That’s not the limit on entities, it’s a limit on addressable memory. You would have a maximum of four-point-two billion bytes of addressable memory.
Let’s say you want to load some voxel data… you previously mentioned that you will be using an octree where each node stores eight pointers to children plus information for color and each leaf stores a reference to block data. Let’s say that is 36 bytes per node (eight 32-bit pointers plus 32 extra bits to hold either the block reference or color information = 9x32 bits = 9x4 bytes = 36 bytes).
Now you want to render 500m^3 of the world. Unless you want your rendering routine to choke, that data needs to be mapped to memory, right? Well each voxel is 0.25m^3. You would need 2000^3 = 8 billion voxels just to cover the leaf nodes.
(The nearest size that actually fits an octree would be one with twelve levels, i.e. 9,817,068,105 nodes total.)
Details
At what depth level will an octree first add at least 8,000,000,000 nodes?
How many nodes will an octree have by that depth?
1. 1 node
2. 1 + 8 = 9 nodes
3. 9 + 64 = 73 nodes
4. 73 + 512 = 585 nodes
5. 585 + 4,096 = 4,681 nodes
6. 4,681 + 32,768 = 37,449 nodes
7. 37,449 + 262,144 = 299,593 nodes
8. 299,593 + 2097152 = 2,396,745 nodes
9. 2,396,745 + 16,777,216 = 19,173,961 nodes
10. 19,173,961 + 134,217,728 = 153,391,689 nodes
11. 153,391,689 + 1,073,741,824 = 1,227,133,513 nodes
12. 1,227,133,513 + 8,589,934,592 = 9,817,068,105 nodes
An octree will first add at least 8,000,000,000 nodes at depth level 12.
By then the octree will contain 9,817,068,105 nodes.
4 gigabytes is only enough room for 119,304,647 voxels at 36 bytes per voxel, and that’s assuming the entire virtual address space is mapped to voxel data. That’s about enough memory to map 128m^3 worth of voxel data.
I may be wrong here, but it is my understanding that the 4GB is a limitation on virtual memory, not physical memory. So even memory-mapped I/O is subject to the limitation. You can still do normal file I/O beyond 4GB but I believe that is considerably slower. I think your render routine would choke if you tried to load the data from disk in realtime.