I suggest you see this presentation on 1GB blocks (tested on test-net) by Peter Rizun; https://www.youtube.com/watch?v=5SJm2ep3X_M And his talk at "Satoshi's Vision" (here he talks about what is called "weak blocks" and how it can improve scaling and wasted PoW; https://www.youtube.com/watch?v=yXFuNkaYcPQ
It's totally feasible and does not require super computers albeit a Raspberry Pi won't do no more. Scaling to VISA level of transactions is possible. I don't think it's really going to change too much in computing power with bigger blocks. The size of the merkle root won't change just because blocks are larger which means the block header size won't change.
Why do you suspect a second-layer is needed?