texture memory faster if compiled with sm13 istead of sm20

Hi

What benefits can I get if I compiled with sm20?
When testing the “simpleTexture” in SDK 3.1, I got the following results:

==GTX 470 with sm_20==
Loaded ‘lena_bw.pgm’, 512 x 512 pixels
Processing time: 0.188277 (ms)
1392.33 Mpixels/sec

==GTX 470 with sm_13==
Loaded ‘lena_bw.pgm’, 512 x 512 pixels
Processing time: 0.151181 (ms)
1733.98 Mpixels/sec

==GTX 470 with sm_20 & fast_math==
Loaded ‘lena_bw.pgm’, 512 x 512 pixels
Processing time: 0.092599 (ms)
2830.97 Mpixels/sec

==GTX 470 with sm_13 & fast_math==
Loaded ‘lena_bw.pgm’, 512 x 512 pixels
Processing time: 0.097954 (ms)
2676.21 Mpixels/sec