Knowing the average number of operations per second on a unit is generally all that is needed to work out its average energy, once the joules per operation are known.
The Xilinx Xpower 28 nm Zynq spreadsheet models about 415 pJ per ARM clock cycle and 23 pJ per DSP multiply. BRAM reads of RAMB18x2 (36 Kbit) units take 8.5 pJ and writes about 10 percent less.
The formula earlier for 45nm, 5.0 + 1.2e-4 / 8.0 * mbits gives 5.5 pJ but twice as much for a write.
But the totals for each component drastically depend on the activity ratios and initial guesses are typically set close to worst case which is conservative, but typically wildly out. Hence SAIF-based or other dynamic trace information must be fed in for an accurate result.
| 46: (C) 2008-18, DJ Greaves, University of Cambridge, Computer Laboratory. |