Commit ·
4d0fd01
1
Parent(s): 03517cc
Add measured Galaxy S26 NPU/GPU encoder rows (#3)
Browse files- Add measured Galaxy S26 NPU/GPU encoder rows (31a227a09f621459e8c30338cee6bfa1022dd39f)
Co-authored-by: Daisuke Majima (MLBoy) <mlboydaisuke@users.noreply.huggingface.co>
README.md
CHANGED
|
@@ -210,6 +210,22 @@ model reproduces the f32 token sequence exactly on the dense test window on
|
|
| 210 |
both Apple devices, and its transcripts are identical between the M4 Max
|
| 211 |
and the Pi 5 on all 12 test clips.
|
| 212 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 213 |
### Accuracy note
|
| 214 |
|
| 215 |
In a 12-clip spot check (LibriSpeech dev-clean samples plus two
|
|
|
|
| 210 |
both Apple devices, and its transcripts are identical between the M4 Max
|
| 211 |
and the Pi 5 on all 12 test clips.
|
| 212 |
|
| 213 |
+
### Snapdragon NPU / GPU — Galaxy S26
|
| 214 |
+
|
| 215 |
+
Measured on a physical Samsung Galaxy S26 (SM-S942Q, Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16) with LiteRT `CompiledModel` 2.2.0 — 5 warm-up runs then 50 timed runs, one accelerator per process, every row at device thermal status `NONE`, delegate placement confirmed from logcat per row. The figures are the **`encode` signature only** (the model's first signature, which the API's default `run()` executes — the encoder over the raw 80,000-sample / 5 s waveform input); not comparable to the end-to-end window totals in the CPU table above. NPU in HTP BURST mode, on-device (JIT) compilation.
|
| 216 |
+
|
| 217 |
+
| File | Compute unit | Encode (median / min) | Load |
|
| 218 |
+
|---|---|---|---|
|
| 219 |
+
| `moonshine_tiny_5s_f32.tflite` | **GPU (Adreno)** | **5.4 ms** / 5.4 ms | 1.7 s |
|
| 220 |
+
| `moonshine_tiny_5s_i8.tflite` | NPU (Hexagon, JIT) — first launch | 1294 ms / 1281 ms | 13.1 s |
|
| 221 |
+
| `moonshine_tiny_5s_i8.tflite` | NPU (Hexagon, JIT) — cached | 1264 ms / 1247 ms | 0.37 s |
|
| 222 |
+
|
| 223 |
+
What the table says:
|
| 224 |
+
|
| 225 |
+
- **On this device the GPU running the f32 file is the only fast accelerated path**: 5.4 ms for the 5 s window's encoder, encoder-only RTF 0.001.
|
| 226 |
+
- **The Hexagon runs the i8 encoder in ~1.26 s** — slower than the Pi 5's CPU in the table above. The two accelerators also refuse each other's file: f32 compiles for the NPU but fails at output-buffer creation, and i8 fails to compile on the GPU. So there is no NPU recommendation to make from this run.
|
| 227 |
+
- `moonshine_tiny_5s_f32_Qualcomm_SM8850.tflite` (the embedded precompiled context) failed to invoke on both accelerators in this harness and is not quoted.
|
| 228 |
+
|
| 229 |
### Accuracy note
|
| 230 |
|
| 231 |
In a 12-clip spot check (LibriSpeech dev-clean samples plus two
|