We consider the following cases.
Somewhat Homomorphic Schemes(SHS), Fully H.S. (FHS) and Partially Homomorphic Schemes (i.e. homomorphic with respect one operation, either addition or multiplication). RSA belongs to the third case, which is multiplicative homomorphic and is fast.
Remark that, the textbook-RSA is multiplicative homomorphic and so insecure, since we do not use a secure padding scheme.
A secure cryptosystem in this class is the Pallier's additive homomorphic cryptosystem.
SHS support both operations, addition and multiplication but up to some level. In these systems, beyond some level, the noise to the ciphertext increases so much as a result the decryption is not correct.
There are many SHS such as, the family based on the Learning With Errors problem. These systems are slow because they use large keys in order to get the necessary security.
FHS start with a SHS (all SHS are noisy encryption schemes). The construction of Gentry, found a way to refresh the ciphertext in order to decrease the noise (bootstrapping). The reason that make them slow is the bootstrapping step. It applies an operation to ciphertext called ${\tt Recrypt}$, which increases the complexity of the system but also manages to control the noise. So allows as many operation we want and the decryption will be correct. I suggest you to read the original paper of Gentry for the case of FHE. Also see this post.