It may have helped if the OP had mentioned system spec, seems they are doing huge renders on a 112GB system.
The absolute limit of pagefile size on either x86 with PAE or x64 is 16TB, a 4GB limit occurs if x86 is run in non-PAE mode.
The 1.5x RAM filesize came about due to the pagefile being the only location that could be known to be still locatable for a full memory dump if Windows crashed. This is because Windows maintains continuous record of the absolute sector addresses for the pagefile. Other files are managed by the file system, which in the case of a crash may be compromised, e.g. NTFS.SYS may have stopped running. A full memory dump requires RAM size plus an overhead, so RAM x 1.5 was used as a convenient way to make sure extra was there, and never changed.
piercazzo, your situation is well beyond my experience. Mulling it over again, my first thought was whether at those sizes the allocations had to be made in powers of 2, and the total array size was just under 4TB, so 2TB became the limit. That wouldn't seem to explain the 1TB occurrence though. The only other possibility I can come up with so far is maybe there is a clash between something in the RAID array, such as block allocation, and segment alignment in the pagefile? Is there any way you can get the use of a single 4TB drive to see if it behaves differently to the RAID? Again, that may not help if the problem is because the 4TB size has to be a full 4TB and the drive is just under (as they usually are).