Utilizing the Double-Precision Floating-Point Computing Power of GPUs for RSA Acceleration

Table 8

Throughput of operations per second.

	Neves and Araujo [22]	Henry and Goldberg [23]	Jang et al. [19]	Emmart and Weems [7]	Yang [24]	Jeffrey and Robinson [21]	Ours
CUDA platform	GTX 260	M2050	GTX 580	GTX780Ti	GT 750 m	GTX TITAN
SM number	24	14	16	15	2	14
Shader Clock (GHz)	1.242	1.150	1.544	0.876	0.967	0.837
Int Mul/SM (/Clock)	8 (24-bit)	16	16	32	32	32
Int Mul (G/s)	134 (238)	258	395	420	62	375
Throughput scaling factor	0.357	0.688	1.053	1.12	0.165	1
Latency scaling factor	1.484	1.374	1.845	1.047	1.155	1
1024-bit MulMod (ops/s)	—		—	—	—
MulMod (scaled) (ops/s)	—		—	—	—
RSA-1024 (ops/s)	41,426	—	—	—	34,981	—	—
RSA-2048 (ops/s)	—	—	12,044	62,365	5,244	—	42,211
RSA-2048 (ms)	—	—	13.83	60.07	195.27	—	21.22
RSA-2048 (scaled) (ops/s)	14,504	—	11,438	55,683	31,782	—	42,211
RSA-2048 (scaled) (ms)	—	—	25.51	62.87	225.60	—	21.22
RSA-4096 (ops/s)	—	—	—	5,257	—	—	5,790
RSA-4096 (scaled) (ops/s)	—	—	—	4,693	—	—	5,790

et al. also report the latency of RSA-2048 decryption is 6.5 ms (after scaled 6.8 ms) when the Batch Size is 1, at the moment the throughput is 154. peak 2048-bit RSA throughput, when Threads/RSA is , window size is 6, Max Reg. is 127, and Batch Size is . peak 4096-bit RSA throughput, when Threads/RSA is , window size is 6, Max Reg. is 127, and Batch Size is .