Skip to content

Latest commit

 

History

History
84 lines (72 loc) · 3.32 KB

README.md

File metadata and controls

84 lines (72 loc) · 3.32 KB
Raw

TensorFlow Lite on VisionFive 2 Benchmarks

Environment

HW

  • StarFive VisionFive 2

OS

  • 202302 Release Linux geerling-image-69-vf2 5.15.0-starfive #1 SMP Mon Feb 27 14:03:14 EST 2023 riscv64 GNU/Linux

SW

  • TensorFlow Lite v2.12.0 (build from source)
    build option 
      cmake \
    -DTFLITE_ENABLE_XNNPACK=ON \
    -DTFLITE_ENABLE_GPU=ON \
    -DXNNPACK_ENABLE_RISCV_VECTOR=OFF

How to benchmarks

  • XNNPACK delegate 
    benchmark_model \
  --graph=_PATH_TO_TF_LITE_MODEL_FILE_ \
  --use_xnnpack=true \
  --num_threads=`N` (number of threads)
  • GPU delegate 
    benchmark_model \
  --graph=_PATH_TO_TF_LITE_MODEL_FILE_ \
  --use_gpu=true

Models

Results

Latency mean (ms)

XNNPACK delegate

Model Input Num threads FP32 INT8
EfficientDet-lite0 320x320 1 1800.89 10311.40
2 967.87 5286.31
3 696.26 3848.13
4 551.99 2926.33
EfficientDet-lite1 384x384 1 3537.45 20653.00
2 1914.71 10537.50
3 1313.44 7390.17
4 1035.22 5676.72
EfficientDet-lite2 448x448 1 5904.75 35009.50
2 3132.86 17824.50
3 2202.07 12737.20
4 1729.77 9450.72
EfficientDet-lite3 512x512 1 13194.90 76227.10
2 6861.08 38823.50
3 4791.45 26623.60
4 3798.13 20067.00
EfficientDet-lite3x 640x640 1 22824.20 141339.00
2 11799.90 71824.70
3 8341.26 49117.90
4 6369.91 36906.40
EfficientDet-lite4 640x640 1 32395.60 202239.00
2 16824.10 102514.00
3 11574.00 70122.80
4 8932.28 52694.40

GPU delegate

Model Input FP32 INT8
EfficientDet-lite0 320x320 436.08 451.45
EfficientDet-lite1 384x384 712.76 716.21
EfficientDet-lite2 448x448 1073.37 1095.76
EfficientDet-lite3 512x512 1817.58 1873.07
EfficientDet-lite3x 640x640 3280.28 3364.11
EfficientDet-lite4 640x640 4384.98 4407.78