The representative performance specifications below are results averaged over
many repetitions of each function call. Data is in cache. Results
were measured using a 233 MHz Pentium® II processor PC with a
256K second level cache and Windows* NT* 4.0. Complete performance numbers for all the functions
can be found in the release--check the CSV files under plsuite\examples\rltest.
Selected Functions |
Single Precision Floating Point |
16-Bit Integer |
Units |
256 point Real FFT |
34.4 |
18.2 (16-bit calculations) |
microseconds |
128 point Complex FFT |
26.8 |
10.3 (16-bit calculations) |
microseconds |
64 dim. Dot product |
3.90 |
2.44 |
clocks per element |
32 dim. Euclidean Distance Squared |
6.70 |
8.23 |
clocks per element |
32 dim. Maholonobis Distance with Diagonal Matrix |
7.13 |
5.48 |
clocks per element |
32 dim. Maholonobis Distance with Full Matrix |
97.2 |
55.5 |
clocks per element |
10 Cepstral Coefficients with Bandpass 200 and FFT order 8 |
5599.02 |
5023.66 |
clocks per cepstral coefficient |
20 Dim. x 8 Gaussian Mixtures with Diagonal Matrix |
249.72 |
180.64 |
clocks per Gaussian |
20 Dim. x 8 Gaussian Mixtures with Full Matrix |
2940.16 |
982.92 |
clocks per Gaussian |
Discrete Constrained Jump HMM with 12 States 256 Symbols 1000 Observations |
13.82 |
7.26 |
clocks per lattice point |
Discrete Left-Right HMM with 12 States 256 Symbols 1000 Observations |
62.06 |
26.41 |
clocks per lattice point |
155x184x165x16 MLP Neural Network |
846 |
5630 |
clocks per neuront |