In the 1M benchmark roughly half is spent decoding rows (in [1]). The rest is spread over interpreter overhead, GC in particular.

[1] https://github.com/MagicStack/asyncpg/blob/master/asyncpg/pr...