I got some confusion about the number of arithmetic units on Nvidia GPU. Focusing on Tesla, my understanding is like this:
For C1060, one MP (which has 8 SPs) has one double-precision unit and 2 special function units (SF). Then each core or each SP has one integer/floating-point unit and some special ALUs. This integer/floating-point unit can support only one integer operation or one floating-point operation at the same time. It cannot support both. By the way, is this floating-point unit single-precision or double-precision?
For C2050 or Fermi, one MP has 4 SFs. Then one SP or one core has one integer unit, one single-precision unit and one double-precision unit.
Is this correct?