drkohler said:
Basically, it means that they studied how PC games actually run on a PC system (they probably had a state machine in their lab that analysed some modern games down to the last clock cycle). The key bottlenecks in PC gaming are a) when cpu and gpu have to talk with each other, they basically use "morse code" - incredibly slow by today's speed standards (a single, unique "command" thread has to talk to the gpu over the PCIe-bus). b) your typical gpu on a pc graphics card may show a high number of TFlops raw power on the advertisement sheet. However, up to a quite high percentage of that incredible power cannot be used at various moments, depending on what you do (because that giant TFlop number is basically calculated as "when all transistors are switching" - basically a state that doesn't exist at any time during a game). So the solution to a) is build a separate data path between cpu and gpu (apparently 20GB/s which is very fast. And it doesn't mess with the 176GB/s you have to the gddr5). A solution to b) doesn't exist, obviously, because graphics processing can't be reeinvented. However what you can do is use ALL the spare gpu time to do compute tasks like physics, for example, or preprocessing graphics (like the spus on the cell). This way the gpu operates at close to maximum raw power. |
ahh alright thanks for the explanation.
