. 2022 Dec 24;10(1):28. doi: 10.3390/bioengineering10010028

Table 3.

Compared with other implementations.

	[19]	[20]	[21]	[31]	This Paper
Platform	Arria10	ZU2	Arria10	Virtex7	XC7Z020
CNN Model	MobileNetV2
Frequency	133 MHz	430 MHz	200 MHz	150 MHz	150 MHz
DSP Usage	1278	212	1220	2160	176 (248 ¹)
On-Chip Memory Usage	1844 M20K (3.07 MB)	145 BRAM	15.3 Mb (1.91 MB)	941.5 BRAM	119.5 BRAM (524.25 KB)
Speed	266.2 FPS	205.3 FPS	1050 FPS (Throughput)	302.3 FPS	70.94 FPS
FPS/MHz per DSP per KB ( $10^{- 6}$ )	0.50	3.54	2.20	0.23	5.13 (3.64 ¹)

¹ As mentioned above, we use LUTs to conduct the multiplications in DWC because of insufficient resources. This will cause unfair results. Here, we assume every multiplication is conducted by DSPs and calculate the relevant results.