Here is a gist with benchmarks for one polynomial in 2 and one with 3 variables.
It is using the new MultivariatePolynomials API from @blegat but I also checked against Master with the same results.
You can see that the evaluation time and allocations explode if we go from 2 to 3 variables, whereas using DynamicPolynomials we only have a slight increase.
Any ideas what could causing this?