Is CPU physics worth it or is moving to GPU a better choice?
-
mrheisenberg
- Posts: 1
- Joined: Tue Nov 13, 2012 2:47 am
Is CPU physics worth it or is moving to GPU a better choice?
The GPU demos of Bullet run way more smoothly than the CPU ones,however I don't know if the GPU ones could be used over a network for example.I mean if players have physics running on their GPUs,the server has to have a massive GPU farm to perform the same physics calculations for the <amount of online users,which can be from thousands to a million> and then send it to them,which breaks the standart of just stacking Xeons on the servers,making things more complicated.What is your opinion on Bullet's use in online games?
-
MaxDZ8
- Posts: 149
- Joined: Fri Jun 24, 2011 8:53 am
Re: Is CPU physics worth it or is moving to GPU a better cho
GPU mode doesn't change anything.
When it comes to result validation, CPU FPUs have enough divergence to make up for a different bitstream.
When it comes to the amount of data transferred, I'd say your scenario isn't currently very likely.
First of all, not many games have "thousands to a million players". Those which do are typically distributed over a network.
Even if all the players are on the same server - which is unlikely - nobody says the physics must be pushed for every particle or debris.
Servers are often not concerned with some detail of the physics process.
For example, say you blow up a pillar in a game. The client-side of things is that the pillar spawns debris falling down.
The server-side of thing is: update collision model to (maybe) precomputed model: damaged pillar[14].
But even supposing we need to be extra super accurate (say we want to allow thin boxes to run everything) I see no problem in multiprocessing or stacking GPUs in servers as well. They're starting to be there.
Client-side effects are the key.
When it comes to result validation, CPU FPUs have enough divergence to make up for a different bitstream.
When it comes to the amount of data transferred, I'd say your scenario isn't currently very likely.
First of all, not many games have "thousands to a million players". Those which do are typically distributed over a network.
Even if all the players are on the same server - which is unlikely - nobody says the physics must be pushed for every particle or debris.
Servers are often not concerned with some detail of the physics process.
For example, say you blow up a pillar in a game. The client-side of things is that the pillar spawns debris falling down.
The server-side of thing is: update collision model to (maybe) precomputed model: damaged pillar[14].
But even supposing we need to be extra super accurate (say we want to allow thin boxes to run everything) I see no problem in multiprocessing or stacking GPUs in servers as well. They're starting to be there.
Client-side effects are the key.