Slashdot Mirror
Asus Crams Three GPUs onto a Single Graphics Card
Barence writes "PC Pro has up a look at Asus' concept triple-GPU graphics card. It's a tech demo, so it's not going to see release at any point in the future, but it's an interesting look at how far manufacturers can push technology, as well as just how inefficient multi-GPU graphics cards currently are. 'Asus has spaced [the GPUs] out, placing one on the top of the card and two on the underside. This creates its own problem, though: attaching heatsinks and fans to both sides of the card would prevent it from fitting into some case arrangements, and defeat access to neighbouring expansion slots. So instead, Asus has used a low-profile heat-pipe system that channels the heat to a heatsink at the back of the card, from where it's dissipated by externally-powered fluid cooling pipes.'"
The technology for multi-GPU processing is already out there (SLI, Crossfire), and now the companies are trying to increase the number of GPUs that can be daisy-chained (CrossfireX, 3-way SLI).
However, it seems with all of these methods, the weak link is always driver support. I think that drivers will have to develop further before anything like this can take true form and be useful.
As an aside, did anyone notice that half of the Slashdot description sounded like an advertisement for Asus GPU cooling?
Fuck Everything, Were doing five cores.
So does that mean we can play Crysis now?
Remember when All razors had a single blade? Then double blade razors were all the rage. These days, Triple and quad blade razors are around. Soon we will have 5 blades but I would call that a cheese grater.
Same thing with CPUs and now GPUs. Problem is, at what point dose it become a pissing contest rather than a way to provide more performance for an application that needs it.
And speaking of those demanding applications. Am I the only one who notice that some of the latest video games running on the best available hardware provide no improvement in appearance or game-play over older games of a similar type running on older hardware?
It's bad enough that I am tempted to think the programmers are just adding fat to make sure the game demands a more expensive video card.
Kevin.
--= Isn't it surprising how badly I spell ?
How many power supplies are required? Does it come with a 12 KV step-down transformer and 220V three-phase power hookup? Can I heat the basement with it?
The end of the article makes a good point. While we have dual- and tri-core graphics setups now, the programs are not designed to exploit them. This is the same issue that is being faced in the general CPU market as well. If you don't have a multi-threaded app for multiple CPUs, you only gain in multitasking, but not in a single program. A serial program can ONLY run serially. There's only so much parallelism that a CPU can infer. And at that point, you have out-of-order execution to make up for mutli-cycle instructions.
The two ways to do multi-core graphics would be to have each GPU generate a separate frame. This would be done in the driver, but since video is dependant on the previous frame there's not as much gain as is theoretically possible because you're still dependant upon the result from the previous frame. The other option is in the app itself, which "knows" what's going to be where at a certain frame and can just send that to each GPU. This would create tremendous speedups because now each frame is generated independently. Thus a 3X setup would achieve 3X performance (again theoretically). These options are probably mutually exclusive though because an app written to take advantage of mutli-GPUs could be foiled by a driver written to take advantage of them. It's a choose your poison option. And currently we're not getting either.
It's not necessairly a limit of the board design, but a limit to what game engines can be optimized for. Most game engines do not scale well beyond two cards, as can be seen here:
http://www.xbitlabs.com/articles/video/display/zotac-9800gx2.html
While there are a few key games that get no boost out of 2-way SLI, the vast majority of games do see improvement. 3-way, on the other hand, can actually cause WORSE performance.
It probably has to do with limitations on how the SLI/Crossfire drivers can fake-out the game engine. There are probably limits to how many frames the game engine allows to be in-flight at once, limiting how much performance boost you can get from AFR SLI. And although you can get around game engine limitations with split-screen rendering, this mode needs specific game support, and shows less potential performance increase. Plus, split-screen rendering and has to be selected explicitly in Crossfire (AFR is the default).
Man is the animal that laughs.
And occasionally whores for Karma.
Why can't we put our efforts toward more efficient GPUs? Just as most users won't ever be able to push their current CPUs at their maximum, most aren't even using the full power of their GPUs.
I want a fanless, 5W GPU with the power of GPUs from about 3 years ago. Can the new smaller transistors allow for this or am I asking for too much?
If ATI and nVidia keep pushing for raw power, they'll get beaten to the low-power finish line by the likes of intel and VIA.
The FA states that multi-threaded gaming is fundamentally flawed. How is this a valid statement? They are testing a multi-cored GPU using games that were most likely only developed to use, at MOST, two cores. Regardless of how many core you throw at it, the application (ie. your game) will NEVER use any of those cores.
In fact, gaming and graphics scale amazingly well as a multi-threaded application. In fact, as many in the graphics/gaming community have been stating recently, ray tracing would benefit greatly from more GPUs. Being able to trace multiple rays at a time would speed up rendering.
They state that it is fundamentally flawed when they should have said that it would be ignorant to assume that an application designed to use a single-core or dual-core GPU would benefit from extra GPUs.
-SaNo
As opposed to raytracing, which is so extremely parallel that it scales nearly linearly. In other words, your 3-way real-time raytracer graphics card (if/when such a beast is ever made) would perform at about 2.8x the one-GPU variant. And unless the Rapture^Wsingularity keeps you from getting a 192-GPU card, it'd render at 170x the reference or so.
(Of course, there's the question of global illumination. I don't know if those can be parallelized as easily, but there was a story about distributed photon mapping here some time back, where they used Blue Gene.)
" so it's not going to see release at any point in the future,"
The future is a really long time.
The Kruger Dunning explains most post on
"Fuck everything, We're doing three GPUs!"
...this is old news. They've already done five cores some time ago.
Big! Strong! Wow! Tada-O!
Wow! A beowulf cluster of graphics cards! That certainly ought to be enough to run Duke Nukem Forever!
This is a terrific advance in the field of cramming. I'm looking forward to seeing their presentation at the Cramming and Stuffing conference later this year.
I fail to see why they even bothered to build this thing. Anyone who's been paying attention would know that multi-GPU technology is a clusterfuck. Nobody has a stable, reliable implementation that actually yields respectable performance. A 20-30% increase for GPU-bound applications is simply pathetic... often times that increase is 0%, if a game is not SLI-enabled by its developer or by the graphics drivers.
When they come up with a multi-GPU system that appropriately virtualizes the whole thing, enabling ALL apps to benefit, I'll buy. Right now, it's just double the money for none of the fun.
-Billco, Fnarg.com