Without cache coherence, how can we get reliable memory use out of GPU's?
Metalbird
Getting reliable memory use out of GPU's comes down to the programmer ensuring that there program can be run in such a way that you don't need to rely on cache coherence. So GPU code has to be designed such that maybe a line only goes to one cache and isn't in multiple cache lines so that there is no chance of having a coherence problem.
zale
This fits well with the data-parallel model of computation anyways, since conceptually SM cores should probably all be using different data.
Without cache coherence, how can we get reliable memory use out of GPU's?
Getting reliable memory use out of GPU's comes down to the programmer ensuring that there program can be run in such a way that you don't need to rely on cache coherence. So GPU code has to be designed such that maybe a line only goes to one cache and isn't in multiple cache lines so that there is no chance of having a coherence problem.
This fits well with the data-parallel model of computation anyways, since conceptually SM cores should probably all be using different data.