← Back to the chapter · Previous detail

Technical detail · Energy Cost of Common Operations

Calculate the local-processing break-even

Now connect that hierarchy to an illustrative radio cost of 10 nJ per transmitted bit:

10 nJ/bit=10,000 pJ/bit.10\ \text{nJ/bit}=10{,}000\ \text{pJ/bit}.

At that radio cost, suppressing one transmitted bit creates room for roughly 10,000/640=15.610{,}000/640=15.6 DRAM-word reads, or many thousands of simple integer operations. That does not mean every compression pass wins. The decision must include the reads and writes used by the algorithm itself:

Elocal=NopEop+NSRAMESRAM+NDRAMEDRAM,E_{local}=N_{op}E_{op}+N_{SRAM}E_{SRAM}+N_{DRAM}E_{DRAM}, Esaved=ΔB Eradio/bit−Elocal.E_{saved}=\Delta B\,E_{radio/bit}-E_{local}.

Local processing is an energy win only when Esaved>0E_{saved}>0 and the extra active time, latency, accuracy loss, and memory footprint remain acceptable. For example, a filter that performs 2,000 integer adds at 0.1 pJ, makes 100 SRAM reads at 5 pJ, and avoids 80 radio bits at 10 nJ/bit has an idealized compute-and-memory cost of only 0.2+0.5=0.70.2+0.5=0.7 nJ against 800 nJ of avoided transmission. If the same filter spills 2,000 words to DRAM, data movement becomes 2,000×6402{,}000\times640 pJ =1.28 μ=1.28\ \muJ and reverses the result.

Use this as the review sequence: count bits avoided; count every memory level crossed; count arithmetic at its actual width; add radio startup, listening, acknowledgement, retry, and tail energy; then confirm the result with a whole-cycle trace. The numerical hierarchy tells you where to look. The measured trace decides whether the optimization survives contact with the real platform.

← Back to the chapter · Previous detail