In doing some research I found that in addition to directory-based and snooping there is a third, less popular approach to cache coherence called snarfing. In snarfing, cache controllers watch addresses and data simultaneously to update their own copies when they are modified in main memory. Seems slow, but involves less internode communication.
Split_Personality_Computer
I understand how this could work for a 2 or 4 core CPU, but do GPUs acutally do this? Or do they just read and write straight from memory?
gogogo
I'd imagine that GPUs would do the same thing. I don't see why GPUs would operate significantly different from CPUs - they have similar functions.
anamdev
We learned in last lecture that GPUs don't implement cache coherency, which is why I'm inclined to think they read straight from memory.
In doing some research I found that in addition to directory-based and snooping there is a third, less popular approach to cache coherence called snarfing. In snarfing, cache controllers watch addresses and data simultaneously to update their own copies when they are modified in main memory. Seems slow, but involves less internode communication.
I understand how this could work for a 2 or 4 core CPU, but do GPUs acutally do this? Or do they just read and write straight from memory?
I'd imagine that GPUs would do the same thing. I don't see why GPUs would operate significantly different from CPUs - they have similar functions.
We learned in last lecture that GPUs don't implement cache coherency, which is why I'm inclined to think they read straight from memory.