Skip to content

Latest commit

 

History

History
237 lines (204 loc) · 7.07 KB

README.md

File metadata and controls

237 lines (204 loc) · 7.07 KB
Raw

Snapshot SDK Benchmark

The results are obtained using benchmark_snapshot.py.

  • SDK is started with one worker per vCPU.
  • The image used is assets/car-4k.jpg resized to target resolution.
  • API is called 50 times for each configuration.
  • API calls are parallelized. The client code makes 4 calls concurrently.
git clone https://github.com/parkpow/deep-license-plate-recognition.git
cd deep-license-plate-recognition
python -m benchmark.benchmark_snapshot

Notes

  • Speed is duration / number_of_calls.
  • All numbers are in milliseconds.
  • In fast mode, number of detection steps is always 1. May result in lower accuracy when using images with small vehicles.

Google Cloud Instance - Snapshot 1.3.17

n1-standard-4 (4 vCPUs, 15 GB memory), 1 x NVIDIA Tesla T4

Mode Resolution Speed
regular 800x600 31.5
fast 800x600 21.5
regular 1280x720 43.7
fast 1280x720 23.7
regular 1920x1080 45.6
fast 1920x1080 27.0
regular 2560x1440 49.7
fast 2560x1440 29.9

n2-standard-8 (8 vCPUs, 32 GB memory)

Mode Resolution Speed
regular 800x600 27.1
fast 800x600 16.9
regular 1280x720 40.8
fast 1280x720 21.1
regular 1920x1080 50.2
fast 1920x1080 28.2
regular 2560x1440 49.9
fast 2560x1440 27.7

c2-standard-4 (4 vCPUs, 16 GB memory)

Mode Resolution Speed
regular 800x600 37.4
fast 800x600 23.8
regular 1280x720 57.7
fast 1280x720 30.0
regular 1920x1080 69.3
fast 1920x1080 40.9
regular 2560x1440 67.5
fast 2560x1440 38.7

n2-standard-4 (4 vCPUs, 16 GB memory)

Mode Resolution Speed
regular 800x600 42.0
fast 800x600 26.8
regular 1280x720 63.9
fast 1280x720 33.1
regular 1920x1080 75.7
fast 1920x1080 45.5
regular 2560x1440 74.7
fast 2560x1440 42.9

n1-standard-4 (4 vCPUs, 15 GB memory)

Mode Resolution Speed
regular 800x600 61.4
fast 800x600 37.8
regular 1280x720 93.8
fast 1280x720 47.3
regular 1920x1080 111.1
fast 1920x1080 64.4
regular 2560x1440 110.3
fast 2560x1440 59.4

n2-standard-2 (2 vCPUs, 8 GB memory)

Mode Resolution Speed
regular 800x600 72.1
fast 800x600 46.9
regular 1280x720 107.8
fast 1280x720 58.2
regular 1920x1080 136.4
fast 1920x1080 86.0
regular 2560x1440 131.3
fast 2560x1440 82.8

e2-standard-2 (2 vCPUs, 8 GB memory)

Mode Resolution Speed
regular 800x600 94.2
fast 800x600 61.0
regular 1280x720 145.0
fast 1280x720 76.8
regular 1920x1080 179.7
fast 1920x1080 108.6
regular 2560x1440 171.0
fast 2560x1440 98.8

Bare Metal - Snapshot 1.3.17

Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz

Mode Resolution Speed
regular 800x600 38.3
fast 800x600 21.8
regular 1280x720 55.3
fast 1280x720 33.0
regular 1920x1080 73.0
fast 1920x1080 37.4
regular 2560x1440 79.6
fast 2560x1440 46.7

Raspbery Pi 4 64 bits (1 worker)

Mode Resolution Speed
regular 800x600 686.7
fast 800x600 401.7
regular 1280x720 1181.0
fast 1280x720 554.0
regular 1920x1080 1373.4
fast 1920x1080 677.8
regular 2560x1440 1587.1
fast 2560x1440 816.9

Raspberry Pi 3/4 (estimates) 32 bits

Mode Resolution Speed
regular 1280x720 1300
fast 1280x720 1000

Jetson Nano (estimates)

Mode Resolution Speed
regular 1280x720 700
fast 1280x720 350

LattePanda Alpha (estimates)

Mode Resolution Speed
regular 1280x720 170
fast 1280x720 119

LattePanda V1 (estimates)

Mode Resolution Speed
regular 1280x720 1250
fast 1280x720 875

Jetson Orin Devkit - Snapshot 1.37

With 4 workers (-e WORKERS=4):

Mode Resolution Speed
regular 800x600 57.8
fast 800x600 38.7
regular 1280x720 78.2
fast 1280x720 40.3
regular 1920x1080 82.9
fast 1920x1080 43.7
regular 2560x1440 87.1
fast 2560x1440 47.9

With 2 workers (-e WORKERS=2):

Mode Resolution Speed
regular 800x600 74.1
fast 800x600 49.2
regular 1280x720 103.4
fast 1280x720 52.5
regular 1920x1080 112.1
fast 1920x1080 59.6
regular 2560x1440 122.4
fast 2560x1440 69.2

With 1 worker (-e WORKERS=1):

Mode Resolution Speed
regular 800x600 119.8
fast 800x600 83.0
regular 1280x720 170.5
fast 1280x720 89.8
regular 1920x1080 190.9
fast 1920x1080 108.8
regular 2560x1440 220.1
fast 2560x1440 134.3

Jetson Orin Devkit - Snapshot 1.38 (INFERENCE_ENGINE=tensorrt)

With 4 workers (-e WORKERS=4):

Mode Resolution Speed
regular 800x600 36.5
fast 800x600 29.2
regular 1280x720 42.1
fast 1280x720 29.9
regular 1920x1080 50.7
fast 1920x1080 38.1
regular 2560x1440 57.5
fast 2560x1440 43.5

With 2 workers (-e WORKERS=2):

Mode Resolution Speed
regular 800x600 47.2
fast 800x600 36.7
regular 1280x720 57.2
fast 1280x720 41.2
regular 1920x1080 70.3
fast 1920x1080 53.7
regular 2560x1440 93.8
fast 2560x1440 64.1

With 1 worker (-e WORKERS=1):

Mode Resolution Speed
regular 800x600 73.2
fast 800x600 59.7
regular 1280x720 90.2
fast 1280x720 69.2
regular 1920x1080 113.6
fast 1920x1080 89.1
regular 2560x1440 153.3
fast 2560x1440 110.8