AMD Plans 1,000-GPU Supercomputer For Games, Cloud

arcticstoat writes "AMD is planning to use over 1,000 Radeon HD 4870 GPUs to create a supercomputer capable of processing one petaflop, which the company says will make 'cloud' computing a reality. When it's built later this year, the Fusion Render Cloud will be available as an online powerhorse for a variety of people, from gamers to 3D animators. The company claims that it could 'deliver video games, PC applications and other graphically-intensive applications through the Internet "cloud" to virtually any type of mobile device with a web browser.' The idea is that the Fusion Render Cloud will do all the hard work, so all you need is a machine capable of playing back the results, saving battery life and the need for ever greater processing power. AMD also says that the supercomputer will 'enable remote real-time rendering of film and visual effects graphics on an unprecedented scale.' Meanwhile, game developers would be able to use the supercomputer to quickly develop games, and also 'serve up virtual world games with unlimited photo-realistic detail.' The supercomputer will be powered by OTOY software, which allows you to render 3D visuals in your browser via streaming, compressed online data."

Only 1.000?

[Duncan3]

Folding@home is at 1.007 PFLOPS of just ATI GPUs :)

(which is an entirely different sort of "computer", but still)

What about latency in gaming?

[WiiVault]

I'm all for cloud gaming- it would be great to not have to upgrade my GPU all the time to play new games, however I wonder how this could be accomplished in a way where lag was so minimal as to not affect gameplay. It seems this would be especially hard if one were to play online games. Correct me if I'm wrong but it would seem one would need to add the lag from the client to the cloud AND the lag from the player to player (or server) in the multiplayer networking. That seems like a too much lag for most FPS's, which I'm assuming are one genre which would gain the most from such a supercomputer.

Re:Contest, the rematch...

[Zephiris]

Nvidia's GTX 295 was around 1.7 teraflops I believe, while the (similarly priced) 4870X2 is 2.4. The 'mere' 295 supposedly beats the 4870X2 by 15% average.
The difference is? Nvidia always has pretty good drivers. ATI struggles to allow games to take >50% advantage of even the lowly 3870 (as measured by the card's own performance counters)...let alone a 2.4 tflop card...let alone a massive array of 4870s.

Plus, wouldn't a 1000 GPU 4870 cloud...only allow some 1000 users some fractional percentage of one 4870 capped by latency and other overhead?

Or...are we talking about providing a larger number of mobile devices the equivalent capabilities and speed of 1999's Geforce 256?

Either way...I don't think it'll catch on, and will be a huge money sink for AMD when it needs to be fixing its processor and video card issues for the average, real consumers who are losing faith in AMD's ability to provide reasonable and usefully competitive products.

Ah, the Big Iron versus micros war again....

[macraig]

Figures. See, most people thought that war had been won long ago. Perhaps it was, but now the Big Iron camp has a new ally: Big Software, who REALLY wants to do away with one-time licenses and purchases and substitute the far more lucrative "Web apps" and the subscription licensing and fees that paradigm will allow. They want to re-brand software as "content" and they want consumers to willingly buy into that. Their latest sneaky flanking maneuver is what you know as Web apps, but the objective is the same.

If you say yes to either one, centralized computing or software subscriptions, you're actually saying yes to BOTH.

Nancy Reagan had the better advice: Just Say No... to both.

Re:No, latency

[i.of.the.storm]

For some people, maybe. But professional RTS gamers can have between 300-400 actions per minute, and some ridiculously good ones have 500, and if they had that much lag I wager they would notice. And of course, that's on top of the amount of time it takes for the supercomputer to generate the image.

Practical benifits

[cee--be]

One practical use for this would be to run staggeringly complex real-time physics calculations in real time. One example would be doing the necessary calculations to render a physically realistic sea with weather conditions into an animation. You could then send this to users in a sea MMO for example. There are many other cool game related things you could do with it, rather than wastefully rendering some uncanny valley mobile phone game at 2 FPS.