Tensorflow Benchmark

Google社がディープラーニング向けに開発したフレームワーク「Tensorflow」について、ベンチマークテスト tf_cnn_benchmarks を継続的に実施しています。さまざまな構成による実機を用いて実行したテスト結果を公開していますので、それぞれの機械学習モデルに対するシステム性能を比較することができます。

ベンチマークの値(total images/sec)の大きいものが、性能が高いことを示します。

ベンチマーク 結果抜粋

AMD Instinct MI210

System: HPC-ProServer DPeR7625
CPU:(2) AMD EPYC 9354 3.25G 32C/64T
Mem : 384GB (24) 16GB RDIMM, 4800MT/s
OS : RockyLinux8.10
GPU : (1) AMD Instinct MI210
ROCm : 6.2.0
AMD Infinity Hub : Tensorflow 2.15 (docker)

model batch size total images/sec
inception4 32 188
inception4 64 211
inception3 32 364
inception3 64 417
alexnet 32 3,119
alexnet 64 4,259
resnet50 32 544
resnet50 64 603
resnet50_v2 32 557
resnet50_v2 64 615
resnet101 32 317
resnet101 64 347
resnet101_v2 32 323
resnet101_v2 64 353
resnet152 32 220
resnet152 64 239
resnet152_v2 32 224
resnet152_v2 64 243
vgg16 32 276
vgg16 64 278
vgg19 32 227
vgg19 64 230
trivial 32 32,996
trivial 64 60,147
NVIDIA A100 Tensor Core 40GB

System : DGX A100
CPU : (2) Dual AMD Rome 7742 128 cores total, 2.25 GHz(base), 3.4 GHz (max boost)
Mem : 1TB
GPU : (8) NVIDIA A100 Tensor Core 40 GB
OS : DGX OS 5.0
Driver : 450.80.02 / CUDA : 11.2
Anaconda 4.9.2
tensorflow-gpu 2.4.1
Date : 2021-03

model Batch
Size
GPUs Speed Up Ratio
    1 2 4 8 1gpu/1gpu 2gpu/1gpu 4gpu/1gpu 8gpu/1gpu
inception4 32 252 447 761 1224 1.0 1.8 3.0 4.9
inception4 64 286 539 978 1625 1.0 1.9 3.4 5.7
inception3 32 474 786 1321 2089 1.0 1.7 2.8 4.4
inception3 64 535 993 1826 3169 1.0 1.9 3.4 5.9
alexnet 32 4644 8798 16578 22505 1.0 1.9 3.6 4.8
alexnet 64 6537 12799 24830 39477 1.0 2.0 3.8 6.0
resnet50 32 678 1226 2121 3213 1.0 1.8 3.1 4.7
resnet50 64 786 1492 2726 4491 1.0 1.9 3.5 5.7
resnet50_v2 32 698 1249 2219 3462 1.0 1.8 3.2 5.0
resnet50_v2 64 805 1528 2819 4957 1.0 1.9 3.5 6.2
resnet101 32 414 724 1151 1701 1.0 1.8 2.8 4.1
resnet101 64 488 901 1596 2660 1.0 1.8 3.3 5.4
resnet101_v2 32 423 739 1200 1836 1.0 1.7 2.8 4.3
resnet101_v2 64 498 917 1638 2712 1.0 1.8 3.3 5.4
resnet152 32 279 486 781 1172 1.0 1.7 2.8 4.2
resnet152 64 340 627 1066 1831 1.0 1.8 3.1 5.4
resnet152_v2 32 295 502 820 1271 1.0 1.7 2.8 4.3
resnet152_v2 64 346 633 1115 1883 1.0 1.8 3.2 5.4
vgg16 32 566 1094 2172 3824 1.0 1.9 3.8 6.8
vgg16 64 609 1196 2343 4407 1.0 2.0 3.8 7.2
vgg19 32 480 955 1770 3055 1.0 2.0 3.7 6.4
vgg19 64 515 1021 1960 3640 1.0 2.0 3.8 7.1
trivial 32 34166 50915 83643 125229 1.0 1.5 2.4 3.7
trivial 64 57025 89881 152530 236302 1.0 1.6 2.7 4.1

NVIDIA A40

System : HPC-ProServer DPeR750XA
CPU : (2) Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz
Mem : Total 512GB (16) 32GB DDR4 3200MHz
GPU : (4) NVIDIA A40
OS : Ubuntu 20.04
Driver : 510.47.03 / CUDA : 11.6
NGC Docker : 22.01-tf1-py3 , Tensorflow 1.15
Date : 2022-02

model Batch
Size
GPUs Speed Up Ratio
    1 2 4 1gpu/1gpu 2gpu/1gpu 4gpu/1gpu
inception4 32 169 329 645 1.0 1.9 3.8
inception4 64 176 348 1161 1.0 2.0 6.6
inception3 32 316 594 1161 1.0 1.9 3.7
inception3 64 324 607 1273 1.0 1.9 3.9
alexnet 32 2241 4288 1921 1.0 1.9 0.9
alexnet 64 3077 6583 3972 1.0 2.1 1.3
resnet50 32 397 773 1476 1.0 1.9 3.7
resnet50 64 439 867 1663 1.0 2.0 3.8
resnet50_v2 32 404 793 1489 1.0 2.0 3.7
resnet50_v2 64 447 883 1715 1.0 2.0 3.8
resnet101 32 240 463 880 1.0 1.9 3.7
resnet101 64 265 519 1002 1.0 2.0 3.8
resnet101_v2 32 244 472 877 1.0 1.9 3.6
resnet101_v2 64 268 526 1040 1.0 2.0 3.9
resnet152 32 168 322 602 1.0 1.9 3.6
resnet152 64 184 355 705 1.0 1.9 3.8
resnet152_v2 32 169 325 623 1.0 1.9 3.7
resnet152_v2 64 185 362 716 1.0 2.0 3.9
vgg16 32 254 466 618 1.0 1.8 2.4
vgg16 64 273 517 900 1.0 1.9 3.3
vgg19 32 220 412 588 1.0 1.9 2.7
vgg19 64 234 455 761 1.0 1.9 3.2
trivial 32 28122 34912 16928 1.0 1.2 0.6
trivial 64 51270 70641 33871 1.0 1.4 0.7

NVIDIA RTX A4000

System : HPC-ProServer DPrR7920
CPU : (2) Intel(R) Xeon(R) Silver 4110 CPU @ 2.10GHz
Mem : Total 256GB (16) 16GB DDR4 2933MHz
GPU : (2) NVIDIA RTX A4000
OS : Ubuntu 20.04
Driver : 515.65.01 / CUDA : 11.7
NGC Docker : 22.09-tf1-py3 , Tensorflow 1.15
Date : 2022-10

model Batch
Size
GPUs Speed Up Ratio
    1 2 1gpu/1gpu 2gpu/1gpu
inception4 32 108 209 1 1.94
inception4 64 111 221 1 1.99
inception3 32 206 386 1 1.88
inception3 64 213 408 1 1.91
alexnet 32 1483 1383 1 0.93
alexnet 64 1789 2530 1 1.41
resnet50 32 279 530 1 1.90
resnet50 64 297 580 1 1.95
resnet50_v2 32 283 535 1 1.89
resnet50_v2 64 304 590 1 1.94
resnet101 32 167 320 1 1.92
resnet101 64 179 351 1 1.96
resnet101_v2 32 170 324 1 1.91
resnet101_v2 64 182 358 1 1.96
resnet152 32 117 223 1 1.91
resnet152 64 126 248 1 1.96
resnet152_v2 32 118 227 1 1.92
resnet152_v2 64 128 250 1 1.96
vgg16 32 156 243 1 1.56
vgg16 64 164 294 1 1.79
vgg19 32 132 219 1 1.66
vgg19 64 138 250 1 1.81
trivial 32 15063 9285 1 0.62
trivial 64 26877 17236 1 0.64

NVIDIA GeForce RTX 3090

System : HPC-ProServer DPeT640
CPU : (2) Intel Xeon Gold 6226R 2.9GHz 16C/32T, 10.4GT/s, 22 M キャッシュ, ターボ, HT (150W)
Mem : 384GB (12) 32GB RDIMM, 2933MT/s, 2R
GPU : (2) NVIDIA GeForce 3090
OS : CentOS v7.9
Driver : 460.32.03 / CUDA : 11.2
Anaconda 4.9.2
tensorflow-gpu 2.4.1
Date : 2021-03

model Batch
Size
GPUs Speed Up Ratio
    1 2 1gpu/1gpu 2gpu/1gpu
inception4 32 152 278 1 1.8
inception4 64 166 321 1 1.9
inception3 32 300 519 1 1.7
inception3 64 340 643 1 1.9
alexnet 32 2650 1068 1 0.4
alexnet 64 3705 1967 1 0.5
resnet50 32 440 774 1 1.8
resnet50 64 496 927 1 1.9
resnet50_v2 32 444 788 1 1.8
resnet50_v2 64 503 947 1 1.9
resnet101 32 241 380 1 1.6
resnet101 64 294 553 1 1.9
resnet101_v2 32 257 451 1 1.8
resnet101_v2 64 297 550 1 1.9
resnet152 32 173 306 1 1.8
resnet152 64 205 382 1 1.9
resnet152_v2 32 171 318 1 1.9
resnet152_v2 64 206 384 1 1.9
vgg16 32 294 326 1 1.1
vgg16 64 311 451 1 1.5
vgg19 32 252 307 1 1.2
vgg19 64 262 387 1 1.5
trivial 32 25999 9339 1 0.4
trivial 64 44249 18098 1 0.4

ベンチマーク 個別結果

AMD Instinct MI210 : AMD Infinity Hub Tensorflow 2.15

NVIDIA RTX A4000 Tensorflow 1.15 NGC Docker

NVIDIA A40 Tensorflow 1.15 NGC Docker

NVIDIA GeForce RTX 3090 Tensorflow 2.4.1 Anaconda 4.9.2

NVIDIA A100 Tensor Core 40GB (DGX A100) Tensorflow 2.4.1 Anaconda 4.9.2

NVIDIA TITAN RTX (DPeR740) Tensorflow v1.12.0 Anaconda3 CuDNN v7.5

NVIDIA Tesla V100 PCIe 32G Tensorflow v1.12.0 Nvidia-Docker

NVIDIA GeForce RTX 2080Ti (DPeR740) Tensorflow v1.12.0 pip

NVIDIA GeForce GTX 1080Ti 4枚構成 (DPeT640) Tensorflow v1.12.0 pip

NVIDIA GeForce RTX 2080 (DPeR740) Tensorflow v1.11.0 Anaconda/Nvidia-Docker

NVIDIA TITAN V (DPeR740) Tensorflow v1.10 Source/Anaconda/Nvidia-Docker

NVIDIA GeForce GTX 1080Ti (DPeR730) Tensorflow v1.10 CuDNN v7.1

AMD Radeon Vega Frontier Edition (DPrT7910) Tensorflow v1.3 ROCm v1.8.118