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."

Comments

Just to get it out of the way...

[im_thatoneguy]

"But can it run Crysis?"

*Ducks*

Re:Just to get it out of the way...

[dtml-try MyNick]

The graphic cards might be able to handle it. But you'd have to run Vista in order to get maximized settings..
Now there is a problem.

Re:Just to get it out of the way...

[kaizokuace]

no it isn't. Duke Nukem Forever will be released when a powerful enough computer is assembled. The game will just manifest itself in the machine one powered up. But you have to have downloaded 20TB of porn and covered the internals with a thin layer of cigar smoke first.

So that's what happens

[chabotc]

When gamers grow up and go to college.. blue leds and bling in the server room!

Alright!

[religious freak]

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)

Just how specialized is GPU hardware?

[MR.Mic]

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.

Re:Just how specialized is GPU hardware?

[hansraj]

Not really. Not every problem gains from a gpu.

As a rule of thumb, if you problem requires solving many instances of one simple subproblem which are independent of each other then a gpu helps. A gpu is like a cpu with many many cores where each cpu is not as general purpose as your intel, rather each core is optimized for some solving small problem (without optimizing for frequent load/store/switching operations etc that a general cpu can handle quite well).

So if you see an easy parallelization of your problem, you might think of using a gpu. There are problems that are believed to not be efficiently parallelizable (Linear Programming is one such problem). Also, even if your problem can be easily made parallel it might be tricky to benefit from a gpu as each subroutines might be too complex.

I don't program but my guess would be that if you can see the solution to your problem consisting of a few lines of codes running on many processors and gaining anything, a gpu might be the way to go.

Perhaps someone can explain it better.

Re:Just how specialized is GPU hardware?

[moteyalpha]

I have been using my own GPU to do this very same thing by automatically converting images to vertex format and use the GPU to scale, shade, etc and in this way I can have a shape recognition by simply measuring the closest match on the frame buffer. There are more complex ways to use the GPU to do pseudo computation in parallel, I still think that a commonly available CAM or near CAM would increase neural like computations by being essentially a completely parallel process. It would be better to allow more people to experiment with the methods because the greatest gain and cost is the software itself and specialized hardware for a single purpose allows better profit but limits innovation.

Re:Just how specialized is GPU hardware?

[TheLink]

You also need to make sure the I/O to/fro the GPU is good enough.

No point being able to do calculations really fast but not be able to get the results or keep feeding the GPU with data.

I think not too long ago graphics cards were fast, but after you added the problem of getting calculation results back, it wasn't really worth it.

Re:Just how specialized is GPU hardware?

[TapeCutter]

I think you did a good job explaining, one point thought. The sub-problems need not be independent.

Many problems such as weather prediction use finite element analysis with a "clock tick" to syncronise the results of the sub-problems. The sub-problems themselves are cubes representing X cubic kilometers of the atmosphere/surface, each sub-problem depends on the state of it's immediate neighbours. The accuracy of the results depends on the resolution of the clock tick, the volume represented by the sub-problems and the accuracy of the initial conditions. This is usefull in all sorts of simulations from designing molds to minmise air pockets that can plague the process of metal casting, to shooting Cassini through the rings of Saturn, twice!

The technique can be thought of as brute force integration with respect to time, space, matter, energy, etc, for a wide range of physical problems. The more power and raw data you throw at these types of problems the more realistic the "physics" in both video games and scientific simulations. IMHO we have only just scratched the surface of what computers can tell us about the real world through these types of simulations and much of this is due to scientists in many fields confusing "computer simulation" with "artists impression".

BTW climate and weather modelling use the same sort of algorithm but get very different results because weather != claimte.

Re:Say no to proprietary NVIDIA hardware

[cheater512]

I want to support ATI and AMD, but nVidia just works.
Their drivers are very nice.

Until that changes I'm a nVidia guy.
A third party open source driver should fix the problems.

Re:It's coming

[Anonymous Coward]

Jesus doesn't approve of you doing that.

Re:Say no to proprietary NVIDIA hardware

[ya really]

Tom's Hardware did a pretty good job detailing the ups and downs of ATI and Nvidia with many of the major games of last year (BioShock, World in Conflict, etc). Overall, both companies faired well, but they reported quite a few crashes due to the ATI drivers. I've had an ATI card before, the 9800xt when Nvidia was producing their horrible 5xxx series back in 2003-04 that was totally worthless. The 9800xt was a good card for everything (gaming, graphical aps, etc). Sorry, I should have cited sources. Wasn't trolling on purpose, though I know that writing anything positive about Nvidia on slashdot is borderline blasphemy.

Fascinating

[AlienIntelligence]

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

Re:Fascinating

[blahplusplus]

"I think this part of the computing timeline is going to be one that is well remembered. I know I find it fascinating."

Well remembered? Perhaps... but I wouldn't sing their praises just yet. Advances in memory are critically necessary to keep the pace of computational speed up. The big elephants in the room are: Heat, memory bandwidth and latency. Part of the reason the GPU's this time round were not as impressive is because of increasing memory bandwidth linearly will start not have the same effects sooner or later, there is a point at which the geometry of information will come into play and the law of diminishing returns on a kind of architecture or memory will take place for a price that people can afford to pay.

Next up, 32-bit addressing is starting to be a real pain in the ass. The move to 64-bit operating systems is critical if we are to expect GPU's to keep increasing their memory (1GB+ of local memory on a card now).

Supreme commander was one of the few games to hit the 4GB addressing limit and more and more games will definitely do so in the future. I know you were talking about other areas of computing, but without the games market, I don't see any serious reason for any regular person to upgrade their computers video at all. The many who donate to distributed computing did so as an afterhtought, not as the main reason they bought the card... As for the wider non-gaming market we'll have to see whether not GPU computing is going to be moer widely adopted.

Lastly, let us not forget that one of the primary reasons GPU computing is so fast is memory bandwidth, delay's in better memory technology will have big impacts on GPU performacne. As we've seen with this generation of GPU's, Nvidia's lack of DDR5 and a smaller process for the GT280 hurt them a lot.

machine or machines?

[MoFoQ]

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)

Re:machine or machines?

[sam0737]

That's one machine for simulating one eye. That's why they need 2 * 8 for simulating human-level vision, or else you won't get the 3D vision.

Re:Say no to proprietary NVIDIA hardware

[TheRaven64]

A video card driver typically has three major components:

• The parts specific to the windowing system (including context switching / multiplexing).
• The parts specific to the 3D API.
• The parts specific to the hardware.

ATi could conceivably steal parts from the first two from nVidia, but it's doubtful that they could steal anything from the last part since their hardware designs are sufficiently different to make this hard.

The problem nVidia are going to have is that the new Gallium architecture means that the first two parts are abstracted away and reusable, as is the fall-back path (which emulates functionality any specific GPU might be missing). This means that Intel and AMD both get to benefit from the other company (and random hippyware developers and other GPU manufacturers / users) improving the generic components, while nVidia are stuck developing their own entire alternative to DRI, DRM, Gallium, and Mesa. The upshot is that Intel and AMD can spend a tiny fraction of the time (and, thus, money) developing drivers that nVidia do. In the long run, this means either smaller profits or more expensive cards for nVidia, more bugs in nVidia drivers (since they don't have the same real-world coverage testing).

Now, if you're talking just about specs, then you're just plain trolling. Intel doesn't lose anything to AMD by releasing the specs for the Core 2 in a 3000 page PDF, because the specs just give you the input-output semantics, they don't give you any implementation details. Anyone with a little bit of VLSI experience could make an x86 chip, but making one that gives good performance and good performance-per-Watt is a lot harder. Similarly, the specs for an nVidia card would let anyone make a clone, but they'd have to spend a lot of time and effort optimising their design to get anywhere close to the performance that nVidia get.

DX10 vs DX9

[DrYak]

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.

Is that a helicopter?

[Anonymous Coward]

God, they stuck so many fans into that box that I bet it takes off the ground when it boots.

becoming less specialized...

[Mad Hughagi]

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.