Benchmarks¶
This page provides information about expected performance for the various cores available on the MitySOM-AM62/A/P family of devices.
FFT Performance¶
The table below lists the FFT kernel processing time, in milliseconds, for:
- the Cortex A-53 cores running at 1.2 GHz using the FFTW single precision (i.e., "float" types) using 1, 2, 3, or 4 threads.
- the C7x integrated DSP running at 850 MHz using the TI provided FFTLIB function
The table measurements were made using on a MitySOM-AM62A with a Quad Core A-53 option at the "V" processor speedgrade. The measurements only include the FFT kernel processing time, they do not include FFT setup time (e.g., twiddle generation) nor data transfer time (in the case of the C7x measurements) between the ARM and the DSP. For both the ARM and the DSP, the FFT data was stored in external LPDDR4 memory and the linux remoteproc framework was used to command the DSP to start and time the FFT operation.
The A-53 measurements should be fairly consistent with and of the MitySOM-AM62, MitySOM-AM62A and the MitySOM-AM62P processor options.
| FFT Size | A53 1 thread, 1.2 GHz | A53 2 threads, 1.2 GHz | A53 3 threads, 1.2 GHz | A53 4 threads, 1.2 GHz | C7x 1 thread, 850 MHz |
|---|---|---|---|---|---|
| 56.0 | 0.001 | 0.054 | 0.073 | 0.051 | 0.013 |
| 512.0 | 0.016 | 0.059 | 0.103 | 0.052 | 0.065 |
| 1024.0 | 0.035 | 0.068 | 0.088 | 0.061 | 0.048 |
| 2048.0 | 0.077 | 0.092 | 0.122 | 0.074 | 0.089 |
| 4096.0 | 0.212 | 0.153 | 0.138 | 0.108 | 0.209 |
| 8192.0 | 0.504 | 0.329 | 0.245 | 0.211 | 0.504 |
| 16384.0 | 1.26 | 0.793 | 0.801 | 0.735 | 1.4 |
| 32768.0 | 4.71 | 2.64 | 3.26 | 1.85 | 3.114 |
| 65536.0 | 16.11 | 9.95 | 8.02 | 6.14 | 7.896 |
| 131072.0 | 43.62 | 23.62 | 17.63 | 14.38 | 17.31 |
| 262144.0 | 94.43 | 57.89 | 43.73 | 30.62 | 40.73 |
| 524288.0 | 225.0 | 119.7 | 81.62 | 68.02 | 85.45 |
| 1048576.0 | 459.98 | 244.04 | 188.63 | 136.8 | 194.0 |
The two figures below graph the above data over different ranges. As can be seen, in some cases it could be advantageous to simply use a Quad Core ARM to perform FFT related functions.
Go to top
