Current server products give you 4 GPUs into 1U or 8GPUs into 4U (depends on the storage and CPU requirements). That's pretty good but when people optimize for this it could also increase.
Those GPUs tend to have more memory. Given average VRAM usage you could have five or six gamers on a GPU before memory is a problem. The other big issue is compute but there are tricks there too.
Not all gamers are in areas of high complexity all the time and not all gamers are running at the same resolution all the time and not all games require the same frame rates. Different service levels would give different quality settings I assume. You might be able to have two 4K gamers on a GPU or 8 1080p gamers on it. With clever load balancing you could mix and match much like we do with virtual machines that have very different usage patterns.
We also have clever techniques like variable rate shading and 'radeon chill' that reduces overall load allowing more simultaneous users. And in the future things will get even more clever with distributing load over multiple GPUs for even more fine grained control.
I can envisage 100,000+ simultaneous users supported in only a moderate 200 rack installation. A large data center might have 1,000 to 2,000 racks.
It is a way to lock you into a subscription model but there are also reasons not to hate it.
1. Roughly speaking there's about a billion gaming consoles on the planet. More if you include handheld. That's a lot of hardware being produced that ultimately spends a lot of time unused. Consolidating it into managed clusters which can also be used for other things during off-peak is highly efficient. Less e-waste and that's not a bad thing.
2. People would be able to play really high-end games on commodity hardware. No more $3-4,000 gaming PCs or even a need for higher end consoles. Anybody with a browser, screen, and internet connection can have a similar experience. Sure there will be the HD, 4K, 8K services at different price points but there will be more than just Google's service so competition I hope will keep prices in line. It's computationally much more expensive than Netflix but shouldn't cost an order of magnitude more.
3. Graphical and feature upgrades happen magically. Your display hardware becomes obsolete at a much slower rate. You'll one day want an 8K screen upgrade but everything else from ray-tracing to advanced AI can just happen in the background without a new top of the line GPU being required.
4. Patching just happens.
5. No install times.
6. Almost no load times either conceivably.
7. Developers would be free to use very powerful backend services. Not being limited to a single system they could run graphics on one cluster, AI on another, physical simulations on another. Microsoft and Cloudgine showed a demo of something like this in 2015 with Crackdown 3. They showed off an entirely destructible landscape with multiple players possible because that expensive computation was running in their cloud service.
8. That brings us to new features. That demo was pretty cool but imagine a GTA online style game where the entire world and events within can be permanent rather than just localized and respawning. It opens up new storytelling abilities. Entirely new modes of play. Far more immersive and advanced world simulations would be possible.
9. And as you say you're instantly cross-platform and you're also portable.
Ultimately it's going to be convenience that wins.
Slashdot Mirror
Current server products give you 4 GPUs into 1U or 8GPUs into 4U (depends on the storage and CPU requirements). That's pretty good but when people optimize for this it could also increase. Those GPUs tend to have more memory. Given average VRAM usage you could have five or six gamers on a GPU before memory is a problem. The other big issue is compute but there are tricks there too. Not all gamers are in areas of high complexity all the time and not all gamers are running at the same resolution all the time and not all games require the same frame rates. Different service levels would give different quality settings I assume. You might be able to have two 4K gamers on a GPU or 8 1080p gamers on it. With clever load balancing you could mix and match much like we do with virtual machines that have very different usage patterns. We also have clever techniques like variable rate shading and 'radeon chill' that reduces overall load allowing more simultaneous users. And in the future things will get even more clever with distributing load over multiple GPUs for even more fine grained control. I can envisage 100,000+ simultaneous users supported in only a moderate 200 rack installation. A large data center might have 1,000 to 2,000 racks.
It is a way to lock you into a subscription model but there are also reasons not to hate it. 1. Roughly speaking there's about a billion gaming consoles on the planet. More if you include handheld. That's a lot of hardware being produced that ultimately spends a lot of time unused. Consolidating it into managed clusters which can also be used for other things during off-peak is highly efficient. Less e-waste and that's not a bad thing. 2. People would be able to play really high-end games on commodity hardware. No more $3-4,000 gaming PCs or even a need for higher end consoles. Anybody with a browser, screen, and internet connection can have a similar experience. Sure there will be the HD, 4K, 8K services at different price points but there will be more than just Google's service so competition I hope will keep prices in line. It's computationally much more expensive than Netflix but shouldn't cost an order of magnitude more. 3. Graphical and feature upgrades happen magically. Your display hardware becomes obsolete at a much slower rate. You'll one day want an 8K screen upgrade but everything else from ray-tracing to advanced AI can just happen in the background without a new top of the line GPU being required. 4. Patching just happens. 5. No install times. 6. Almost no load times either conceivably. 7. Developers would be free to use very powerful backend services. Not being limited to a single system they could run graphics on one cluster, AI on another, physical simulations on another. Microsoft and Cloudgine showed a demo of something like this in 2015 with Crackdown 3. They showed off an entirely destructible landscape with multiple players possible because that expensive computation was running in their cloud service. 8. That brings us to new features. That demo was pretty cool but imagine a GTA online style game where the entire world and events within can be permanent rather than just localized and respawning. It opens up new storytelling abilities. Entirely new modes of play. Far more immersive and advanced world simulations would be possible. 9. And as you say you're instantly cross-platform and you're also portable. Ultimately it's going to be convenience that wins.
I'd rather talk to somebody who might know something about it.