1 of 1 people found this helpful
yes it is more explicit form of zero copy. when you create normal buffer you can't be 100% sure if it is shared. with SVM you know that it is shared between host and device. another difference is that pointers can be shared between device and host.
My understanding is that with course grain SVM, they allocate a buffer on the GPU and a buffer on the CPU at the same virtual address. i.e. so the GPU and CPU see it at them at same addresses and all the pointers within it point at the correct spot.
When you do a sync it copies the relevant buffer across.
With zero copy the GPU reads directly from the CPU's memory. It maybe that course grain SVM uses a zero copy buffer by then it may not.
It may depend on whether the GPU is discrete or not. i.e. with an APU use a zero copy buffer, with a dGpu use separate buffers and copy them back and forth.
(me, I want to see a dGPU connected via Hyper Transport... or a really big APU in a multi socked motherboard.)