Given the results for f32 and f64 memory usage provided by a third party benchmark repo, should we consider updating the "2x lower memory usage" claim in the readme?
To make it easier to read the memory usage data, I took out the other columns to do with timing for both tables:
f32
Sequence length
rustfft memory
phastft memory
8
368 B
32 B
16
624 B
96 B
32
1.152 KB
224 B
64
3.808 KB
480 B
128
6.368 KB
992 B
256
11.48 KB
2.016 KB
512
21.72 KB
4.064 KB
1024
42.20 KB
8.160 KB
2048
83.16 KB
16.35 KB
4096
165.0 KB
32.73 KB
8192
328.9 KB
65.50 KB
16384
656.6 KB
131.0 KB
32768
1.311 MB
262.1 KB
65536
2.622 MB
524.2 KB
131072
5.244 MB
1.048 MB
262144
10.48 MB
2.097 MB
524288
20.97 MB
4.194 MB
1048576
41.94 MB
8.388 MB
2097152
83.88 MB
16.77 MB
4194304
167.7 MB
33.55 MB
8388608
335.5 MB
67.10 MB
f64
Sequence length
rustfft memory
phastft memory
8
320 B
64 B
16
480 B
192 B
32
816 B
448 B
64
3.968 KB
960 B
128
7.040 KB
1.984 KB
256
13.18 KB
4.032 KB
512
25.47 KB
8.128 KB
1024
50.04 KB
16.32 KB
2048
99.20 KB
32.70 KB
4096
197.5 KB
65.47 KB
8192
394.1 KB
131.0 KB
16384
787.3 KB
262.0 KB
32768
1.573 MB
524.2 KB
65536
3.146 MB
1.048 MB
131072
6.292 MB
2.097 MB
262144
12.58 MB
4.194 MB
524288
25.16 MB
8.388 MB
1048576
50.33 MB
16.77 MB
2097152
100.6 MB
33.55 MB
4194304
201.3 MB
67.10 MB
8388608
402.6 MB
134.2 MB
Just eyeballing the results, it seems PhastFT has ≈ 5x lower memory usage for f64 ≈ 3x lower memory usage for f64. Maybe it would be better compute the ratio of rustfft to phastft memory usage for all sequence sizes and then take the average/median/min? Perhaps we should re-run the benchmarks on the original benchmark machine before anything.
@Shnatsel
Given the results for
f32andf64memory usage provided by a third party benchmark repo, should we consider updating the "2x lower memory usage" claim in the readme?To make it easier to read the memory usage data, I took out the other columns to do with timing for both tables:
f32rustfftmemoryphastftmemoryf64rustfftmemoryphastftmemoryJust eyeballing the results, it seems
PhastFThas ≈ 5x lower memory usage for f64 ≈ 3x lower memory usage for f64. Maybe it would be better compute the ratio ofrustffttophastftmemory usage for all sequence sizes and then take the average/median/min? Perhaps we should re-run the benchmarks on the original benchmark machine before anything.Curious to hear your thoughts!