It's own RAM. The memory bandwidth of my GPU (616 GB/s) is about 23 times higher than that of my system memory (26.5 GB/s). And I have applications that utilize 90% of that GPU memory bandwidth.
Thanks! I've written code for discrete GPUs but not integrated, and so I'm unsure what the advantages would be. Would there be lower latency for transfers, then, if not higher throughput?