It's a trade-off. The off-chip GPU will have the most memory bandwidth and the most crunching power (larger thermal envelope). The on-chip GPU will have wildly

It's a trade-off. The off-chip GPU will have the most memory bandwidth and the most crunching power (larger thermal envelope). The on-chip GPU will have wildly superior latency, lower cost, and lower power consumption. It also has access to the on-die L2 cache now, although I'm not sure what that means for bandwidth- how does CPU L2 compare to GPU L2? High-end on-chip GPU's can compete with midrange off-chip GPU's.