by tnt » Thu Aug 29, 2013 4:05 pm
it seems pretty clear that you can't have a full 4096 core ... the (0,0) is always mapped to the local one and any directly mapped external memory is just accessed by the same routing N/S/W/E that any other core and so you need to reserve space for that. (by stealing some from possible cores).
Your "magical" unregister of a core just doesn't make much sense in the current chips and it would imply quite a bit of logic in each router node to know and keep track of this. Any way to grow beyond the current model would need to sacrifice the 'uniform' address space view and would most likely be implemented by translation layers between 'islands' of cores, avoiding to modify the mesh and cores themselves and just adding logic at the edge of large groups of them.