Practically, it can be (quite often) a limitation of number of qubits/hardware, but also it is a hyperparameter to play with. So it may be that using more qubits gives you better results or worse.
Also, in the QSVM, there is or may be a parameterized part you have to optimize over. So increasing the number of qubits results in more optimization (more parameters), that makes it harder. You may need to play on the depth of the variational part to improve results (so more parameters to optimize).
But if you are limited in the number of qubits, you can change the data encoding. For instance, in this paper Fig.2, they use a quantum circuit with $17$ qubits and loaded $67-$dimensional data without dimensionality reduction. This results in a deeper circuit.