Skip to main content

View full-text article in PMC

. 2025 Aug 22;8:539. doi: 10.1038/s41746-025-01929-z

Table 4.

Inference latency in milliseconds for selected models in three hardware configurations

		Apple M2	RTX 3060Ti		NVIDIA A100
Backbone	Model	B=1	B=1	B=10	B=1	B=10	B=100
HRNet-W32-256	Depth	129	11.4^a	3.79^a	29.4^a	2.94^a	0.956^a
HRNet-W32-256	IIF-2	316	18.2^a	14.7^a	29.6^a	3.96^a	3.20^a
HRNet-W32-384	LIF-3	208	12.5^a	5.87^a	36.2	3.64^a	1.52^a
HRNet-W48-384	RGB	629	31.7^a	24.6^a	91.8	8.87^a	6.53^a
HRNet-W48-384	LIF-3	1070	51.7	44.8	89.9	11.9^a	9.59^a
HRFormer-S	IIF-2	432	23.3^a	16.5^a	55.9	7.60^a	6.49^a
HRFormer-B	Depth	784	52.4	40.0	47.0	16.6^a	15.0^a
HRFormer-B	IIF-2	1090	59.3	52.0	57.0	21.2^a	19.2^a

The three hardware configurations represent a cot-side small PC, on-site desktop PC, or off-site server.

Bold indicates suitability for 1fps inference.

B batch size.

^a30fps inference.