This library claims it achieved 2.5x speeds by implementing a custom BSON decoder:
https://deepkit.io/library/orm (the benchs are a little bit below)

The code for the Mongo client is here: https://github.com/deepkit/deepkit-framework/tree/master/packages/mongo/src/client and it is tightly coupled to the framework.

BSON Package which does the magic is here:
https://github.com/deepkit/deepkit-framework/tree/master/packages/bson 

So their claim is 10x improvement over the official implementation. Is this something that can be integrated in the official mongo drivers or apply the same technique for parsing/serialising ?