Game Streaming On-Demand Will Eventually Popularize - Looking at GeForce Now

Video on demand has become ubiquitous. Netflix, Hulu, Amazon Prime, and even ISP-bound options offer some form of movies or TV in exchange for service charges. It is no longer the “future of TV” – it is TV. This is the reality of the modern, cable-cutter era, and it is only a matter of time until all traditional cable services are axed in favor of streamed-via-internet options. Cable television will die.

The same could be true for gaming. The data is bigger, the latency demands are greater – but these problems are conquerable and will only diminish as high-speed internet proliferates. Input latency is the most critical. The time from button-press to photon-on-screen dictates whether a game is playable; it's more significant even than frametimes. Long poll times or slow frame encode/decode will create a mismatch between the player's actions and the perceived outcome, resulting in frustration that we've all experienced at some point with server lag in traditional gaming.

(Side note: We mention some GameWorks updates in the above video. You can find that content here.)

NVIDIA's been developing its “GRID” technology for a few years now. It's been through an excruciatingly long beta period, and now has become officially available to consumers under the “GeForce Now” brand. GeForce Now streams triple-A desktop games down to compatible devices – which is presently only the Shield console ($200) – and uses massive server racks to perform all game logic and graphics computations remotely. The packaged frames are then shipped down the pipe to the user, who's in turn sending responding input up the datastream. This function is similar to other streaming services we've discussed – encode/decode occurs on the GPUs, but heavy-duty processing is all remote. Interesting use cases arise where low-end platforms – like the Shield, which hosts only a Tegra GPU – can accept Ultra-quality rendered output.

Total latency target is <150ms from controller to photon. Total input latency is comprised of game engine processing time (<60ms), raw input/button-press-to-Shield latency, and ping-time from Shield to the server (<30ms target). Servers are located in Tokyo, Singapore, the Bay area, Virginia, Oregon, and Dublin. A total of 100-150ms target is around “console range,” from what the engineering team tells us.

The service is $8/mo, in-line with on-demand video, and offers access to all games on the service as long as the account is open. That's important, of course – like most on-demand movies and TV, access is lost once the service is terminated. A few sample games presently available include:

• Android apps/games (Shield runs on Android)

• The Witcher 3

• Tomb Raider

• DiRT / GRID

• Borderlands

• Sleeping Dogs

• Massive Chalice by Double Fine

• Metro: Last Light Redux

• Trine 3

• LEGO titles

• F1 2015

• A pile of Sonic games – presumably at least one of them is not terrible.

The largest and closest “competitor” (though the word “alternative” may be more accurate) in the space would be Steam, which works much differently. It's possible to use Steam's “in-home streaming” to render content on a primary gaming desktop (e.g. an upstairs PC), but play that content on another system (e.g. a downstairs, TV-attached PC). This works over LAN and naturally has reduced latency. It's possible to do this with several other services as well, including NVIDIA's own game-streaming tools for owners of the company's GPUs. All of this necessitates that the user own the game and the hardware required to render that game's content. In the case of GeForce Now and other internet-streaming solutions, a remote server allocates dedicated system resources (similar to how a VPS would work) to the user, then processes data. Each user is assigned dedicated CPU threads and GPU processing power, with some level of scalability as needed. The Shield console uses its on-chip encoder/decoder to manage the received frame data and dispatched input data and will be updated to Maxwell this year, which will further increase encode/decode speed.

None of these technologies are really ready for competitive gaming just yet. It's getting there, but an extra 100-150ms from time of input (and the 50ms variance is major, too – consistency is key) kills the potential of playing CSGO or Dota2 at a high-level. Anyone who's competitive needs to be as close to the processing as physically possible, and that's always going to be a desktop.

While talking to this point, it seems worth noting that game streaming on-demand – although we do think it is eventually going to “happen” – is not a replacement for desktop enthusiast setups. These services will expand the gaming audience to the folks who'd never build or maintain their own rig, and will grow gaming as a whole. There's no threat to DIY-style PC gaming.

Out of pure curiosity, we asked NVIDIA how they run servers that can sustain such respectable latencies and ping-times while at a trade show. Anyone in the industry knows that convention center internet is atrociously expensive to vendors – we're talking thousands to tens of thousands of dollars for a few days – and also atrociously slow. It's nigh impossible to get an outbound signal from PAX East, for instance. Any vendor who can establish a functioning, server-oriented setup from the bowels of a trade show has some engineering experience, and we wanted to know how the GeForce Now team does it.

Chad Cooper of NVIDIA, an engineer on the project, installs and maintains what is “basically a monster gaming PC” as a GeForce Now server-in-a-closet. It gets installed in a tiny room within nVidia's booth, then is locally routed to all the demo systems. That “monster PC” is outfitted with several of nVidia's M5000 GPUs (GM204 Maxwell class), about equivalent to a GTX 980. The rig runs several VMs, but speaks to Amazon's AWS “cloud” service for game saves and validation.

From an op-ed standpoint, we do believe that game streaming will eventually roll-out through traditional ISPs and other solutions. NVIDIA's GeForce Now is certainly one pathway. Other services will rise and compete, too, and we'd expect competition (or licensing of GeForce Now / similar services) to eventually propagate in TimeWarner, AT&T, and similar plans. Whether or not GeForce Now is the end-all solution, we do feel strongly that gaming on-demand is only as far away as more widespread internet speed hikes -- and those are rolling-out globally.

Note: GeForce Now SDK available here, for developers and publishers interested.

Editorial: Steve “Lelldorianx” Burke
Video Production: Keegan “HornetSting” Gallick