Battery, Thermals & Scheduling

The energy accounting for an inference inverts intuition: an int8 multiply-accumulate costs picojoules; fetching its operands from DRAM costs 100–1000× more. The battery cost of a model is dominated by memory access, not arithmetic.

This explains the levers: quantization saves energy by shrinking bytes-moved; batching amortizes a weight-read across many inputs (energy per inference falls); keeping tensors in SRAM-resident tiles avoids DRAM. When you optimize an edge model for energy, you're optimizing a memory-movement budget — the compute is nearly free.