First picture with cascaded shadow generator, 1024x1024x4 cascades, so same overall texture size than in the second picture (for fair comparison).
Second picture with standard shadow generator, 2048x2048 texture.
I’m afraid BABYLON.RenderTargetTexture.REFRESHRATE_RENDER_ONCE can’t be used for CSM because the projection matrices (and so the depth maps) are recalculated each frame according to the camera view, so matrices (/ shadow maps) calculated (/ rendered) for frame X are generally not suitable for frame X+t, except if you keep t low so that the changes are minimal (so REFRESHRATE_RENDER_ONEVERYTWOFRAMES should work, albeit you may experience some rendering artifacts depending on your scene and movements).
If you want a one shot shadow rendering, you are better off using baked / pre-calculated shadows I think.
CSM is doing N shadow map rendering, N being the number of cascades (4 by default), so it’s expected it takes a little more time than the standard generator to render, that does a single rendering. However, we tried to optimize as much as possible, removing code from the shaders depending on the options you choose: have a look at the “Optimizing for speed” section in the documentation for detail.
Also, proper culling of the shadow casters is not implemented yet, all objects (shadow casters) are currently sent for rendering in each shadow map. So, if you have a lot of shadow casters in your scene and only a small number is relevant for a given frame, it might affect the performances.