Using voxel hierarchies as a generic 3D scene representation makes ray marching, antialiasing, and LOD easy. The drawback is the huge amount of memory required to store voxels, even with empty space compression. Still, GigaVoxels showed that by using a ray-guided cache to produce and store only visible voxels bricks on demand, it is possible to walk through very large and detailed worlds with real-time performance in bounded GPU memory. However, on-demand production of data during rendering is still challenging in terms of synchronization and starvation of GPU cores. We propose a new GPU-driven algorithm using dynamic parallelism (DP) to minimize these, and a "GPU-cores timeline" profiling tool to analyze them. We validate our model with timings (2× gain) and we illustrate it on various scenes.