Benchmarks

The  Hailo-8™ AI Processor brings industry-leading neural processing throughput and power efficiency to support a wide range of AI applications.

 

The table below demonstrates some of Hailo-8’s best-in-class capabilities in the foundational neural tasks of object detection and classification. It achieves very high FPS and power efficiency (FPS/W) across a number of industry-standard neural network models, while also offloading postprocessing from the host processor its working with.

 

Benchmarks

 NN Model Input Resolution FPS Power [W] FPS/ W
Classification
ResNet_v1_50 224×224 1,328 3.1 428
MobileNet_v2_1.0 224×224 2,617 2.1 1246
MobileNet_v3 * 224×224 3,519 1.9 1852
Regnetx-800MF 224×224 3,507 2.7 1298
EfficientNet_M 240×240 891 3.2 278
Object Detection
SSD_MobileNet_v1 300×300 1,055 2.3 453
Tiny_YOLOv3 416×416 1,010 3.2 315
YOLOv5m 640×640 218 4.2 53
Multi stream object detection (8 streams)
YOLOv3 ** 608×608 58 4 14.5
YOLOv4 ** 512×512 83 3.5 23.4

* MobileNet_v3 – the benchmarked model flavor is MobileNet_v3 Large Minimalistic

** Figure for processing 8 simultaneous video streams

Notes:

  • Based on SDK version 3.20.0 (SW Suite version 2022-10)
  • Measurements are done at room temperature on a single Hailo-8™ device through PCIe interface on a Hailo evaluation board
  • System host: Intel® Core™ i5-9400 CPU @ 2.90GHz
  • Hailo-8™ outperforms other edge AI processors on the market on common benchmarks.

Click here and here to see how the Hailo AI Processor compares.