Nvidia just finishing telling us about how it's going to stick a Kepler GPU in the cloud: now, CEO Jen-Hsun Huang is telling us how it will use distributed graphics to stream low-latency video games from the internet to computers that don't have one themselves. Nvidia's partnered with cloud gaming provider Gaikai, and claims that the GeForce Grid GPU has reduced latency of streaming games to just ten milliseconds by capturing and encoding game frames rapidly, and in a single pass, and promises that the enhanced Gaikai service will be available on TVs, tablets and smartphones running Android and iOS.

David Perry from streaming game company Gaikai is on stage to discuss and demo the technology now; Gaikai also announced that it's working in partnership with Nvidia to produce content for GeForce Grid. Perry showed off an unreleased FPS shooter called Hawken running on an Asus Transformer Prime powered by the new GeForce Grid GPU. Perry played against an opponent who was running the game on an LG Cinema TV without the need for an external console — Huang made a big deal of showing off an ethernet cord hooked up to the TV, saying that the cord was the console. One of the most noteworthy aspects of the GeForce Grid demo is the fact that gamers can play across platforms of varying power, thanks to the Grid doing all the heavy lifting. Also of note in that demo was that the server hosting the gameplay was a good 10 miles away — but the gameplay was lag-free, as you'd expect to see when playing on a local machine.

From a hardware perspective, the GeForce Grid features two Kepler GPUs with 3,072 CUDA cores on each processor, providing 4.7 teraflops of 3D shader performance. The number of CUDA cores in the Grid is the same as Nvidia's GTX 690 graphics card; it seems that Nvidia is putting its latest consumer-facing graphics card technology right into the Grid, though this card has a 250W TDP compared to the GTX 690's 300 watts.

Following the presentation, Jen-Hsun Huang was able to provide a little more detail on how exactly GeForce Grid works, in response to a question by PC Perspective:

Where we used to render from frame buffer and copy to the CPU for compression and streaming, here it's already streaming right out of the GPU, saving encode time, not to mention copy time. Compressing and streaming in parallel... we've taken maybe a couple hundred milliseconds of lag, and reduced it to something that's the same performance and snappiness as a game console.

Nothing special is needed at the client-end, just a standard H.264 hardware video decoder. Anything that can play YouTube can run GeForce Grid-streamed games, Huang told us. He did say hat there's still optimizations possible at the client end to further reduce latency, though, but wouldn't confirm whether Nvidia or partners were working on a Geforce Grid-specific client for the task.

We also heard that Nvidia was considering offering Netflix-like subscriptions to GeForce Grid, with something like a $10 monthly fee for access to the streaming game catalog — such a move would put the service in direct competition with OnLive's streaming game service. In addition to Gaikai, Nvidia has partnerships lined up with Ubitus and Playcast, two other streaming game service providers. There's also a few developers pledging support from the start, including Epic, Capcom, and THQ. Unfortunately, while Nvidia appears to have the right technology and some partners to kick things off, there's no real timeframe yet for when GeForce Grid might make its way down to consumers.

Update: This post originally conflated Nvidia's VGX virtual GPU with the GeForce Grid streaming-friendly GPU, but they're different things.

Nathan Ingraham contributed to this report.