Ada Audits the Offload Break-Even

Ada keeps bytes, radio charge, setup, and tail in one ledger before deciding where the workload belongs

foundations
math-foundations
energy
offloading
intermediate
Ada ADA · CALCULATION AUDIT

Ada Audits the Offload Break-Even

The chapter gives a node a 6 mA radio on a 50 kbps link and a local inference that costs 30 mA-s, then asks when it is cheaper to send data than to compute it. With no setup overhead the crossover lands near 31,250 bytes, about 30 KB; add an 8 mA-s connection setup and it drops to about 22 KB. This audit works the offload break-even to the byte, so a placement decision survives a battery review.

Companion to the chapter Compute Offloading and Placement — every number here comes from that chapter.

Offloading is not a cloud magic trick; it is an energy equation. Keep local compute, payload bytes, setup charge, and radio tail in the same ledger before deciding where the workload belongs.

1. The BLE example derives its per-byte radio charge from current and throughput

The chapter gives a 6 mA active radio and an effective 50 kbps link, so one byte costs:

e_byte = 6 mA x (8 bit / 50000 bit/s) = 0.00096 mA-s per byte

2. The no-setup break-even is about 31,250 bytes

The local inference costs 15 mA for 2 seconds, or 30 mA-s. Dividing by the per-byte radio charge gives:

D* = 30 mA-s / 0.00096 mA-s per byte = 31250 bytes, about 30 KB

3. Setup overhead moves the decision boundary downward

If the link spends 8 mA-s before payload transfer, only 22 mA-s remains for bytes at the break-even point:

D* = (30 - 8) / 0.00096 = 22916.7 bytes, about 22 KB
Review item Arithmetic shown Decision signal
2 KB feature vector 2048 x 0.00096 = 1.97 mA-s Offload can beat the 30 mA-s local run before fixed overheads dominate.
80 KB raw input 81920 x 0.00096 = 78.64 mA-s Compute locally because radio charge alone is already more than 2.6x the local compute charge.
Cellular tail example 25 + (120 x 0.4) + (40 x 8) = 393 mA-s Setup and tail flip the payload-only estimate against offloading when local compute is 90 mA-s.

What the audit buys you: a placement decision that survives a battery review. Payload size, throughput, and radio state timing are separate measured terms; round only after the comparison is complete.

Every number above is taken from this chapter's own worked example and re-derived step by step.