Compute Overhead

Skeletal tracking at 90Hz is computationally expensive. Running it constantly for a keyboard drains batteries fast.

NPU Allocation

Modern headsets offload ML inferencing to dedicated Neural Processing Units (NPUs). However, the NPU budget is shared with environment meshing, eye tracking, and face tracking. Hand tracking for high-speed typing requires priority scheduling.

Common Mistakes

  • Running inferencing on the GPU: This drastically reduces the thermal budget available for rendering the application itself.
  • Uncapped tracking rates: Hand tracking should dynamically scale its framerate based on hand velocity to save power when hands are resting.

Thermal Budget per Inference

ProcessorPower Draw (Continuous)
Snapdragon XR2 Gen 2 (NPU)~800mW
Apple M2 (Neural Engine)~500mW
Generic GPU Inference>2.5W

Interactive Tool: Battery Drain Estimator

Estimated Tracking Runtime (Hours)
4.2

Heuristic estimation assuming base device load of 10h at 0Hz tracking.

FAQ

Why can't we use standard CPU processing?

CPUs are highly inefficient for the massive matrix multiplications required by neural network inferencing. Continuous CPU tracking would drain a headset battery in under an hour.