I take that to mean the 1284P has 2 8 bit timers and 2 16 bit times, where as the non-P has two 8 bit timers but only one 16 bit timer.
You disagree, if so why?
because i have both chips and assure you both have a timer3. unbelievers will find the code below blinks any leds attached to port d.
reset: ldi r23,0xff
l1: lds r22,tcnt3l
atmel data sheets are rife with misinformation. there was an almost identical situation with the 328 couple years back. another interesting but related issue is they use the same die for supposedly radically different parts. ie many of the low end 6 pin are actually relabeled higher end parts. in that case i dont complain because i get 2x or 4x the memory for a few cents less.
In the early days of microprocessors, 8080 and 6800, there were unused opcodes. That is, there were combinations of 8 bit codes that were not documented. People began exploring them, testing what they did and started using the undocumented opcodes. They was a lot of supposition that there were problems with the unused opcodes and the manufacterer just did not want to support them. Later processors locked out the unused opcodes so that they were just noops.
I suspect that the non-P timer3 may have a problem with the timer. Rather than make a new die masks, they just removed it from the spec. An engineer would be foolish to use an unsupported feature. A hobbyist can do anything they want.
Regarding the extra memory. The manufacturer probably had extra of the high end parts and needed low end parts. Rather than firing up the fab line for those parts, they took the higher end parts and just branded them with a lower end part number.