Hi,

That will mean coding work, right now there are still places using
fixed allocated tables per thread. With 8 threads max the memory
overhead is acceptable, on the edge.
For more threads I would really make it dynamic allocated tables.

And before you start doing that, realize that each thread in Blender
renders an entire tile. That also means it has to allocate memory for
the tiles, so having 256 threads will give a lot of overhead extra.

And... not to mention the mutexes in code, which will halt all 256
threads with a bit bad luck.