Benchmarks for EasyLocate

Test conditions

□

These numbers are only indicative and represent only the memory required for the neural network.

□

Your actual memory requirements may be bigger or lower according to your GPU model.

□

The GPU must have more memory than the indicated amount to work because storing images and results may require additional GPU memory and because of memory fragmentation.

□

The training time is approximately twice the inference time per image. An iteration is equivalent to a loop over all the images in the dataset.

□

The GPU memory requirements indicated below are approximate and can vary according to the GPU model.

- These values were obtained for a NVIDIA GeForce 3080 Ti on Windows 11.

- The GPU inference can be 10 to 50% faster on Linux for GeForce GPUs.

□

On Windows:

- When using the WDDM driver mode (always on for a GeForce GPU), the inference times can vary quite a lot.

- When using the TCC mode on a Quadro GPU, the inference times are more stable.

□

In the tables below 'n/a' means that the value could not be computed for this specific configuration (for example because there is not enough memory).

□

In the tables below, a '=' means that the value is equal to the one above it.

□

The benchmarks were obtained using EasyLocate Axis Aligned Bounding Box.

□

For EasyLocate Interest Point, the training and inference speeds are approximately the same. The small variations (a few percent slower or faster) in the processing speed depend on the parameters of the tool.

Capacity Small

Image size	Batch	Inference time / image (ms)
Image size	Batch	GPU NVIDIA GeForce 3080Ti	CPU Intel Core i9 7900X	GPU NVIDIA Jetson Xavier NX (ARM)	CPU Raspberry Pi 4 Model B
128 × 128	1	4.19	30.13	18.55	504
	4	4.44	=	7.43	=
	16	1.81	=	4.16	=
	64	0.41	=	3.45	=
256 × 256	1	4.85	84	32.02	1 959
	4	6.88	=	16.74	=
	16	1.45	=	13.51	=
	64	1.37	=	14.19	=
512 × 512	1	11.32	341	66.10	9 314
	4	5.70	=	60.85	=
	16	5.38	=	53.89	=
	64	–	=	56.28	=

Image size	Batch	GPU memory for inference (MB)	GPU memory for training (MB)
128 × 128	1	175	n/a
	4	241	354
	16	503	879
	64	1 553	2 979
256 × 256	1	241	n/a
	4	503	879
	16	1 553	2 979
	64	5 884	11 511
512 × 512	1	503	n/a
	4	1 553	2 979
	16	5 884	11 511
	64	23 455	45 885

Capacity Normal

Image size	Batch	Inference time / image (ms)
Image size	Batch	GPU NVIDIA GeForce 3080Ti	CPU Intel Core i9 7900X	GPU NVIDIA Jetson Xavier NX (ARM)	CPU Raspberry Pi 4 Model B
128 × 128	1	3.75	32	19.03	645
	4	2.43	=	8.14	=
	16	1.90	=	4.48	=
	64	0.42	=	3.69	=
256 × 256	1	7.00	91	32.36	2 717
	4	6.83	=	18.14	=
	16	1.59	=	14.88	=
	64	1.54	=	15.47	=
512 × 512	1	8.93	391	71.63	12 646
	4	5.62	=	66.94	=
	16	5.39	=	58.83	=
	64	–	=	61.78	=

Image size	Batch	GPU memory for inference (MB)	GPU memory for training (MB)
128 × 128	1	178	n/a
	4	248	369
	16	528	929
	64	1 648	3 168
256 × 256	1	248	n/a
	4	528	929
	16	1 648	3 168
	64	6 256	12 255
512 × 512	1	528	n/a
	4	1 648	3 168
	16	6 256	12 255
	64	24 937	48 849

Capacity Large

Image size	Batch	Inference time / image (ms)
Image size	Batch	GPU NVIDIA GeForce 3080Ti	CPU Intel Core i9 7900X	GPU NVIDIA Jetson Xavier NX (ARM)	CPU Raspberry Pi 4 Model B
128 × 128	1	6.31	76	48.57	1 194
	4	4.99	=	14.32	=
	16	1.08	=	10.05	=
	64	0.85	=	7.34	=
256 × 256	1	8.80	205	65.37	5 694
	4	3.46	=	38.04	=
	16	2.25	=	30.14	=
	64	2.32	=	29.54	=
512 × 512	1	25.93	866	168.90	26 320
	4	9.11	=	128.96	=
	16	8.52	=	115.15	=
	64	–	=	–	=

Image size	Batch	GPU memory for inference (MB)	GPU memory for training (MB)
128 × 128	1	288	n/a
	4	421	714
	16	952	1 776
	64	3 075	6 023
256 × 256	1	421	n/a
	4	952	1 776
	16	3 075	6 023
	64	11 701	23 144
512 × 512	1	952	n/a
	4	3 075	6 023
	16	11 701	23 144
	64	46 450	91 874