GPU Skin Cache's efficiency

With 300 characters your bottlenecks are more likely to be updating the actual animations than skinning the geometry. You should profile to identify what is costing performance:

Also, compute shader skinning still does the same calculations as the vertex shader skinning. The only advantage is that the results are cached for re-use on different passes in the same frame (shadows, planar reflections, scene capture) and they can use async compute on consoles.