deeplearning benchmark 6800XT RoCM 5.4.3

gistfile1.txt

Step    Img/sec total_loss
1       images/sec: 177.0 +/- 0.0 (jitter = 0.0)        7.765
10      images/sec: 179.0 +/- 0.3 (jitter = 1.1)        8.049
20      images/sec: 178.6 +/- 0.3 (jitter = 1.6)        7.808
30      images/sec: 178.7 +/- 0.3 (jitter = 1.6)        7.976
40      images/sec: 178.8 +/- 0.2 (jitter = 1.7)        7.591
50      images/sec: 178.8 +/- 0.2 (jitter = 1.6)        7.549
60      images/sec: 29.6 +/- 4.1 (jitter = 1.7) 7.820
70      images/sec: 33.6 +/- 3.5 (jitter = 1.9) 7.821
80      images/sec: 37.4 +/- 3.1 (jitter = 1.9) 7.847
90      images/sec: 41.0 +/- 2.7 (jitter = 2.0) 8.028
100     images/sec: 44.4 +/- 2.5 (jitter = 2.1) 8.029
----------------------------------------------------------------
total images/sec: 44.37
----------------------------------------------------------------

after amdgpu-clocks tuning:

/etc/default/amdgpu-custom-states.card0

OD_SCLK:
0: 500Mhz
1: 2499Mhz
OD_MCLK:
0: 97Mhz
1: 1000Mhz
OD_VDDGFX_OFFSET:
-100mV
OD_RANGE:
SCLK:     500Mhz       2800Mhz
MCLK:     674Mhz       1075Mhz

# Force power limit (in micro watts):
FORCE_POWER_CAP: 300000000
FORCE_PERF_LEVEL: manual

results

Step    Img/sec total_loss
1       images/sec: 180.8 +/- 0.0 (jitter = 0.0)        7.765
10      images/sec: 182.2 +/- 0.8 (jitter = 1.6)        8.049
20      images/sec: 181.9 +/- 0.5 (jitter = 1.0)        7.808
30      images/sec: 182.3 +/- 0.4 (jitter = 1.1)        7.976
40      images/sec: 182.3 +/- 0.3 (jitter = 1.2)        7.591
50      images/sec: 182.5 +/- 0.3 (jitter = 1.3)        7.549
60      images/sec: 182.3 +/- 0.3 (jitter = 1.4)        7.819
70      images/sec: 182.4 +/- 0.3 (jitter = 1.5)        7.820
80      images/sec: 182.4 +/- 0.3 (jitter = 1.4)        7.848
90      images/sec: 182.4 +/- 0.2 (jitter = 1.6)        8.027
100     images/sec: 182.4 +/- 0.2 (jitter = 1.6)        8.028
----------------------------------------------------------------
total images/sec: 182.40
----------------------------------------------------------------

pytorch's vgg16 eval at fp32: 24.5ms avg
pytorch's vgg16 train at fp32: 105.4ms avg
pytorch's resnet152 eval at fp32: 35.0ms avg
pytorch's resnet152 train at fp32: 136.8ms avg
pytorch's densenet161 eval at fp32: 37.8ms avg
pytorch's densenet161 train at fp32: 152.6ms avg

pytorch's vgg16 eval at fp16: 392.6ms avg
pytorch's vgg16 train at fp16: 456.8ms avg
pytorch's resnet152 eval at fp16: 48.4ms avg
pytorch's resnet152 train at fp16: 120.1ms avg
pytorch's densenet161 eval at fp16: 32.8ms avg
pytorch's densenet161 train at fp16: 143.4ms avg