So, here's a data point. On my 4-core MacBookPro (Intel Core i7 3,1GHz, 16GB RAM), flambda is slower than vanilla on coq-stdpp, which does not use native-compute AFAIK (and comes from many of the same authors as lambda-rust).
To install this coq-stdpp version, you'll have to run first:
opam repo add iris-dev git+https://gitlab.mpi-sws.org/iris/opam.git
With ocaml-base-compiler.4.07.1:
$ $(which time) opam install -y coq-stdpp.dev.2019-09-19.1.9041e6d8
61.50 real 105.10 user 12.38 sys 61.53 real 105.35 user 12.16 sys 61.00 real 106.94 user 12.78 sys 61.74 real 106.84 user 13.91 sys 64.05 real 110.46 user 14.66 sys 63.41 real 109.63 user 14.32 sys 62.06 real 106.74 user 13.93 sys 62.51 real 108.20 user 14.19 sys
With ocaml-variants.4.07.1+flambda+no-flat-float-array:
$ $(which time) opam install -y coq-stdpp.dev.2019-09-19.1.9041e6d8
68.63 real 134.18 user 12.44 sys 68.56 real 132.68 user 12.29 sys 68.16 real 133.00 user 12.42 sys 70.29 real 134.81 user 14.28 sys 69.39 real 133.14 user 14.07 sys
The slowdown in user time is here around ~25% (flambda/vanilla - 1).