Slashdot Mirror
MIT Artificial Vision Researchers Assemble 16-GPU Machine
lindik writes "As part of their research efforts aimed at building real-time human-level artificial vision systems inspired by the brain, MIT graduate student Nicolas Pinto and principal investigators David Cox (Rowland Institute at Harvard) and James DiCarlo (McGovern Institute for Brain Research at MIT) recently assembled an impressive 16-GPU 'monster' composed of 8x9800gx2s donated by NVIDIA. The high-throughput method they promote can also use other ubiquitous technologies like IBM's Cell Broadband Engine processor (included in Sony's Playstation 3) or Amazon's Elastic Cloud Computing services. Interestingly, the team is also involved in the PetaVision project on the Roadrunner, the world's fastest supercomputer."
"But can it run Crysis?"
*Ducks*
AMD/ATI have released the specs for their hardware. Why haven't the proprietary NVIDIA engineers done the same? What do they have to hide?
The day when self modification/upgrade enthusiasts start overclocking themselves and bragging about how many fps their eyes get watching the superbowl.
I noticed they didn't have 8GB of RAM though. Very sad.
When gamers grow up and go to college.. blue leds and bling in the server room!
One more step to the last invention man ever need make... hooker bot. (mine would be a Buffy Bot, but that's just personal preference)
If you can read this... 01110101 01110010 00100000 01100001 00100000 01100111 01100101 01100101 01101011
...throw more computing power at it.
I miss '60s-'70s AI research.
Oh, and get off my lawn.
I keep seeing all these articles about bringing more types of processing applications to the gpu, since it handles floating point math and parallel problems better. I only have a rudimentary understanding of programming compared to most people on this site, so the following may sound like a dumb question. But how do you determine what types of problems will perform well (or are even possible to be solved) through the use of GPUs, and just how "general purpose" can you get on such specialized hardware?
Thanks in advance.
All of the hardware (save the case) is stock. It's just two machines (with 4 video cards each) in one physical case.
I'm still eager to see PhysX running on my dual 8800M GTX laptop. I've run all the drivers from 177.35 up and I'm running the 8.06.12 PhysX drivers as required. :(
Apparently it's just the mobile versions
"We know what happens to people who stay in the middle of the road. They get run over." - Aneurin Bevan
I think this part of the computing timeline is going to be
one that is well remembered. I know I find it fascinating.
This is a classic moment when tech takes the branch that
was unexpected. GPGPU computing will soon
reach ubiquity but for right now it's the fledgling that is being
grown in the wild.
Of course I'm not earmarking this one particular project
as the start point but this year has gotten 'GPU this' and
'GPGPU that' start up events all over it. Some even said
in 2007, that it would be a buzzword in 08.
And of course there's nothing like new tech to bring out
a naysayer.
Folding@home released their second generation
GPU client in April 08. While retiring the GPU1 core in
June of this year.
I know I enjoy throwing spare GPU cycles to a distributed
cause and whenever I catch sight of the icon for the GPU
client it brings the back the nostalgia of distributed clients
of the past. [Near the bottom].
I think I was with United Devices the longest.
And the Grid.
Now we are getting a chance to see GPU supercomputing
installations from IBM and this one from MIT.
Soon those will be littering the Top 500 list.
I also look forward most to the peaceful endeavors the new
processing power will be used for... weather analysis,
drug creation, and disease studies.
Oh yes, I realize places like the infamous Sandia will be using
the GPU to rev up atom splitting. But maybe if they keep their
bombs IN the GPU it'll lessen the chances of seeing rampant
proliferation again.
Ok, well enough of my musings over a GPU.
-AI
For me, it is far better to grasp the Universe as it really is than to persist in delusion
> 8x9800gx2s donated by NVIDIA.
;)
I wonder how many BSODFLOPS (Blue screens of death per second) it can generate?
http://byronmiller.typepad.com/byronmiller/2005/10/stupid_windows_.html http://www.google.com.au/search?q=nvidia+'blue+screen+of+death'+nv4_disp
... can it run Crysis on Vista at full settings? I kid, I kid!
My last sig was ridiculed
wouldn't that many gpu (8 on each ?) on a single motherboard fight for contention on the bus ? depends on the algorithm I suppose ... seems like a good excuse to try out the latest doom anyway
is it me or do I see two separate mobos...which means it's two machines, 8 per machine in one box....not 16?
now...if it was 16 in one...now that would be amazing....otherwise...it's not...'cuz there was that other group that did 8 in 1 (aka...16/2 => 8/1)
a system with 16 x 4870x2s. they will draw less energy too.
Read radical news here
Read the major mandate of the Los Alamos research centre: "Our highly skilled management team of nuclear experts and industry leaders are focused on making Los Alamos the premier national security laboratory of the 21st century."
This isn't for gaming, this is for planning how to more cost effectively kill humans.
There are 2 main differences between DX9 and DX10 :
I - The shaders offered by the two APIs are different (shader model 3 vs 4). None of the DX9 screen shot does self-shading. This is specially visible on the rocks (but even in action on the plancks of the fences). So there *are* available under Vista additional subtleties
II - The driver architecture is much more complex in Vista, because it is built to enable cooperation between several separate processes all using the graphics at the same time. Even if Vista automatically disables Aero when games are running full-screen (and thus the game is the only process accessing the graphic card), the additional layers of abstraction have an impact on performance. It is specially visible at low quality settings where the software overhead is more noticeable.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
God, they stuck so many fans into that box that I bet it takes off the ground when it boots.
Anyone please ammend the following to the quote of the article:
http://tech.slashdot.org/article.pl?sid=08/05/31/1633214
Do you know the human brain has about 100 billion neurons? Each neuron can be represented as a weighted average of its inputs, a typical human neuron has some 1000 inputs and does around a hundred operations per second.
So, yes, *maybe* there could be some very smart algorithm that mimics human reasoning, but that's not how it's done in the human brain. It's raw computing power all the way.
Thats an easy problem to solve! Just wait for the technology to mature before purchas...Oh.
Nvidia has to worry about ATI stealing their drivers {...} ATI knows that, and that's why their drivers are open.
We are not speaking about releasing source code of current drivers. In fact ATI/AMD's fglrx *IS NOT* open. At all. What is open are 2 *separate* drivers projects, which are done using the *technical data* released by AMD.
You're confusing the situation with Intel. (They paid Thungsten Graphics to write an open source drivers for i8xx/i9xx to begin with. There's no such thing as a proprietary intel drive on linux. Only an opensource driver written by TG)
What we want is not nVidia releasing the source of their drivers. What we want is nVidia providing enough technical data to the Nouveau project, so they can develop alternative open source drivers.
I think Nvidia's rationale for not giving up their specs is reasonable. Now, if they only cared more about their drivers for Linux, proprietary or not.
Their main rationale isn't about ATI peeking inside their spec. After all, what the open source projects are asking for (and obtaining from AMD and finally from VIA, but still not from nVidia) is not the *code*, but the *specs*. And the specs are only a description of how to interface the hardware. As both Radeon HDs and GeFroce use radically different designs, specs won't help much beyond have a little better idea of what the other hardware is doing under the hood. But there's no way the knowledge of which hardware registers does what that will help stealing driver code.
It fact that won't help at all because ATI and most open source projects are using Linux&BSD's standart Xorg + DRI stack, whereas nVidia use their own structure to handle 3D.
The most probable reason for not releasing the hardware specs, is that modern GPU hardware is horribly complicated. With a big number of teams working on an insane amount of components. Significant portions of technology going into a graphic cards might have been subcontracted or might be licensed from 3rd party providers.
It can be a real legal nightmare to track all the license dependency to clear a release. For nVidia which has covered a pretty big chunk of the market (Windows and some Linux running proprietary drivers), going through all legal hoops just to please a last few corner cases (linux running non-x86 hardware, other opensource systems, etc...) which only bring tiny fractions of market simply isn't worth at all.
This complexity can also be seen in AMD/ATI release scheme : /. a couple of months ago. They plan to use a different design on future product, having clearly separate video and hdcp units so releasing the specs of the former won't violate the licensing terms of the later. (They've also told in other places that they will move to more open-source friendly designs overall).
- They release the specs one piece at a time, with long delay because of the necessary check with the legal department to obtain clearance. (Sometimes criticized on irc channels for the slowness of the process)
- Some specs won't be releasable at all. In current generations of GPUs, the video decompression acceleration is intricate with the HDCP encryption. And the licensing terms of the later prohibits the release of the specs of the current video+hdcp unit. This was mentioned in an interview on
The situation is somewhat different with VIA as : /. that their chips have video decompression unit, but doesn't provide any software middleware
- Their chips are rather small projects combining less technologies and mostly done in house. The licensing complexity is smaller.
- They don't license 3rd party technologies, but instead count on OEMs and system integrators to license what they needs themselves. This may poses problem for Windows development (it was recently mentioned on
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
So it's about one fan per GPU? Seems annoying and inefficient. Why not build it more spread apart, or use a "Central Air" system like people use in their homes.
Not using water cooling I understand, 'cause there'd be around 30 tubes snaking in and out of the box - something would fail/leak.
..........FULL STOP.
They should use the quantum computer described a few posts above, it seems to be especially designed for pattern matching that computer vision might require.
I think it's great to see that we can finally start using GPUs to do things beyond gaming, but I also don't see it as the Great Second Coming of high-speed computing. GPUs are designed to tackle only one kind of problem, and a highly parallel problem at that. If you are a researcher and you can see huge gains in performance by using GPUs, then great! But GPUs are hardly general purpose, and will simply not address most of our computing needs. I see the rise of GPUs as similiar to computing in the 60's(?). Figure out what kind of software you need to run, and then design a computing platform around it. If you need to perform small operations on a highly parallel data structure, then a GPU cluster is an excellent way to go. -One beginning computer architect's opinion
http://acceleware.com/newsEvents/newsreleasearchive/20080617clustersolution.cfm
The GPU architecture has been progressively moving to a more "general" system with every generation. Originally the processing elements in the GPU could only write to one memory location, now the hardware supports scattered writes, for example.
As such I think the GPGPU method of casting algorithms into the GPU APIs (CUDA et. al) are going to die a quick death once Larabee comes out and people can simply run their threaded codes on these finely-grained co-processors.
UBU
Pathetic. Not my field, I can't comment either. But all those upmodded jokes and irrelevant and uninformed side discussions about stream processing on GPUs shows that /. modders often mod crap just because they have nothing else to mod up.
Hey, if all one has in a forum is total crap, mod nothing.
But does it run Crysis?
On June 30 of this year, The New Yorker magazine published a fascinating, if at moments disturbing article entitled The Itch. The article discusses, among other things, the human mind's perception of the reality of its environment based on the various nervous inputs it has, vision included. Apparently this is an oft debated topic among the scientific community, but it was new information to me.
One of the things I found intriguing was the note that the bulk (80%) of the neural interconnections going into the visual cortex of the human brain come not from the optic nerves themselves, but from other areas of the brain including "functions like memory." The suggestion is that the eyes provide visible light input, but that the brain's processing of what it is looking at is primarily an act of abstract object/pattern reconstruction.
If you couple this notion with the limitations of the human eye itself -- such as the fact that the finest resolution/detail comes from an incredibly narrow region directly in the center of the FOV with rapidly decreasing information towards the extremities that ranges from soft profiles to mere suggestions of color, brightness and movement -- it strikes me that if researchers at MIT wish to replicate the model of human vision on any level close to reality, their input and processing systems should actually be modeled a bit like the real deal.
I believe that without a solid neural net with strong pattern recognition, instant recall, massive parallel processing - the finer things of the visual cortex - that human vision will not be possible with semiconductors. Is 16, 32 or 100 GPU's substantial enough to pull it off? Probably not, I think - too much overhead, not enough interconnects... to coin a pun: "too RISCy". I must admit though that the research discussed in the technologyreview link is very interesting.
I looked through each of TFA's linked in the story, and I don't see any technical details on this system. Whereas when the FASTRA people at Univ. of Antwerp put together their 4 9800-GX2 system for CUDA, they published all the nitty gritty down to specific parts, etc. The pictures are interesting but not enough.
So which one of those guys in the pics is Miles Dyson?
I had eight Quadro Plex units where I used to work for CAD/CAM/FEA/CFD...a year ago.
I think this system needs to be redefined as a 16 gpu cluster. Each gpu card has it's own fully functional 3d generator ( call this the OS ) in it's own rom. In stead of a lan/(i)scsi-bus/optical-fibre interconnecting the cards, they are using pci-express to interconnect the cards. It's even actually 2 clusters, seeing as there are 2 computers in the system with each a cluser o 8 gpu-systems.
1) Your task has to be highly parallel. You really need something that can be made parallel to a more or less infinite level. Current GPUs have hundreds of parallel shader paths (which are what you use for GPGPU). So you have to have a problem that can be broken down in to a bunch of small parallel processes.
2) Your task needs to be single precision floating point. The latest nVidia GPUs do support double precision, but they are the only ones, and they take a major, major speed penalty (way over 50%) to do it. Thus your task needs to be 32-bit FP numbers, as that's what the GPUs like to crunch.
3) Your task needs to have a minimal amount of branching. GPUs now can do if-then sort of logic, but it incurs a pretty big penalty. So your task needs to be largely free of that. Needs to be the kind of thing where you are doing lots of predictable calculations, not a whole lot of indeterminate tests.
4) Your task needs to fit in to the available RAM. GPUs have a lot of RAM, but not nearly as much as normal computers. The most a consumer GPU has right now is 1GB for the nVidia GTX 280. The vast majority of them have only 512MB (ATi 4800 series, nVidia 9000 series and most of their 8000 series). You can get specialised boards with more, nVidia calls them Tesla, but the cost more and they still cap out at 4GB. So you have to be able to describe the problem in that amount of memory. While GPUs can access system RAM is is much, MUCH slower. The PCIe bus is only a couple GB/sec and video memory is over 100GB/sec in the high end GPUs. Thus if your task won't fit in GPU memory, you are going to lose much of the performance to swapping.
Those are the principal limitations. If your task is highly parallel, 32-bit FP, mostly linear, and under 1GB, then you'll likely find a GPGPU solution works well. If not, you'll want to do more research and testing first. It isn't that they might not help, but it also might not be the big gain you were hoping for.
CAE's Tropos image generators use 17 GPUs per channel in a commercially available package. Each image channel (there are usually at least 3 in a flight simulator) uses 4 quad-GPU Radeon 8500 cards in addition to the onboard GPU which is only used for the operator interface. I've been working on these things for a couple of years now.
You laugh, but it seems like my eyes have gotten faster.
I used to not care about 60hz refresh rate, but now I can't stand it. Look straight ahead at a CRT monitor running 60 hertz looks like a rapidly flickering/shimmering mess. 70 hz is still annoying b/c my peripheral vision picks it up.
I attribute my increased sensitivity to flicker to playing FPS's.
Oh and when a decent brain-computer interface comes out I'll be getting one installed.
They ARE out to get you simply because They are in it for themselves and they don't care about you.