Add RP2040 microcontroller feasibility note (measured)
Browse files
README.md
CHANGED
|
@@ -144,6 +144,10 @@ Measured 2026-10-12 on Raspberry Pi 4 Model B Rev 1.4 (4 GB, Cortex-A72 @ 1.8 GH
|
|
| 144 |
|
| 145 |
Measured 2026-10-12 on Raspberry Pi 5 Model B Rev 1.0 (8 GB, Cortex-A76 @ 2.4 GHz), Raspberry Pi OS Bookworm (aarch64), Python 3.11, ONNX Runtime 1.31.0, identical method to the Pi 4 run above but on a near-idle board (~0.6 GB RAM in use, SoC 44 °C, no throttling). At 4 threads a single face costs ~8.7 ms (about 115 fps): the Pi 5 runs this 1M model comfortably in real time, roughly 7.7x faster than the in-service Pi 4.
|
| 146 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
- **v1**: clean crops, standard augmentation. Best on well-lit photos.
|
| 148 |
- **v2**: v1 recipe + low-light augmentation (random gamma darkening, contrast crush, sensor noise). Meant for indoor webcam use. Slightly lower on the clean test set.
|
| 149 |
- INT4 gives no kernel speed-up on CPU/Pi and costs 2-4 points, so **ship INT8**.
|
|
|
|
| 144 |
|
| 145 |
Measured 2026-10-12 on Raspberry Pi 5 Model B Rev 1.0 (8 GB, Cortex-A76 @ 2.4 GHz), Raspberry Pi OS Bookworm (aarch64), Python 3.11, ONNX Runtime 1.31.0, identical method to the Pi 4 run above but on a near-idle board (~0.6 GB RAM in use, SoC 44 °C, no throttling). At 4 threads a single face costs ~8.7 ms (about 115 fps): the Pi 5 runs this 1M model comfortably in real time, roughly 7.7x faster than the in-service Pi 4.
|
| 146 |
|
| 147 |
+
### Microcontrollers (RP2040)
|
| 148 |
+
|
| 149 |
+
This model **cannot run** on a Raspberry Pi Pico / RP2040 (264 KB SRAM) in any quantization: its largest intermediate activation alone is 312 KB INT8 — larger than the chip's entire SRAM — so no tensor-arena arrangement can hold it (ONNX shape-inference measurement, 112x112 input). The INT8 weights (~988 KB) would fit the 2 MB flash, but activations rule the chip out. For reference, a ~100K-parameter model of the same operator mix measured on a real RP2040 at 133 MHz runs INT8 inference in about 1.09 s/face using 171 KB of arena; RP2040-class deployment needs an architecture in that size class or smaller. This model targets Raspberry Pi 4/5-class CPUs and up.
|
| 150 |
+
|
| 151 |
- **v1**: clean crops, standard augmentation. Best on well-lit photos.
|
| 152 |
- **v2**: v1 recipe + low-light augmentation (random gamma darkening, contrast crush, sensor noise). Meant for indoor webcam use. Slightly lower on the clean test set.
|
| 153 |
- INT4 gives no kernel speed-up on CPU/Pi and costs 2-4 points, so **ship INT8**.
|