At the time the roots were generated/submitted to root programs, there was no requirement to use P-384 root keys (let alone intermediates). I haven't checked the current regs, but the choice to use P-384 was intentional.
There was some discussion about the choice of algorithms back when the hierachy was still being discussed: Let's Encrypt new hierarchy plans
The (major) speed difference between P-256 and P-384 in OpenSSL comes from the fact that OpenSSL ships a highly optimized P-256 implementation, but only offers a generic plain-C implementation for P-384. Mileage with other crypto libraries may differ. A certain drop is expected everywhere though: There's more to compute, and EC H/W acceleration (where available) typically won't go beyond including P-256.
Still much better than a webserver that has to sign 4096-bit RSA 
.