Both uses cublas under the hood. So I think it is similar for prefilling (of course, this framework is too early and don't have FP16 / BF16 support for GEMM it seems). Hand-roll gemv is faster for token generation hence llama.cpp is better.