Your data has to be a scalar field on a 3D grid for this to be offered at all. Image data and rectilinear grids work directly; unstructured meshes are supported but slower, and a common move is to resample onto a uniform grid first with Resample To Image.

The transfer function maps scalar values to a colour and an opacity. Colour is mostly taste. Opacity is the entire result: leave everything semi-opaque and you get fog, make everything sharp and you have reinvented the isosurface badly. What you want is a narrow band of opacity around the values you care about, and near-zero everywhere else.

Look at the histogram behind the curve in the Color Map Editor before touching anything. If most of your cells sit in a narrow spike, a linear opacity ramp will be dominated by that spike and you will see nothing else. Drag the control points so the ramp lives where the interesting values are, not where the bulk of the data is.

Shrink the data first — Extract Subset or a coarser resample beats every rendering option. Then raise the Scalar Opacity Unit Distance, which is the sampling step: larger is faster and grainier. Only then worry about shading, which adds depth cues but costs a normal calculation at every sample. On a rented machine with a GPU, the card does the raycasting, so the thing that actually limits you is usually how much of the dataset fits in its memory.