View Source Benchmarks on MacBook Air M3

Elixir YOLO

EPS: coreml

Ultralytics YOLO v11 nano - yolo11n

> mix run benchmarks/ultralytics_yolo.exs coreml models/yolo11n.onnx

Operating System: macOS
CPU Information: Apple M3
Number of Available Cores: 8
Available memory: 16 GB
Elixir 1.18.3
Erlang 27.3.3
JIT enabled: true

| Name | ips | average | median | 99th % | --- | --- | --- | --- | | preprocess | 975.38 | 1.03 ms | 0.99 ms | 1.38 ms | postprocess | 276.86 | 3.61 ms |3.56 ms | 3.98 ms | run | 64.07 | 15.61 ms | 15.63 ms | 17.66 ms

Around 20ms from the frame to the detections, which means ~50 FPS, not bad!

Ultralytics YOLO v11 medium - yolo11m

> mix run benchmarks/ultralytics_yolo.exs coreml models/yolo11m.onnx

Operating System: macOS
CPU Information: Apple M3
Number of Available Cores: 8
Available memory: 16 GB
Elixir 1.18.3
Erlang 27.3.3
JIT enabled: true

| Name | ips | average | median | 99th % | --- | --- | --- | --- | | preprocess | 892.35 | 1.12 ms | 1.09 ms | 1.67 ms | postprocess | 265.45 | 3.77 ms | 3.70 ms | 4.36 ms | run | 25.20 | 39.68 ms | 39.68 ms 43.92 ms

Around 22FPS.

Python Ultralytics

Let's see how performs, on the same machine, what I consider the benchmark - the official Ultralytics library:

Ultralytics YOLO v11 nano - yolo11n

> python benchmarks/ultralytics_yolo.py models/yolo11n.pt cpu

0: 384x640 17 persons, 4 bicycles, 7 cars, 1 truck, 1 traffic light, 25.8ms
Speed: 1.2ms preprocess, 25.8ms inference, 0.5ms postprocess per image at shape (1, 3, 384, 640)

Pre and post-processing are faster (not by much) and inference is definitely slower than the elixir version, probably due to Ortex coreml acceleration.

When running the benchmark with mps acceleration (if you have a Mac with Apple Silicon), inference is much faster (even than Elixir YOLO with ONNX model), but pre and post-processing are slower.

> python benchmarks/ultralytics_yolo.py models/yolo11n.pt mps

0: 384x640 17 persons, 4 bicycles, 7 cars, 1 truck, 1 traffic light, 9.2ms
Speed: 1.4ms preprocess, 9.2ms inference, 5.3ms postprocess per image at shape (1, 3, 384, 640)

Ultralytics YOLO v11 medium - yolo11m

> python benchmarks/ultralytics_yolo.py models/yolo11m.pt mps

0: 384x640 16 persons, 7 bicycles, 10 cars, 1 truck, 4 traffic lights, 3 backpacks, 1 handbag, 23.3ms
Speed: 1.5ms preprocess, 23.3ms inference, 11.8ms postprocess per image at shape (1, 3, 384, 640)

← Previous Page Benchmarks