Comparative benchmarking results

Run 10 h WAV-files on Intel i7-8565U (4-Core, 16 GB RAM, Windows 10)

  • Input: 10 × 60-minute WAV-files

  • Disk: NVMe SSD (Intel SSDPEKKF010T8L)

  • Device: Lenovo ThinkPad X1 Carbon 7th Gen

Result: 49 × real-time (RTF: 0.02047110)

$ birdnet-benchmark test-dataset/test_dataset_100x60min -p 5

-------------------------------
------ Benchmark summary ------
-------------------------------
Start time: 07/09/2025 03:47 PM
End time:   07/09/2025 04:00 PM
Wall time:  0:12:16.959720
Input: 10 file(s) (WAV)
  Total duration: 10:00:00
  Average duration: 1:00:00
  Minimum duration (single file): 1:00:00
  Maximum duration (single file): 1:00:00
Producer(s): 1
Buffer: 8.0/8 filled slots (mean)
Busy workers: 3.9/4 (mean)
  Average wait time for next batch: 0.257 ms
Memory usage:
  Program: 1596.43 M (total max)
  Buffer: 4.39 M (shared memory)
  Result: 0.34 M (NumPy)
Performance:
  49 x real-time (RTF: 0.02047110)
  16 segments/s (0:00:48.849346 audio/s)
Worker performance:
  50 x real-time (RTF: 0.01992776)

Run 100 h WAV-files on AMD Ryzen 7 3800X (8-Core, 64 GB RAM, Linux)

  • Input: 100 × 60-minute WAV-files

  • Disk: NVMe SSD (Samsung MZVLB1T0HBLR-00000)

Result: 416 × real-time (RTF: 0.00240288)

$ birdnet-benchmark test-dataset/test_dataset_100x60min -f

-------------------------------
------ Benchmark summary ------
-------------------------------
Start time: 07/09/2025 03:15 PM
End time:   07/09/2025 03:30 PM
Wall time:  0:14:25.035450
Input: 100 file(s) (WAV)
  Total duration: 4 days, 4:00:00
  Average duration: 1:00:00
  Minimum duration (single file): 1:00:00
  Maximum duration (single file): 1:00:00
Producer(s): 1
Buffer: 15.4/16 filled slots (mean)
Busy workers: 8.0/8 (mean)
  Average wait time for next batch: 0.044 ms
Memory usage:
  Program: 1849.19 M (total max)
  Buffer: 8.79 M (shared memory)
  Result: 2.94 M (NumPy)
Performance:
  416 x real-time (RTF: 0.00240288)
  139 segments/s (0:06:56.167916 audio/s)
Worker performance:
  417 x real-time (RTF: 0.00239808)

Run 100 h WAV-files on NVIDIA Titan RTX (24 GB, Linux)

  • Input: 100 × 60-minute WAV-files

  • Disk: NVMe SSD (Samsung MZVLB1T0HBLR-00000)

Result: 2702 × real-time (RTF: 0.00037009)

Disk speed:

$ sudo hdparm -tT /dev/nvme0n1
  Timing cached reads: 27324 MB in 1.99 seconds = 13707.90 MB/sec
  Timing buffered disk reads: 6880 MB in 3.00 seconds = 2293.17 MB/sec

$ birdnet-benchmark test-dataset/test_dataset_100x60min --device GPU --backend pb -w 1 -p 5 -s 1025 --prefetch-ratio 5

-------------------------------
------ Benchmark summary ------
-------------------------------
Start time: 11/05/2025 01:54 PM
End time:   11/05/2025 01:56 PM
Wall time:  0:02:13.231957
Input: 100 file(s) (WAV)
  Total duration: 4 days, 4:00:00
  Average duration: 1:00:00
  Minimum duration (single file): 1:00:00
  Maximum duration (single file): 1:00:00
Producer(s): 5
Buffer: 1.7/6 filled slots (mean)
Busy workers: 0.9/1 (mean)
  Average wait time for next batch: 0.001 ms
Memory usage:
  Program: 8543.88 M (total max)
  Buffer: 3295.93 M (shared memory)
  Result: 5.58 M (NumPy)
Performance:
  2702 x real-time (RTF: 0.00037009)
  901 segments/s (0:45:02.054441 audio/s)
Worker performance:
  2869 x real-time (RTF: 0.00034853)
  956 segments/s (0:47:49.160292 audio/s)

Run 100 h FLAC-files on NVIDIA Titan RTX (24 GB, Linux)

  • Input: 100 × 60-minute FLAC-files

  • Disk: NVMe SSD (Samsung MZVLB1T0HBLR-00000)

Result: 2487 × real-time (RTF: 0.00040214)

$ birdnet-benchmark test-dataset/test_dataset_100x60min_flac --device GPU --backend pb -w 1 -p 5 -s 1025

-------------------------------
------ Benchmark summary ------
-------------------------------
Start time: 07/09/2025 03:07 PM
End time:   07/09/2025 03:09 PM
Wall time:  0:02:24.770857
Input: 100 file(s) (FLAC)
  Total duration: 4 days, 4:00:00
  Average duration: 1:00:00
  Minimum duration (single file): 1:00:00
  Maximum duration (single file): 1:00:00
Producer(s): 5
Buffer: 1.9/2 filled slots (mean)
Busy workers: 1.0/1 (mean)
  Average wait time for next batch: 0.051 ms
Memory usage:
  Program: 9304.52 M (total max)
  Buffer: 1126.11 M (shared memory)
  Result: 2.94 M (NumPy)
Performance:
  2487 x real-time (RTF: 0.00040214)
  829 segments/s (0:41:26.688317 audio/s)
Worker performance:
  2602 x real-time (RTF: 0.00038431)