Performance issues are subtle, and depend a lot on implementation details, but the following can still be said.
RSA encryption and RSA decryption both use modular exponentiation. There are many algorithms for that, but for typical RSA key sizes, Montgomery multiplication in a square-and-multiply algorithm are typical (the "multiply" steps can be further reduced with window-based optimizations). As a rough approximation, time to compute a modular exponentiation with a modulus of $n$ bits and an exponent of $k$ bits will be proportional to $k·n^2$.
The CRT replace one modular exponentiation with two, but these two exponentiations use half-size modulus and exponents, so each of them is about eight times faster than the non-CRT exponentiation. Thus, CRT speeds up RSA decryption by a factor of about 4. CRT requires knowledge of modulus factorization, so it cannot be applied to encryption, only to decryption.
On the other hand, RSA encryption uses the public exponent, which can be extremely small. A traditional RSA public exponent is 65537, thus 17 bits long. Exponentiation to the power 65537, a 17-bit integer, should be about 60 times faster than exponentiation to a 1024-bit power $d$ (the private exponent). Even with the CTR speed-up, RSA encryption should still be about 15 times faster than RSA decryption.
In practice there are some extra overhead costs in both encryption and decryption (conversions to and from Montgomery representations, CRT reassembly, masking with a random value to protect against timing attacks...) so the "15x" figure can vary quite a lot. Things will also vary depending on the modulus size (you would still use 65537 as public exponent with a 2048-bit modulus, for instance). A 15x ratio is still typical.