Compute or Communicate? A Break-Even Criterion for Splitting Wearable AI Between Microcontrollers and Body-Coupled Links

Motivation

Wearables offload most of a modern neural network because they lack the memory and energy to run it. The largest of the 19 microcontrollers we evaluate has 4 MB of memory, whereas recent wearable foundation models exceed 100 million parameters. Most of the network therefore runs on a phone or an on-body hub.

The common design response is to compute part of the network on the wearable so that it transmits fewer bits. The reason is the cost of radios. Chatterjee et al. estimate that transmitting a bit costs about 10⁴ times the energy of switching a bit in a 45–65 nm chip, and that BLE-class radios draw two to three orders of magnitude more power than sensing and computation [1]. Splitting a network and transmitting an intermediate feature map applies the same in-sensor analytics principle.

However, computation on a commercial microcontroller is not free: it costs tens to hundreds of picojoules per MAC, far more than switching a single bit. A split therefore saves energy only when the transmission it removes costs more than the computation it adds, and this comparison depends on the link. Figure 2 shows it for one network on one microcontroller.

Expressed in MACs, one bit transmitted over BLE costs the STM32U385 as much energy as 135 MACs, and one bit over Wi-R as much as 9, including the MCU time spent transmitting. SleepFM's tokenizer spends 89 MACs per bit it removes. The tokenizer therefore saves energy over BLE but not over Wi-R. The next section formalizes this comparison.

Break-even condition

We split the network after layer k. The wearable then computes C_k MACs and transmits B_k bits instead of the raw B₀ bits. The split reduces energy when the following holds.

Interactive explorer

The explorer applies the break-even condition to any of the 114 network–MCU pairs. Select a network, a microcontroller and the BLE keep-alive interval. The two panels evaluate BLE and Wi-R with the full model, including sleep and link maintenance between inferences. The chart below them evaluates an arbitrary link from its energy per bit.

Results across six networks and 19 microcontrollers

Across all 114 network–MCU pairs, on-wearable computation minimizes energy in 11 pairs over BLE and in 0 over Wi-R. We define a network's crossover as the energy per bit above which computing first reduces energy. Only SleepFM and EnsembleSleepNet, which compress their data in early layers, have crossovers near BLE's energy per bit. Every crossover lies above Wi-R's.

The best cut spends the fewest MACs per bit removed (1/ρ). Crossover = (e_MAC − P_sl/Θ) / ρ on the STM32U385, where BLE costs 10.9 and Wi-R 0.73 nJ/bit; † marks cuts that exceed its memory. The last two columns count the MCUs, of 19, on which computing first minimizes energy; each MCU uses its own link energy.

Robustness

We test each main result against parameter error and against the modeling choices most likely to change it.

Results from the model's parameter sweeps (robustness.py, sweeps.js). The sweeps test parameter values, not the structure of the model.

Direction of each modeling assumption

Assumptions

All numbers are model estimates based on datasheets, Ixana's public power calculator [2] (screenshot below) and MLPerf Tiny benchmarks; we have not measured them on hardware. Ixana makes Wi-R, and one of the authors (S. Sen) founded it.

The numbers apply to the parameter values below and are not fixed properties of BLE, Wi-R or any MCU; the calculator recomputes them for any setting. The method carries over to other designs: compare the bits a split removes per MAC with the MCU's energy per MAC and the link's energy per bit.

Source of the radio values · Ixana Wi-R Power Profiler, default view · 20 Oct 2026

Every radio current in our model appears in this calculator. We do not use its energy-per-bit, battery-life or ratio outputs, which follow its own default settings.

1 2 3 4 5 6 Not used

	In the calculator	In our model
1	BLE: typical nRF52-class SoC	The external BLE radio for MCUs without one, such as the STM32U385. MCUs with built-in BLE use their own datasheet currents.
2	BLE sleep current, 1.9 µA	Radio sleep current, 1.9 µA
3	BLE peak current, 9000 µA at +8 dBm	The same transmit-current table. We transmit at 0 dBm, where it gives 4.6 mA.
4	Wi-R data: YR23 datasheet v1.3	Ixana YR23, normal mode (2.2 V I/O)
5	Wi-R sleep current, 2.6 µA	Deep sleep, 3.08 µW (2.3 µA at the 1.1 V core + 0.25 µA at the 2.2 V I/O)
6	Wi-R peak current, 1127 µA	Transmit, 2.32 mW (0.146 mA at 1.1 V + 0.981 mA at 2.2 V)
	Average current, energy per bit, battery life, “15.0× lower current” and “65.3× lower energy”	Not used. They follow the calculator's default settings: +8 dBm, 128-byte packets and one packet every 50 ms. At our settings, BLE costs 10.9 and Wi-R 0.73 nJ per sent bit (SleepFM on the STM32U385).

Code and data

• paper_numbers.json: every number in the paper, with plain names, units and meanings
• results.json: every split for every MCU and setting, at full precision
• cost_model.js: the energy model, the same code this page runs
• reproduce.js and robustness.py: rerun with node analysis/reproduce.js and python3 analysis/robustness.py
• model_profiles.json: per-layer data size, MACs and memory of the six networks

Cite

@misc{roychowdhury2026compute,
  title  = {Compute or Communicate? A Break-Even Criterion for Splitting Wearable AI
            Between Microcontrollers and Body-Coupled Links},
  author = {Roy Chowdhury, Meghna and Datta, Gourav and Sen, Shreyas},
  year   = {2026},
  note   = {Manuscript under review}
}

References

1. B. Chatterjee, P. Mohseni and S. Sen, “Bioelectronic Sensor Nodes for the Internet of Bodies,” Annual Review of Biomedical Engineering, vol. 25, 2023. doi:10.1146/annurev-bioeng-110220-112448
2. Ixana, “Wi-R Power Profiler and Tools,” 2026; calculator based on the YR23 datasheet v1.3. https://ixana.ai/resources/tools. Accessed 6 Oct 2026; rechecked 21 Oct 2026. Screenshot of the default view, 20 Oct 2026.