case with a POWER chip and get to work. It frustrates the hell out of me to, as I think it's crucially important to have alternatives to the x86 hegemony.

(And yeah, I know you can order that Tyan server for $3,000 or whatever. If you can actually get one. But $3,000 is not exactly chump change).

I agree, and I bet if a 500 dollar, dual socket, openpower motherboard appeared along with a retail CPU or two, IBM might be surprised by the number of orders...

For what it's worth, I'm in the trading world and we evaluated Power8. It's cool. But reality is, it needs to become ubiquitous before it's all that useful.

You're looking at $6k+ for a 2 socket barebones system... preposterous. If you want FLOPS, you can add a few Titan X's to a dual socket E5 server and absolutely destroy the Power9. Want ECC? Throw in a single Tesla. Don't want to rewrite code for CUDA? Use an Intel Xeon Phi with the Intel C/C++ compiler. Just need more CPU power? Throw in top of the line Xeon E7's and they'll still outperform.

Now, the math on this would change if Power8 actually managed to become ubiquitous and competitive price-wise.

You are correct in the sense that the power platform is not the absolute best in most case, but i think it's the best compromise : Titan/Xeon phi may have more FLOPS, but that's only for a restricted domain of application. For a general purpose chip i think power8/power9 still give a good value ( power consumption not withstanding).

I am curious as to how you guys evaluated the platform, care to ping me ?

I'm with you on that. I recall back when there were affordable PPC systems really competing with Intel running alternative OS's. Fun stuff. I've even told people about each new POWER chip because they're often badass in some way. Yet, it's constantly out of reach for most of us because the numbers don't add up at these prices.

Further, I speculate this is harming it in the long-run because so much code out there comes from the scratching an itch phenomenon. People usually do that on whatever boxes they like working with. If basically none have POWER machines, then basically none of that code is going to be for POWER architecture. I don't have the data to test or prove this but I'm sure there's network effects there.

Just look at the explosion of code for ARM following the explosion of cheap devices. And Intel before it. And the major microcontrollers. And so on.

Just look at the explosion of code for ARM following the explosion of cheap devices. And Intel before it. And the major microcontrollers. And so on.

Exactly. This is why I never understood why IBM didn't do more to appeal to the experimenter / hobbyist / consumer market, even if their primary market was big enterprise accounts. Getting hardware out there for the tinkerers, OSS developers, etc., to seed the software space could only be beneficial to them.

I'm sure everyone involved in this OpenPOWER game is aware of the "attack of the killer micros". So the question is, why try again with the same recipe that already failed?

Getting something out there which is bang-for-buck competitive with Xeons is really the key here. But without massive volume, it's probably very hard to do. And the Intel division that makes Xeons apparently runs at 50% gross margins, so if they feel threatened the are certainly able to dump prices to make life for any competitor even more difficult.

But if they were serious about making a go at it, they'd, say, sell the motherboards and CPUs at cost and/or perhaps twice what Xeon systems go for. But everything I've heard before and in this discussion puts them at $3K per socket (I think for a barebones system), and that's a non-starter for something that isn't 10 times better than the competition (a common metric for what it takes to get attention when there's an entrenched competent competition).

Intel Xeon funding and lots of ecosystem it uses came from other CPU's that cost less. So, I'd say POWER needs such benefits that come from lower-end if Xeon needed it to get where it is. POWER actually needs more since it's way the heck behind Xeon in balance of price, performance, and capabilities in SW.

Right behind you on this.

Industry as a whole needs to come together to support this. OCP should add open CPU architectures as well into the mix.

OCP is great but there is no supply chain to support it at a reasonable cost.

I think this was to coincide with the Open Power summit going on right now, there a link of the hardware announced at the top of this:

http://openpowerfoundation.org/openpower-summit-2016/

Wow, thanks for sharing. Now this is encouraging. Looks like there are several servers available (well, announced), including ones from Penguin Computing and Mark III.

Maybe this does bode well for the eventual availability of affordable POWER based kit in the not-too-distant future.

Reply on news.ycombinator.com