Skip to content

Instantly share code, notes, and snippets.

@mwrnd

mwrnd/benchmark_results_RX580_ROCm2.8.13

Last active October 23, 2019 20:19
Show Gist options
  • Save mwrnd/f59a61d35344325530c4ed8693352b4a to your computer and use it in GitHub Desktop.
Save mwrnd/f59a61d35344325530c4ed8693352b4a to your computer and use it in GitHub Desktop.
Download ZIP
AMD Radeon RX 580 Tensorflow benchmarking results with ROCm 2.8.13
Raw
benchmark_results_RX580_ROCm2.8.13
AMD Radeon RX 580 8GB tensorflow/benchmarks Results
By Matthew Wielgus 2019-10-22
Video Card: MSI Radeon RX 580 8GB ARMOR OC (rocm-smi -v VBIOS version: Unable to get)
Motherboard: MSI X570-A Pro with 32GB DDR4-3000 BIOS H.40
Processor: AMD Ryzen 5 3600X
OS: Ubuntu 18.04.2 no apt dist-upgrade
Kernel: 4.18.0-15-generic
rocm-dkms: 2.8.13 installed through apt
tensorflow-rocm: 1.14.2 installed through pip
tensorflow benchmarks: abb1aec2f2db4ba73fac2e1359227aef59b10258
https://codeload.github.com/tensorflow/benchmarks/zip/abb1aec2f2db4ba73fac2e1359227aef59b10258
tensorflow_models: 1.13.0
https://codeload.github.com/tensorflow/models/zip/v1.13.0
Legend:
X means XLA was enabled
export TF_XLA_FLAGS=--tf_xla_cpu_global_jit
R means ROCm Fusion was enabled
export TF_ROCM_FUSION_ENABLE=1
F means 16-Bit Floating Point was used (--use_fp16)
C means MIOpen Compute Unit optimations do not exist for this GPU
na means the benchmark did not run or batch size was too large
Note: - for performance gain calculations, baseline 0s are runs with no flags enabled
- trivial models and na are excluded
==== images/sec results for various dataset and model combinations ====
imagenet
model batchsize = 016 016F 016XR 016XRF 032 032F 032XR 032XRF 064 064F 064XR 064XRF 128 128F 128XR 128XRF 256 256F 256XR 256XRF
alexnet 223 237 na 239 318 325 309 327 396 401 400 401 446 445 448 445 466 463 469 466
googlenet 135 138 136 138 155 160 156 161 165 174 166 175 166 180 167 181 166 181 165 182
inception3 25.7 29.7 26.5 30.2 26.6 31.3 27.4 32.0 27.5 32.2 28.3 32.7 32.5 na na na na na na na
inception4 12.2 14.1 na 14.3 12.6 14.7 12.9 14.9 14.9 na na na na na na na na na na na
lenet5 3920 4104 na 4167 6918 6785 6829 6761 11132 10556 10618 10741 16169 14848 14952 14892 20677 18910 18656 19027
mobilenet 270 219 274 221 338 275 341 273 389 339 396 329 405 383 424 376 453 439 427 413
nasnet 8.3 37.5 7.9 39.1 8.4 8.8 8.5 8.8 9.1 9.1 8.7 9.0 9.1 9.8 9.1 9.1 na na na na
official_ncf 1478 1196 1468 1196 2927 2409 2913 2409 5823 4788 5861 4752 11478 9499 11512 9496 22590 18712 22764 18731
overfeat 66.4 64.3 na 64.3 88.6 88.4 88.4 88.5 106 107 106 108 116 120 116 119 120 125 120 125
resnet101 29.1 32.1 30.4 32.9 33.3 36.0 34.9 36.8 37.9 na na na na na na na na na na na
resnet101_v2 29.3 32.3 na 33.3 33.6 36.2 35.3 37.2 35.7 38.0 37.8 39.1 na na na na na na na na
resnet152 20.3 22.1 na 22.5 23.1 24.6 24.2 25.2 25.9 na na na na na na na na na na na
resnet152_v2 20.5 22.2 na 22.7 23.3 24.8 24.5 25.4 26.0 na na na na na na na na na na na
resnet50 51.4 58.7 na 60.6 57.9 65.8 60.9 67.7 61.0 70.0 64.1 72.2 71.3 na na na na na na na
resnet50_v1.5 47.4 53.3 na 55.1 52.8 59.0 55.4 60.7 57.1 63.0 60.5 64.8 63.6 na na na na na na na
resnet50_v2 51.8 59.2 na 61.4 58.5 66.3 62.1 68.4 61.7 70.7 65.7 73.0 71.8 na na na na na na na
trivial 4012 587 na 588 7128 1162 7106 1157 12965 2270 12521 2270 21281 4350 21587 4354 30196 7975 29805 7959
vgg11 39.7 42.6 39.7 42.6 45.0 47.2 45.0 47.2 47.9 49.9 48.0 49.9 47.8 50.2 48.8 50.3 50.8 na na na
vgg16 20.3 22.0 20.3 21.9 21.9 23.6 21.9 23.5 22.6 24.2 22.7 24.2 24.4 na na na na na na na
vgg19 16.6 17.7 na 17.6 17.8 18.9 17.8 18.9 18.3 19.4 18.4 19.3 19.5 na na na na na na na
Average Gain 0 +5.8% +0.5% +7.3% 0 +4.9% +2.1% +6.2% 0 +3.5% +1.5% +4.4% 0 -0.8% +0.1% -2.0% 0 -2.7% -2.5% -3.3%
Median Gain 0 +8.4% +0.4% +10.7% 0 +6.4% +1.2% +7.3% 0 +5.5% +0.7% +5.5% 0 +1.6% +0.4% -0.1% 0 -1.9% -0.3% -4.0%
cifar10
model batchsize = 016 016F 016XR 016XRF 032 032F 032XR 032XRF 064 064F 064XR 064XRF 128 128F 128XR 128XRF 256 256F 256XR 256XRF
trivial 7199 4994 7887 4998 14904 9573 15664 9371 24906 18810 27120 18856 44463 37492 50211 39096 74973 71439 90453 70917
alexnet 2791 269 na 271 4503 294 4448 291 6058 304 6101 302 7480 313 7508 309 8651 313 8672 311
nasnet na 45.3 na 45.0 53.8 53.0 53.9 52.2 61.0 59.8 60.0 62.1 104 82.7 59.4 na na na na na
resnet110 286 257 na 275 437 430 484 467 596 617 662 663 707 794 740 801 760 888 778 867
resnet110_v2 287 260 323 285 443 438 494 475 597 623 667 672 708 794 714 774 758 884 734 814
resnet20 1415 1264 1507 1339 2234 2132 2392 2251 3036 3086 3282 3288 3639 3964 3761 4004 3937 4462 4018 4380
resnet20_v2 1417 1267 1560 1381 2193 2121 2399 2249 2975 3055 3284 3251 3555 3916 3627 3849 3842 4398 3780 4120
resnet32 930 831 987 888 1453 1382 1569 1485 1963 2013 2162 2154 2351 2595 2439 2611 2532 2907 2582 2845
resnet32_v2 931 834 1025 910 1445 1391 1613 1506 1943 2008 2167 2156 2312 2569 2358 2520 2490 2873 2432 2676
resnet44 695 617 726 656 1065 1030 1162 1106 1451 1497 1597 1601 1735 1924 1804 1946 1865 2147 1907 2105
resnet44_v2 706 630 770 671 1069 1028 1187 1107 1447 1499 1609 1605 1714 1917 1743 1871 1846 2136 1789 1978
resnet56 557 488 591 518 853 822 927 881 1151 1189 1266 1278 1369 1528 1427 1538 1477 1706 1507 1669
resnet56_v2 553 491 610 542 848 823 954 909 1145 1194 1286 1283 1363 1523 1380 1488 1463 1700 1419 1570
Average Gain 0 -10.7% +8.4% -3.9% 0 -3.1% +9.1% +3.5% 0 +2.7% +9.5% +9.8% 0 +8.2% -1.3% +10.4% 0 +15.4% -0.3% +10.0%
Median Gain 0 -10.7% +9.1% -4.2% 0 -3.3% +9.4% +3.6% 0 +3.3% +10.4% +10.9% 0 +11.1% +2.0% +9.7% 0 +15.4% +0.2% +9.4%
* Excluded alexnet and trivial as outliers as their 16-Bit Floating-Point results were low
coco
model batchsize = 016 016F 016R 016XR 016XRF 032 032F 032XR 032XRF 064 064F 064XR 064XRF 128 128F 128XR 128XRF 256 256F 256XR 256XRF
ssd300 17.1 16.4 na 17.4 16.5 17.6 17.1 17.9 17.2 17.8 16.8 18.2 17.0 16.9 na na na na na na na
Percentage Gain 0 -4.09 na +1.75 -3.51 0 -2.84 +1.7 -2.27 0 -5.62 +2.25 -4.49 0
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment