The main approach nowadays is shadow mapping. For pointlight shadows, you can render each face separately into faces of a cubemap texture or texture atlas like you said.

On hardware that can use geometry shaders, you can use the geometry shader to export the SVRenderTargetArrayIndex value that decides inside the sahder which cubemap face you will render into (or SVViewportArrayIndex and decide which viewport to render into). In the most naive approach, you can expand the primitives inside the geometry shader six times.

The better approach is to use instancing, and only render the mesh with so many instances that correspond the faces it will be visible inside. The newest GPUs are capable to export the SVRenderTargetArrayIndex and SV_ViewportArrayIndex from the vertex shader stage, so you don't even have to use the geometry shader for this.

You will also want to determine which cubemap faces (frustums) are visible from the camera, so you will want to cull those too.

So to recap, the state of the art approach of cube shadow maps can:

• render a mesh in a single pass into all visible faces
• using instancing, it will determine which faces the mesh will be visible in (instance buffer contains view-projection matrix index)
• using CPU-side frustum-frustum intersection check, determine which cubemap faces are visible from the main camera
• don't use geometry shader if vertex shader supports writing to SVRenderTargetArrayIndex or SV_ViewportArrayIndex