SGI -- formerly the titanic Silicon Graphics company -- has released a "personal supercomputer" that can handle up to 80 cores and up to a terabyte of RAM. I used to do work for an SGI VAR and we had a running joke about the merged SGI-Cray unit shipping a water-cooled laptop. This isn't that far off.
SGI Unveils Octane III Personal Supercomputer (via The Inquirer)Octane III is office-ready with a pedestal, one-by-two-foot form factor, whisper-quiet operations, easy-to-use features, low maintenance requirements and support for standard office power outlets. While a typical workstation has only eight cores and moderate memory capacity, the superior design of the Octane III permits up to 80 high-performance cores and nearly 1TB of memory for unparalleled performance...
Octane III is easily configurable with single- and dual-socket node choices, and offers a wide selection of performance, storage, graphics, GP-GPU and integrated networking options. Yielding the same leading power efficiencies inherent in all SGI Eco-Logical compute designs, Octane III supports the latest Intel processors to capitalize on greater levels of performance, flexibility and scalability.
RSS Feed
Oh boy! I want one! Now, all they need is an OS X license. Hey, Apple, are you listening?
"I used to do work for an SGI VAR"
Um, still on a retainer from the marketing department by any chance?
Ooh... LOVE to see how badly Vista would screw this machine up...
Though I bet even on this Photoshop would still bog.
I'd have to disagree with the claim "unparalleled performance...". With 80 cores that would be extremely parallel.
I'm sure you think you're joking, but I take that kind of thing deadly serious. You're either being deliberately or accidentally insulting, but you're sure as anything being insulting. If you seriously think that I take bribes to post material here, you can take your readership elsewhere.
Impressive engineering, getting 20 quad core Xeons into something that would sit by the desk and not howl like a chorus of the damned. The atom configuration seems kind of weird. The integration is still nice; but that sort of density with atoms is less dramatic, and atoms are really pitiful in terms of power.
It is sad; but not unexpected, that this is just a PC cluster, albeit a nicely engineered one, not one of old-school SGI's cool tightly interconnected setups.
VMWare ESXi would be a nice addition to the OS list, too.
If the price is "ask a sales rep", the word "personal" probably doesn't belong in the product's name.
There doesn't seem to be much detail on how this works inside. Is it a real multiprocessor computer, or is it actually a bunch of dual-core machines networked somehow? How are they interconnected? And how can this possibly be "whisper quiet" with a maximum power of 4kw? This thing could heat a large office when it's under load.
And Cory, I'm impressed that you're still bothered by trolls.
By "dual core" I meant "dual processor". Also it seems like #4 answered my question while I was still typing it.
That seems pretty flat-out awesome to me. I would have to dream of ways to use it, but it's fun to dream...
Personal weather modelling
CFD in 10 minutes vs. overnight
Finite element analysis in a moment instead of an afternoon.
All that horsepower on the desk is kind of captivating...
...and soon it will be dwarfed by what our phones will be able to do...
Ah desktop supercomputers that's the kind of news that makes you want to shout "WE LIVE IN THE FUTURE!"
Can't imagine how much electricity this thing sucks up and how much heat it puts out.
It is cool, but the word "supercomputer" is meaningless. Even the piece of junk you and I have at work is many times more powerful than the first "supercomputer."
But will it run Crysis 2?
Actually, I can think of much more interesting uses for this, but someone had to say it.
I agree with lobster. The real reason that this is noteworthy is that NOW you can put on your desk the equivalent of a bleeding edge supercomputer from when ~30 year olds were in college, instead of from when me or my father were in college.
That is a thing of beauty! - People can bitch and moan about how much it costs & how you could build one yourself for $x but I've always been impressed with SGI's designs... HOT!
There are three versions - "1-way" 19x Atom dual-core CPUs, "2-way" 20 Xeon quad-core CPUs, or the "Graphics Workstation" 2x Xeon quad-core CPUs.
The power is listed as up to 4x 1000 Watt PSUs, redundant power optional. All configurations are fed by dual 15 Amp mains connections
The two-way configuration (based on Intel Xeon CPUs) supports VMWare ESX 4.0, so ESXi is likely supported.
The Graphics Workstation configuration sports a tremendous number of PCI Express slots.
Here is an online overview, and here is a PDF datasheet.
Some time ago when I worked for a fabless semiconductor company, we had a few Alpha-based servers that were used for silicon simulation tools.
I'm pretty sure that's the sort of thing this is aimed at -- supercomputer problems on a group/personal level (vs. a university or government).
That said, I'm already looking forward 5-8 years when you'll be able to find them for a hundred bucks.
This is really cool, but there are a few flaws with it - or at least with their marketing.
The always popular term "Desktop Supercomputer" is meaningless because it's always going to be far behind the current state of the art. Otherwise, actual supercomputers would be desktop size, but then if they needed more computing power than would fit in a desktop, what would they do? Make it bigger! The current top system has 129,600 cores - do they really think 80 cores is going to rate "super" in comparison? (Especially when their own top system has 150,152 [mildly slower] cores?)
But really they're playing a completely different game, with a system that starts at $5000. How much do you think BlueGene cost?!
The other flaw is the "whisper quiet" claim. I'm calling BS on that until they release decibel readings when full of 80 Xeon cores running full tilt. I see some pretty small fans - those are almost never quiet.
But on the positive side, this is a really cool system, basically a blade server with the integration taken care of for you. Hopefully SGI sells a truck load of them.
The press release says the price is 'from' $8k.
Alas, it seems like SGI's fortunes are in decline. Their products aren't as exciting as they used to be.
The "personal" refers to more the form factor than affordability of the unit.
That said, I'll take two please. I wonder how long brute force security attacks would take with these.
So is that like a render farm? Or are those blades actually in a cluster or something like that?
I imagine this would sit under the desk of a designer at a big studio so they don't have to worry about the farm taking a hit when everyone renders at once. I can also see this being used for smaller shops that need the power but don't have a lot of space.
I do wonder, with gpus being able to handle more and more tasks, is a cpu-only rendering solution the best?
You can find more details, and questions about Why is SGI even trying this, over at VizWorld.com and InsideHPC.com:
* http://www.vizworld.com/2009/09/sgi-unveils-octane-iii-personal-supercomputer/
* http://insidehpc.com/2009/09/21/sgi-launches-new-deskside-octane-3/
The details of this machine are puzzling. Build it with Atoms?? If you get a GPU installed, it drops to 2 processors?? It's just a strange beast all around.
Seeing as there's 120 jobs in queue for the supercomputer I'm using at the moment - I could really use one of those!
I'm having an orgasm looking at this thing....I'm with mattf, if it runs osx, or any UNIX at all...Baby, I'm there!
#18 - GPUs have a very limited use at the moment. Assuming you have custom tools they are still limited to the types of things you can do with them, such as particle acceleration and other effect.
There's Gelato that renders through the GPU but it's in the early stages of development, so it's not really that useful. Best GPU use right now is in Video Editing. HD is quite demanding and there are a lot of encoding tools that will make use of the GPU.
Other then that, I don't think there's much being developed for GPU rendering.
@21: It comes with SuSE or RedHat, so ...
They did say "GP-GPU" which indicates that it should support GPU renders once the software is there.
SGI died last April as it went into bankruptcy and its assets were purchased by Rackable Systems, which assumed the new name and took on at least some of the SGI engineers and products. This looks like a Rackable product, though.
http://en.wikipedia.org/wiki/OpenCL
http://developer.apple.com/mac/snowleopard/opencl.html
But, but, but ... it doesn't have a couch! How dare you call this a supercomputer?
Well, yeah. That's what's holding it up mainly. Not much software supports it and many industry leaders believe it to be a dead end as far as 3d rendering goes. Have to wait and see.
[sniff] An SGI (er, sgi) and it's just a box, not even purple or anything... [sniff] so boxy...:( I worked on SGIs from 1994 through ~2001, Indigo > Indy > Indigo2 Extreme > Indigo2 Impact > O2. I loved that they were never just plain boxes, never "normal" colors. I do miss those cool machines, but not paying >10x more than "regular" hardware for 4x the performance.
Ok someone mentioned using OS X with this box... This is exactly the persion that SHOULDN'T be using this box.
Any flavor of *NIX instead of that heretic piece of garbage.
#22 Just because gpgpu hasn't matured doesn't mean it isn't relevant. Projects like gelato don't match the quality of the big boys, but it demonstrates the potential. 3d rendering can be threaded, which is why render farms have many nodes and cores, like the sgi box above. Once opencl is implemented in the big software renderers I expect the gpu to be just as important as the cpu.
That would complicated the form factor. Maybe they can stack them like this:
..hmm looks like my stack of "less than" symbols got removed.
The only problem I see with this is that the blades are probably custom plugin PCBs specific to this computer, and there most likely is limited uprgadability to this (as opposed to swapping a motherboard). So, you are tied to SGI for any support whatsoever untiol EOL. But for specific business applications, it might be the way to go.
I could use that. Maybe I wouldn't crash Maya as often.
Nah, Maya will crash just as consistently.
You can get adapters that allow you to mount video cards flat, so it might get a U taller.
The GPU rendering has some technical hurdles to get over. First of all, it needs to be able to render with MentalRay and some the other rendering solutions, because as seen with other rendering solutions. Without that support you'd have a hard time getting anyone to adopt the technology, at least at the smaller scale (I'm sure ILM will write some crazy code for themselves, but that doesn't help anyone other then ILM).
Aside from that it would mean that your render farm will be 30% more expensive, so you'd have to make sure it's worth the cost. You might get the same result from just buying two or three more blades.
Trent,
what is the hurdle to Mental Ray using OpenCL?
No challenge, genuinely asking.
@Yeraze: the point of using Atoms would be to get the nodes in the machine to fit within a certain power profile -- either so that you can plug it into a standard wall outlet, or so that you can cool it (so it doesn't let out the magic smoke).
If you have GPU-friendly jobs, then you don't need such fast CPUs, because you'll expect that the GPUs can carry most of the load. High-end GPUs take a lot to cool (hence the ginormous fan on the GPU card). Alternately, if your jobs run better on the CPU, then it's better to spend your cooling budget on hotter CPUs.
Projections for future machines are that for many processors, it's better to have slower but lower-power processors. See e.g., the recent "Exascale Report":
http://users.ece.gatech.edu/mrichard/ExascaleComputingStudyReports/ECS_reports.htm
@hilbertastronaut: Yeah, I guess getting it to run on 110 would be the main thing.. Most "full" clusters would take 220.
But if you look, you can't get it with Atoms & GPU's.. it's GPU's+Xeons, or Atoms.. no mix..
@Yeraze I made the typical mistake of not reading the actual article ;-)
Just Atoms and no GPUs might be good for an old-school MPI code that's communication-bound. I imagine there's a lot of code out there that doesn't scale well past a few dozen processors, but is still useful. That's the perfect application for the "desktop supercomputer."
btw I've seen a few different systems out there that could be called "desktop supers." Cray makes one (it runs Windows even) and I saw at least one other brand when I was at Supercomputing '08 in Austin. NVidia makes a "deskside" plugin box with GPUs inside.
@hilbertastronaut: Yeah, I've seen the Cray system.. In fact the original ads even made it on Photoshop Disasters :)
The NVidia one is the QuadroPlex, 4 Quadro's in a single External box, that consumes 1 PCI2.0 x16 lane in the box.
Unfortunately, I don't think any of this SGI builds support a quadroplex, except for the "Graphics" version which kinda defeats the point.
But yeah, the 80-core build would be great for people doing raw computational MPI stuff.. I just think it's a small market to be targeting so specifically. People who a) Need 10 computers for parallel computational.. b) Can't afford a rack of 10 computers.. c) Can afford this..
See....
I'd love to have something like that if only my ESRI software(Desktop version) would run on windows server ..%$@#$
I'd take my quarter trillion point Lidar data and render a couple of mountain ranges with Millimetre accuracy.
I am more impressed by the Ikea cabinet rendering farm: http://helmer.sfe.se/
Imagine a beowolf cluster of these!
Wait, this isn't Slashdot?
@Mindpowered
"ESRI software(Desktop version)"
*shudder*
What a waste of perfectly good CPU cycles
@Yeraze Say you don't have a machine room, but you need a modest cluster. Machine room infrastructure is expensive and permanent; workstation-sized boxes are portable.
Anyway, I'm not buying one ;-)
Sometimes I like to pretend I'm living 20 years in the past. That way every system I own is a supercomputer.
Sheesh, Cory. That was mildly insulting, but hardly comment-worthy. Now, if someone said, "I hate you and your writing sucks," well, okay, that wouldn't be comment-worthy, either, but I'd expect an IP ban out of that one ;-)
Having said that, I wonder what the demand for this machine is? It might be fun to push the thing to its limit ;-)
I got to use SGI's about 10 years ago to learn video editing/composting.. i remember for the time how much faster they were than the Pc's/macs that were also available..
i'd kill for this machine.. even as a sweet (& totally overpowered) design platform (but, could u run osX on it?)
$8k. Still less than a fully loaded Mac Pro!
nVidia Tesla technology is already there for less. Velocity Micro, for example, will sell you a 240 core machine for around $10K. These and other desktop supercomputers built around GPUs from Colfax, Microway, AMAX, HP and others (and running up to 960 cores) are currently being used for climate modeling. You wait 20 minutes instead of a few seconds on the big IBM, but you also don't have to stand in line.
I don't know enough about how this works but I thought you needed software that was designed to use a certain number of cores in order to get increased performance. In other words, would the software have to be custom-tailored to take advantage of this system?
@#53 POSTED BY ANONYMOUS
I thought the same... but I have a feeling this might not make a workable hackintosh.. maybe with xserve though? I dunno...
#5 Cory:
I think my comment was prompted by the slightly anachronistic subject matter; it just seemed a little out of place to post a feature on so specific a product that seemed out of context with the usual Boingboing fayre.
With hindsight (and reading down into the other posts after the piece) it's clearly something everyone is interested in, so I'm sorry.
I like your work and I'm slightly mortified to have caused you offence.
I'm not a troll and I don't want to take my reading elsewhere.
But.. But how is it COOLED???? ;)
And um, okay I want one. Even, yes even if the memory's not shared.
But sorry it is just not snazzy enough to be called an Octane, what's with that Rackspace?
Matt R.
The fact that 'desktop supercomputer' is inaccurate doesn't mean it is bad marketing. It got your attention, didn't it?