{"id":1124,"date":"2021-08-31T03:10:12","date_gmt":"2021-08-30T18:10:12","guid":{"rendered":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/?p=1124"},"modified":"2022-02-25T10:44:30","modified_gmt":"2022-02-25T01:44:30","slug":"optimize-distributed-ai-training-oneapi-toolkits","status":"publish","type":"post","link":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/2021\/08\/31\/optimize-distributed-ai-training-oneapi-toolkits\/","title":{"rendered":"\u30a4\u30f3\u30c6\u30eb\u00ae oneAPI \u30c4\u30fc\u30eb\u30ad\u30c3\u30c8\u3092\u7528\u3044\u305f\u5206\u6563\u578b AI \u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u306e\u6700\u9069\u5316"},"content":{"rendered":"

\u30a4\u30f3\u30af\u30ea\u30e1\u30f3\u30bf\u30eb\u30fb\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3067\u5927\u5e45\u306a\u6027\u80fd\u5411\u4e0a\u3092\u5b9f\u73fe<\/h2>\n

\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u306e\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u306f\u3001\u6025\u901f\u306a\u30da\u30fc\u30b9\u3067\u5897\u52a0\u3057\u3066\u3044\u307e\u3059\u3002\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u30fb\u30d9\u30fc\u30b9\u306e\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306f\u3001\u5927\u91cf\u306e\u30c7\u30fc\u30bf\u3092\u51e6\u7406\u3057\u3066\u30d1\u30bf\u30fc\u30f3\u3092\u898b\u3064\u3051\u3001\u753b\u50cf\u5206\u985e\u3001\u7269\u4f53\u691c\u51fa\u3001\u6642\u7cfb\u5217\u4e88\u6e2c\u306a\u3069\u3092\u884c\u3044\u307e\u3059\u3002\u30c7\u30fc\u30bf\u306e\u53ef\u7528\u6027\u304c\u9ad8\u307e\u308b\u306b\u3064\u308c\u3001\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u30fb\u30e2\u30c7\u30eb\u306e\u8907\u96d1\u3055\u3082\u5897\u3057\u3066\u304d\u307e\u3057\u305f\u3002ResNet<\/a> (\u82f1\u8a9e) \u3084 VGG (Visual Geometry Group)<\/a> (\u82f1\u8a9e) \u306a\u3069\u306e\u30e2\u30c7\u30eb\u306f\u3001\u6570\u767e\u4e07\u500b\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u30fc\u304c\u3042\u308a\u3001\u6570\u5341\u5104\u56de\u306e\u6d6e\u52d5\u5c0f\u6570\u70b9\u6f14\u7b97\u3092\u884c\u3044\u307e\u3059\u3002GPT-3 (Generative Pretrained Transformer 3)<\/a> (\u82f1\u8a9e)\u3084 BERT (Bidirectional Encoder Representations from Transformers)<\/a> (\u82f1\u8a9e) \u306e\u3088\u3046\u306a\u6700\u8fd1\u306e\u30e2\u30c7\u30eb\u306f\u3001\u6570\u5341\u5104\u304b\u3089 1 \u5146\u500b\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u30fc\u304c\u3042\u308a\u307e\u3059\u3002\u305d\u306e\u305f\u3081\u3001\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u30fb\u30e2\u30c7\u30eb\u306e\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u306f\u3001\u8a08\u7b97\u91cf\u304c\u591a\u304f\u3001\u6642\u9593\u304c\u304b\u304b\u308b\u30d7\u30ed\u30bb\u30b9\u3067\u3059\u3002\u89e3\u6c7a\u307e\u3067\u306e\u6642\u9593\u3092\u77ed\u7e2e\u3059\u308b\u306b\u306f\u3001\u30b7\u30f3\u30b0\u30eb\u30b3\u30a2\u3084\u30de\u30eb\u30c1\u30b3\u30a2\u306e\u6027\u80fd\u3092\u6700\u9069\u5316\u3059\u308b\u4ee5\u5916\u306b\u3001\u8907\u6570\u306e\u30ce\u30fc\u30c9\u3078\u306e\u30b9\u30b1\u30fc\u30eb\u30a2\u30a6\u30c8 (\u5206\u6563\u5b66\u7fd2\u306a\u3069) \u3082\u691c\u8a0e\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u306e\u4f5c\u696d\u306f\u3001\u30e2\u30c7\u30eb\u3092\u5206\u5272\u3059\u308b\u65b9\u6cd5 (\u30e2\u30c7\u30eb\u4e26\u5217)\u3001\u30c7\u30fc\u30bf\u3092\u5206\u5272\u3059\u308b\u65b9\u6cd5 (\u30c7\u30fc\u30bf\u4e26\u5217)\u3001\u3042\u308b\u3044\u306f\u4e21\u65b9\u3092\u7d44\u307f\u5408\u308f\u305b\u305f\u65b9\u6cd5 (\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u4e26\u5217) \u3067\u5b9f\u73fe\u3067\u304d\u307e\u3059\u3002<\/p>\n

\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001\u8a08\u7b97\u30ce\u30fc\u30c9\u306e\u30af\u30e9\u30b9\u30bf\u30fc\u4e0a\u3067\u306e\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u30fb\u30d9\u30fc\u30b9\u306e\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306e\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3068\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u306b\u7126\u70b9\u3092\u5f53\u3066\u3066\u3044\u307e\u3059\u3002\u3053\u3053\u3067\u306f\u3001\u534a\u6559\u5e2b\u3042\u308a\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (S-GAN) \u3092\u7528\u3044\u3066\u753b\u50cf\u3092\u5206\u985e\u3059\u308b\u65b9\u6cd5\u3092\u8aac\u660e\u3057\u307e\u3059\u3002\u307e\u305f\u3001OpenMP\u3001TensorFlow<\/em>\u3001\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u3067\u3055\u307e\u3056\u307e\u306a\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3092\u884c\u3044\u307e\u3057\u305f\u3002\u30b7\u30f3\u30b0\u30eb\u30ce\u30fc\u30c9\u306b\u304a\u3044\u3066\u3001\u30a4\u30f3\u30c6\u30eb\u306e\u6700\u9069\u5316\u306b\u3088\u308a 2 \u500d\u306e\u30b9\u30d4\u30fc\u30c9\u30a2\u30c3\u30d7\u3092\u5b9f\u73fe\u3057\u307e\u3057\u305f\u30028 \u30ce\u30fc\u30c9\u306e\u30af\u30e9\u30b9\u30bf\u30fc\u306b\u62e1\u5f35\u3057\u305f\u5834\u5408\u3001\u5168\u4f53\u3067 16 \u500d\u306e\u30b9\u30d4\u30fc\u30c9\u30a2\u30c3\u30d7\u3092\u9054\u6210\u3057\u307e\u3057\u305f\u3002\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI\u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u306e\u753b\u50cf\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8 (1 \u79d2\u3042\u305f\u308a\u306e\u753b\u50cf\u6570) \u306f\u3001Open MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u3068\u6bd4\u8f03\u3057\u3066\u3001\u5358\u4e00\u30ce\u30fc\u30c9\u3067\u6700\u5927 27\uff05\u30018 \u30ce\u30fc\u30c9\u306e\u30af\u30e9\u30b9\u30bf\u30fc\u3067\u6700\u5927 18\uff05\u3001\u4e00\u8cab\u3057\u3066\u5411\u4e0a\u3057\u307e\u3057\u305f\u3002\u3053\u308c\u3089\u306e\u6027\u80fd\u5411\u4e0a\u306e\u591a\u304f\u306f\u30b3\u30fc\u30c9\u3092\u5909\u66f4\u3059\u308b\u3053\u3068\u306a\u304f\u9054\u6210\u3055\u308c\u3066\u304a\u308a\u3001\u3053\u308c\u3089\u306e\u6700\u9069\u5316\u304c\u307b\u304b\u306e\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306b\u3082\u52b9\u679c\u7684\u306b\u9069\u7528\u3067\u304d\u308b\u3053\u3068\u3092\u793a\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n

S-GAN<\/h2>\n

\u6559\u5e2b\u3042\u308a\u5b66\u7fd2\u3067\u306f\u3001\u5927\u91cf\u306e\u30e9\u30d9\u30eb\u4ed8\u304d\u30c7\u30fc\u30bf\u304c\u5fc5\u8981\u3067\u3059\u3002\u30e9\u30d9\u30eb\u4ed8\u3051\u3084\u30a2\u30ce\u30c6\u30fc\u30b7\u30e7\u30f3\u306f\u5c02\u9580\u5bb6\u304c\u624b\u4f5c\u696d\u3067\u884c\u308f\u306a\u3051\u308c\u3070\u306a\u3089\u306a\u3044\u305f\u3081\u3001\u624b\u9593\u3068\u30b3\u30b9\u30c8\u304c\u304b\u304b\u308a\u307e\u3059\u3002\u534a\u6559\u5e2b\u3042\u308a\u5b66\u7fd2\u306f\u3001\u30e9\u30d9\u30eb\u4ed8\u304d\u30c7\u30fc\u30bf\u3068\u30e9\u30d9\u30eb\u306a\u3057\u30c7\u30fc\u30bf\u306e\u4e21\u65b9\u3092\u4f7f\u3063\u3066\u30e2\u30c7\u30eb\u3092\u5b66\u7fd2\u3055\u305b\u308b\u624b\u6cd5\u3067\u3059\u3002\u901a\u5e38\u3001\u30e9\u30d9\u30eb\u4ed8\u304d\u306e\u30c7\u30fc\u30bf\u30dd\u30a4\u30f3\u30c8\u306e\u6570\u306f\u3001\u30e9\u30d9\u30eb\u306a\u3057\u306e\u30c7\u30fc\u30bf\u30dd\u30a4\u30f3\u30c8\u3088\u308a\u3082\u304b\u306a\u308a\u5c11\u306a\u304f\u306a\u308a\u307e\u3059\u3002\u534a\u6559\u5e2b\u3042\u308a\u5b66\u7fd2\u3067\u306f\u3001\u30c7\u30fc\u30bf\u306e\u30d1\u30bf\u30fc\u30f3\u3084\u50be\u5411\u3092\u5229\u7528\u3057\u3066\u5206\u985e\u3092\u884c\u3044\u307e\u3059\u3002<\/p>\n

S-GAN \u306f\u3001\u751f\u6210\u30e2\u30c7\u30eb\u3092\u7528\u3044\u3066\u30c7\u30fc\u30bf\u30dd\u30a4\u30f3\u30c8\u3092\u751f\u6210\u3059\u308b\u3053\u3068\u3067\u3001\u81a8\u5927\u306a\u91cf\u306e\u5b66\u7fd2\u30c7\u30fc\u30bf\u306e\u5fc5\u8981\u6027\u306b\u53d6\u308a\u7d44\u3093\u3067\u3044\u307e\u3059\u3002Generative Adversarial Network (GAN) \u306f\u3001\u5927\u898f\u6a21\u306a\u975e\u30e9\u30d9\u30eb\u30c7\u30fc\u30bf\u3092\u7528\u3044\u3066\u3001\u753b\u50cf\u8b58\u5225\u30e2\u30c7\u30eb\u3092\u4ecb\u3057\u3066\u753b\u50cf\u751f\u6210\u30e2\u30c7\u30eb\u3092\u5b66\u7fd2\u3059\u308b\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u3067\u3059\u3002GAN \u306f\u3001\u751f\u6210\u30e2\u30c7\u30eb\u3068\u8b58\u5225\u30e2\u30c7\u30eb\u306e 2 \u3064\u306e\u30e2\u30c7\u30eb\u3067\u69cb\u6210\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u4e21\u8005\u306f\u30bc\u30ed\u30b5\u30e0\u30b2\u30fc\u30e0\u3067\u4e00\u7dd2\u306b\u8a13\u7df4\u3055\u308c\u307e\u3059\u3002\u751f\u6210\u5668\u306e\u5f79\u5272\u306f\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u5b58\u5728\u3059\u308b\u3082\u306e\u3068\u540c\u3058\u3088\u3046\u306a\u30c7\u30fc\u30bf\u3092\u751f\u6210\u3059\u308b\u3053\u3068\u3067\u3059\u3002\u8b58\u5225\u5668\u306e\u5f79\u5272\u306f\u3001\u751f\u6210\u3055\u308c\u305f\u30c7\u30fc\u30bf\u306e\u4e2d\u304b\u3089\u5b9f\u969b\u306e\u30c7\u30fc\u30bf\u3092\u8b58\u5225\u3059\u308b\u3053\u3068\u3067\u3059\u3002S-GAN \u306f\u3001GAN \u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc\u3092\u62e1\u5f35\u3057\u3001\u5206\u985e\u30bf\u30b9\u30af\u306b\u6559\u5e2b\u3042\u308a\u8b58\u5225\u5668 (\u5206\u985e\u5668) \u3092\u8ffd\u52a0\u3057\u305f\u3082\u306e\u3067\u3059 (\u56f31)\u3002\u3053\u308c\u306b\u3088\u308a\u3001\u898b\u305f\u3053\u3068\u306e\u306a\u3044\u30c7\u30fc\u30bf\u306b\u5bfe\u3057\u3066\u3082\u5341\u5206\u306b\u4e00\u822c\u5316\u3067\u304d\u308b\u5206\u985e\u5668\u304c\u5f97\u3089\u308c\u307e\u3059\u3002<\/p>\n

\"\"<\/p>\n

\u56f31. S-GAN \u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u30fc<\/p>\n

\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u3068\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2<\/h2>\n

\u3053\u308c\u3089\u306e\u5b9f\u9a13\u3067\u306f\u4ee5\u4e0b\u306e 3 \u3064\u306e\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> oneAPI \u30c4\u30fc\u30eb\u30ad\u30c3\u30c8 (v2021.1) \u3092\u4f7f\u7528\u3057\u307e\u3057\u305f\u3002<\/p>\n

    \n
  1. \u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> oneAPI \u30d9\u30fc\u30b9\u30fb\u30c4\u30fc\u30eb\u30ad\u30c3\u30c8<\/a><\/li>\n
  2. \u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> oneAPI HPC \u30c4\u30fc\u30eb\u30ad\u30c3\u30c8<\/a><\/li>\n
  3. \u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> oneAPI AI Analytics \u30c4\u30fc\u30eb\u30ad\u30c3\u30c8 (\u82f1\u8a9e)<\/a><\/li>\n<\/ol>\n

    \u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> oneAPI \u30de\u30b9\u30fb\u30ab\u30fc\u30cd\u30eb\u30fb\u30e9\u30a4\u30d6\u30e9\u30ea\u30fc<\/a>\u3068\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> oneAPI \u30c7\u30fc\u30bf \u30a2\u30ca\u30ea\u30c6\u30a3\u30af\u30b9 \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc<\/a>\u3092\u4f7f\u7528\u3057\u305f\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> \u30c7\u30a3\u30b9\u30c8\u30ea\u30d3\u30e5\u30fc\u30b7\u30e7\u30f3\u306e Python<\/a> (v3.6) \u306f\u3001Python \u306e\u30b3\u30a2\u6570\u5024\u30d1\u30c3\u30b1\u30fc\u30b8\u3092\u9ad8\u901f\u5316\u3059\u308b\u305f\u3081\u306b\u4f7f\u7528\u3055\u308c\u307e\u3057\u305f\u3002S-GAN \u30e2\u30c7\u30eb\u306f\u3001\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> Optimization for TensorFlow* (v1.15) (\u82f1\u8a9e) <\/a>\u3092\u4f7f\u7528\u3057\u3066\u5b9f\u88c5\u3055\u308c\u307e\u3057\u305f\u3002\u5206\u6563\u5b66\u7fd2\u306b\u306f Horovod (v0.20.2) \u3092\u4f7f\u7528\u3057\u307e\u3057\u305f\u3002Horovod \u306f\u30ce\u30fc\u30c9\u5185\u901a\u4fe1\u306b MPI \u3092\u4f7f\u7528\u3057\u3066\u3044\u308b\u305f\u3081\u30012\u3064\u306e MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc (\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc<\/a>\u3068 Open MPI (v4.0.5)) \u306e\u6027\u80fd\u3092\u6bd4\u8f03\u3057\u307e\u3057\u305f\u3002\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u306e\u89e3\u6790\u306b\u306f\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> VTune™ \u30d7\u30ed\u30d5\u30a1\u30a4\u30e9\u30fc\u3092\u4f7f\u7528\u3057\u307e\u3057\u305f\u3002\u3059\u3079\u3066\u306e\u30c6\u30b9\u30c8\u306f\u3001\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> Xeon\u00ae<\/sup> Platinum 8260L \u30d7\u30ed\u30bb\u30c3\u30b5\u30fc (\u82f1\u8a9e) <\/a>\u3092\u4f7f\u7528\u3057\u3001100Gbps \u3067\u52d5\u4f5c\u3059\u308b\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> Omni-Path \u30d5\u30a1\u30d6\u30ea\u30c3\u30af (\u82f1\u8a9e)\u00a0<\/a>\u3092\u4ecb\u3057\u3066\u63a5\u7d9a\u3055\u308c\u305f\u30a4\u30f3\u30c6\u30eb\u306e Endeavor<\/a> \u30af\u30e9\u30b9\u30bf\u30fc\u4e0a\u3067\u5b9f\u884c\u3055\u308c\u307e\u3057\u305f\uff08\u56f32\uff09\u3002\"\"<\/p>\n

    \u56f32\uff0e\u30ce\u30fc\u30c9\u306e\u69cb\u6210 ($ cpuinfo -g)<\/p>\n

    \u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u65b9\u6cd5<\/h2>\n

    \u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3092\u6700\u9069\u5316\u3059\u308b\u524d\u306b\u3001\u30d9\u30fc\u30b9\u30e9\u30a4\u30f3\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u6e2c\u5b9a\u3059\u308b\u3053\u3068\u306f\u826f\u3044\u7fd2\u6163\u3067\u3059\u3002\u30d7\u30ed\u30d5\u30a1\u30a4\u30ea\u30f3\u30b0\u30fb\u30c4\u30fc\u30eb\u306f\u3001\u30b9\u30ec\u30c3\u30c7\u30a3\u30f3\u30b0\u3001\u30d9\u30af\u30c8\u30eb\u5316\u3001I\/O\u3001\u30de\u30eb\u30c1\u30ce\u30fc\u30c9\u30fb\u30b3\u30df\u30e5\u30cb\u30b1\u30fc\u30b7\u30e7\u30f3\u306a\u3069\u3001\u6700\u9069\u5316\u306e\u53ef\u80fd\u6027\u304c\u3042\u308b\u9818\u57df\u3092\u7279\u5b9a\u3059\u308b\u306e\u306b\u5f79\u7acb\u3061\u307e\u3059\u3002\u3053\u3053\u3067\u306f\u3001\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> VTune™ \u30d7\u30ed\u30d5\u30a1\u30a4\u30e9\u30fc\u306e\u300c\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u30fb\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u30fb\u30b9\u30ca\u30c3\u30d7\u30b7\u30e7\u30c3\u30c8 (\u82f1\u8a9e)<\/a>\u300d\u3092\u4f7f\u7528\u3057\u307e\u3057\u305f\u3002Application Performance Snapshot \u30d7\u30ed\u30d5\u30a1\u30a4\u30eb\u306b\u3088\u308b\u3068\u3001\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306b\u3088\u3063\u3066\u3042\u307e\u308a\u306b\u3082\u591a\u304f\u306e\u30b9\u30ec\u30c3\u30c9\u304c\u751f\u6210\u3055\u308c\u3066\u3044\u307e\u3057\u305f\u3002\u3044\u304f\u3064\u304b\u306e\u30b9\u30ec\u30c3\u30c9\u306f OpenMP* \u304b\u3089\u306e\u3082\u306e\u3067\u3001\u307b\u304b\u306e\u30b9\u30ec\u30c3\u30c9\u306f TensorFlow* \u304c\u547c\u3073\u51fa\u3059 Eigen \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u304b\u3089\u306e\u3082\u306e\u3067\u3057\u305f\u3002OpenMP* \u3068 Eigen \u306e\u30b9\u30ec\u30c3\u30c9\u6570\u306f\u3001\u30ce\u30fc\u30c9\u3042\u305f\u308a\u306e\u8ad6\u7406\u30b3\u30a2\u6570\u3092\u8d85\u3048\u3066\u304a\u308a\u3001\u30ea\u30bd\u30fc\u30b9\u306e\u30aa\u30fc\u30d0\u30fc\u30b5\u30d6\u30b9\u30af\u30ea\u30d7\u30b7\u30e7\u30f3\u304c\u767a\u751f\u3057\u3066\u3001\u901a\u5e38\u306f\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u304c\u4f4e\u4e0b\u3057\u307e\u3059\u3002<\/p>\n

    \u6700\u9069\u306a\u30b9\u30ec\u30c3\u30c9\u6570\u306e\u9078\u629e<\/strong><\/p>\n

    \u6700\u521d\u306b\u884c\u3063\u305f\u306e\u306f\u3001\u6700\u9069\u306a\u30b9\u30ec\u30c3\u30c9\u6570\u3092\u898b\u3064\u3051\u308b\u3053\u3068\u3067\u3057\u305f\u30021 \u3064\u306e\u8a08\u7b97\u30ce\u30fc\u30c9\u4e0a\u3067\u74b0\u5883\u5909\u6570 OMP_NUM_THREADS \u3092\u8a2d\u5b9a\u3057\u3001\u4f4e\u89e3\u50cf\u5ea6\u306e\u753b\u50cf (256\u00d7256\u30d4\u30af\u30bb\u30eb) \u3092\u4f7f\u7528\u3057\uff0c\u6642\u9593\u3092\u7bc0\u7d04\u3059\u308b\u305f\u3081\u306b\u30a8\u30dd\u30c3\u30af\u6570 (2\u56de) \u3092\u5236\u9650\u3057\u3066\u3001\u3055\u307e\u3056\u307e\u306a\u6570\u306e OpenMP* \u30b9\u30ec\u30c3\u30c9\u3092\u30c6\u30b9\u30c8\u3057\u307e\u3057\u305f\u3002\u305d\u306e\u7d50\u679c\u300150 \u30b9\u30ec\u30c3\u30c9\u3067 1 \u79d2\u3042\u305f\u308a 19.67 \u679a\u3068\u3044\u3046\u6700\u9ad8\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u304c\u5f97\u3089\u308c\u307e\u3057\u305f\uff08\u56f33\uff09\u3002<\/p>\n

    \"\"<\/p>\n

    \u56f33\uff0e\u6700\u9069\u306a OpenMP* \u30b9\u30ec\u30c3\u30c9\u306e\u6570\u3092\u898b\u3064\u3051\u308b<\/p>\n

    \u6b21\u306b\u3001TensorFlow* \u306e\u30b9\u30ec\u30c3\u30c7\u30a3\u30f3\u30b0 API \u3092\u4f7f\u3063\u3066\u3001\u30b9\u30ec\u30c3\u30c9\u6570\u3092\u5236\u5fa1\u3057\u307e\u3057\u305f\u3002inter_op_parallelism_threads \u306e\u5024\u306f\u3001\u72ec\u7acb\u3057\u305f\u30ce\u30f3\u30d6\u30ed\u30c3\u30ad\u30f3\u30b0\u6f14\u7b97\u3067\u4f7f\u7528\u3059\u308b\u30b9\u30ec\u30c3\u30c9\u6570\u3092\u6307\u5b9a\u3057\u3001intra_op_parallelism_threads \u306e\u5024\u306f\u3001\u884c\u5217\u306e\u4e57\u7b97\u3084\u30ea\u30c0\u30af\u30b7\u30e7\u30f3\u306a\u3069\u306e\u500b\u5225\u6f14\u7b97\u3067\u4f7f\u7528\u3059\u308b\u30b9\u30ec\u30c3\u30c9\u6570\u3092\u6307\u5b9a\u3057\u307e\u3059\u3002\u56f34 \u306f\u3001\u3053\u306e 2 \u3064\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u30fc\u306e\u5024\u3092\u5909\u3048\u305f\u3068\u304d\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u793a\u3057\u3066\u3044\u307e\u3059\u3002\u6fc3\u3044\u7dd1\u8272\u306e\u30d6\u30ed\u30c3\u30af\u306f\u826f\u597d\u306a\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u793a\u3057\u3001\u6fc3\u3044\u8d64\u8272\u306e\u30d6\u30ed\u30c3\u30af\u306f\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u304c\u4f4e\u3044\u3053\u3068\u3092\u793a\u3057\u3066\u3044\u307e\u3059\u3002\u767d\u306e\u30d6\u30ed\u30c3\u30af\u306f\u5b9f\u884c\u306b\u5931\u6557\u3057\u305f\u3053\u3068\u3092\u793a\u3057\u3066\u3044\u307e\u3059\u3002\"\"<\/p>\n

    \u56f34\uff0einter_op_parallelism_threads \u3068 intra_op_parallelism_threads \u306e\u6700\u9069\u5024\u3092\u6c42\u3081\u308b<\/p>\n

    \u30ce\u30fc\u30c9\u3042\u305f\u308a\u306e MPI \u30e9\u30f3\u30af\u304c 1 \u3064\u306e\u5834\u5408\u3001inter_op_parallelism_threads \u3068 intra_op_parallelism_threads \u306e\u6700\u9069\u5024\u306f\u305d\u308c\u305e\u308c 0 \u3068 45 \u3067\u3042\u308a\u3001\u3053\u308c\u306f 1 \u79d2\u3042\u305f\u308a 12.34 \u679a\u306e\u753b\u50cf\u306b\u76f8\u5f53\u3057\u307e\u3059\u3002\u3053\u308c\u306f\u3001OMP_NUM_THREADS \u3092\u4f7f\u7528\u3057\u3066\u9054\u6210\u3057\u305f 19.67 \u753b\u50cf\/\u79d2\u306e\u5024\u3088\u308a\u3082\u4f4e\u304b\u3063\u305f\u305f\u3081 (\u56f3 3)\u3001TensorFlow* \u306e\u30b9\u30ec\u30c3\u30c9 API \u3092\u4f7f\u7528\u3059\u308b\u4ee3\u308f\u308a\u306b\u3001\u3053\u306e\u74b0\u5883\u5909\u6570\u3092\u4f7f\u7528\u3057\u3066\u30b9\u30ec\u30c3\u30c9\u6570\u3092\u5236\u5fa1\u3057\u307e\u3057\u305f\u3002<\/p>\n

    \u30ce\u30fc\u30c9\u3042\u305f\u308a\u306e\u6700\u9069\u306a MPI \u30e9\u30f3\u30af\u6570\u306e\u9078\u629e<\/h2>\n

    \u3053\u306e\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3067\u306f MPI \u3068 OpenMP* \u306e\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u4e26\u5217\u3092\u4f7f\u7528\u3057\u3066\u3044\u308b\u305f\u3081\uff0c\u30ce\u30fc\u30c9\u3042\u305f\u308a\u306e OpenMP* \u30b9\u30ec\u30c3\u30c9\u3068 MPI \u30e9\u30f3\u30af\u306e\u6700\u9069\u306a\u7d44\u307f\u5408\u308f\u305b\u3092\u898b\u3064\u3051\u308b\u3053\u3068\u304c\u91cd\u8981\u3067\u3059\u3002\u56f3 3 \u306e\u30c6\u30b9\u30c8\u3092\u7e70\u308a\u8fd4\u3057\u884c\u3044\u307e\u3057\u305f\u304c (\u4f8b\u3048\u3070\u30011\u2266OMP_NUM_THREADS\u226696) \u30ce\u30fc\u30c9\u3042\u305f\u308a\u306e MPI \u30e9\u30f3\u30af\u6570\u3092\u5909\u5316\u3055\u305b\u307e\u3057\u305f(21\u2266ppn\u226624\u3001ppn \u306f\u30ce\u30fc\u30c9\u3042\u305f\u308a\u306e\u30e9\u30f3\u30af\u6570\u3001\u56f3 5 \u53c2\u7167)\u3002\u305d\u306e\u7d50\u679c\u3001\u5404\u30ce\u30fc\u30c9\u306b 8 \u3064\u306e MPI \u30e9\u30f3\u30af\u3068 9 \u3064\u306e OpenMP* \u30b9\u30ec\u30c3\u30c9\u3092\u914d\u7f6e\u3057\u305f\u5834\u5408\u306b\uff0c\u6700\u9ad8\u306e\u6027\u80fd (26.3\u679a\/\u79d2) \u304c\u5f97\u3089\u308c\u307e\u3057\u305f\u3002<\/p>\n

    \"\"<\/p>\n

    \u56f35\uff0e1 \u3064\u306e\u30ce\u30fc\u30c9\u3067 OpenMP* \u30b9\u30ec\u30c3\u30c9\u3068 MPI \u30e9\u30f3\u30af\u306e\u6700\u9069\u306a\u7d44\u307f\u5408\u308f\u305b\u3092\u898b\u3064\u3051\u308b<\/p>\n

    \u30e1\u30e2\u30ea\u30fc\u30ea\u30fc\u30af\u306e\u514b\u670d<\/h2>\n

    \u3053\u308c\u307e\u3067\u306e\u5b9f\u9a13\u3067\u306f\u3001\u4f4e\u89e3\u50cf\u5ea6\u306e\u753b\u50cf\u3092\u4f7f\u7528\u3057\u307e\u3057\u305f\u3002\u9ad8\u89e3\u50cf\u5ea6\u306e\u753b\u50cf (1360\u00d71360 \u30d4\u30af\u30bb\u30eb) \u3067\u306f\u3001\u5404\u30ef\u30fc\u30ab\u30fc\u30d7\u30ed\u30bb\u30b9\u306e\u30e1\u30e2\u30ea\u30fc\u30fb\u30d5\u30c3\u30c8\u30d7\u30ea\u30f3\u30c8\u304c\u5927\u5e45\u306b\u5897\u52a0\u3059\u308b\u305f\u3081\u3001\u30ce\u30fc\u30c9\u3042\u305f\u308a 8 \u3064\u306e MPI \u30e9\u30f3\u30af\u3092\u5b9f\u884c\u3059\u308b\u3068\u3001\u30e1\u30e2\u30ea\u30fc\u4e0d\u8db3\u30a8\u30e9\u30fc\u304c\u767a\u751f\u3057\u307e\u3059\u30021 \u3000\u3064\u306e\u30e9\u30f3\u30af\u306e\u30e1\u30e2\u30ea\u30fc\u30fb\u30d5\u30c3\u30c8\u30d7\u30ea\u30f3\u30c8\u306f 129GB \u4ee5\u4e0a\u306b\u306a\u308b\u3053\u3068\u304c\u308f\u304b\u308a\u307e\u3057\u305f\u3002\u5404\u30ce\u30fc\u30c9\u306b\u306f\u7d04 200GB \u306e DRAM \u3057\u304b\u306a\u304f\u3001\u30ce\u30fc\u30c9\u3054\u3068\u306b 1 \u3064\u306e\u30e9\u30f3\u30af\u3057\u304b\u8d77\u52d5\u3067\u304d\u307e\u305b\u3093\u3067\u3057\u305f\u3002\u3053\u308c\u306f\u30c7\u30e5\u30a2\u30eb\u30bd\u30b1\u30c3\u30c8\u306e\u30de\u30b7\u30f3\u3067\u306f (NUMA (Non-Uniform Memory Access)\u306e\u554f\u984c\u304b\u3089) \u6700\u9069\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3067\u3057\u305f\u3002\u3053\u306e\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3092 Python* \u30e6\u30fc\u30c6\u30a3\u30ea\u30c6\u30a3\u30fc\u306e\u30e1\u30e2\u30ea\u30fc\u30fb\u30d7\u30ed\u30d5\u30a1\u30a4\u30e9\u30fc (\u82f1\u8a9e)\u00a0<\/a>\u3067\u5b9f\u884c\u3059\u308b\u3068\u3001\u30e1\u30e2\u30ea\u30fc\u30ea\u30fc\u30af\u304c\u898b\u3064\u304b\u308a\u307e\u3057\u305f\u3002\u7d04 56GB \u306e\u30e1\u30e2\u30ea\u30fc\u304c\u30e2\u30c7\u30eb\u306e\u30d5\u30a9\u30ef\u30fc\u30c9\u30d1\u30b9\u4e2d\u306b\u89e3\u653e\u3055\u308c\u3066\u3044\u307e\u305b\u3093\u3067\u3057\u305f (\u56f36\u3001\u4e0a)\u3002\u3053\u308c\u306f\u3001TensorFlow* \u306e Keras Model.predict \u30e1\u30bd\u30c3\u30c9\u306e\u65e2\u77e5\u306e\u30d0\u30b0 (TensorFlow* issue #33009 (\u82f1\u8a9e)<\/a> \u304a\u3088\u3073 Keras issue #13118 (\u82f1\u8a9e)<\/a>) \u3067\u3042\u308b\u3053\u3068\u304c\u5224\u660e\u3057\u307e\u3057\u305f\u3002TensorFlow* \u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u30fc\u304b\u3089\u306e\u63d0\u8a00\u306b\u57fa\u3065\u304d\u3001Model.predict \u3092 Model.predict_on_batch \u306b\u7f6e\u304d\u63db\u3048\u305f\u3068\u3053\u308d\u3001\u30d7\u30ed\u30bb\u30b9\u3054\u3068\u306e\u30e1\u30e2\u30ea\u30fc\u6d88\u8cbb\u91cf\u304c\u5168\u4f53\u7684\u306b\u6e1b\u5c11\u3057\u307e\u3057\u305f (\u56f36\u3001\u4e0b)\u3002\u30e1\u30e2\u30ea\u30fc\u30ea\u30fc\u30af\u304c\u4fee\u6b63\u3055\u308c\u3066\u3082\u3001\u9ad8\u89e3\u50cf\u5ea6\u753b\u50cf\u306e\u5834\u5408\u306f\u30ce\u30fc\u30c9\u3042\u305f\u308a 2 \u3064\u306e\u30e9\u30f3\u30af\u306b\u5236\u9650\u3055\u308c\u3066\u3044\u307e\u3057\u305f\u3002\u3053\u308c\u4ee5\u4e0a\u306e\u30e1\u30e2\u30ea\u30fc\u6d88\u8cbb\u91cf\u306e\u6700\u9069\u5316\u306f\u884c\u3044\u307e\u305b\u3093\u3067\u3057\u305f\u304c\u3001\u53ef\u80fd\u6027\u306f\u3042\u308a\u307e\u3059\u3002<\/p>\n

    \"\"<\/p>\n

    \u56f36\uff0e\u30e1\u30e2\u30ea\u30fc\u30fb\u30d7\u30ed\u30d5\u30a1\u30a4\u30e9\u30fc\u3067\u691c\u51fa\u3055\u308c\u305f\u30e1\u30e2\u30ea\u30fc\u30ea\u30fc\u30af<\/p>\n

    \u305d\u306e\u4ed6\u306e OpenMP* \u304a\u3088\u3073\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u74b0\u5883\u5909\u6570\u306e\u8a2d\u5b9a<\/h2>\n

    \u307e\u305f\u3001\u307b\u304b\u306e 3 \u3064\u306e\u74b0\u5883\u5909\u6570 (KMP_BLOCKTIME\u3001KMP_AFFINITY\u3001I_MPI_PIN_DOMAIN) \u306e\u6700\u9069\u306a\u8a2d\u5b9a\u306b\u3064\u3044\u3066\u3082\u691c\u8a0e\u3057\u307e\u3057\u305f\u3002\u3053\u308c\u307e\u3067\u306e\u63a8\u5968\u6027\u80fd (\u82f1\u8a9e)\u00a0<\/a>\u306b\u57fa\u3065\u3044\u3066\u3001\u8868 1 \u306b\u793a\u3059 6 \u3064\u306e\u5b9f\u9a13\u3092\u884c\u3044\u307e\u3057\u305f\u3002\u306a\u304a\u3001\u30bb\u30c3\u30c8 #5 \u306e\u74b0\u5883\u5909\u6570\u306e\u8a2d\u5b9a\u306f\u3001\u307b\u304b\u306e 5 \u3064\u306e\u30bb\u30c3\u30c8\u3067\u3082\u4f7f\u7528\u3057\u307e\u3057\u305f\u3002\u30bb\u30c3\u30c8 #3 \u304c\u6700\u3082\u512a\u308c\u305f\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u6027\u80fd\u3092\u793a\u3057\u307e\u3057\u305f (\u56f37\uff09\u3002<\/p>\n

    \u88681. \u305d\u306e\u4ed6\u306e OpenMP* \u304a\u3088\u3073 \u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc \u306e\u74b0\u5883\u5909\u6570\u306e\u5b9f\u9a13\u7684\u8a2d\u5b9a<\/p>\n

    \"\"<\/p>\n

    \"\"<\/p>\n

    \u56f37. \u8868 1 \u306e 6 \u3064\u306e\u5b9f\u9a13\u306e\u6027\u80fd\u6bd4\u8f03<\/p>\n

    MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u306e\u9078\u3073\u65b9<\/h2>\n

    \u6b21\u306b\u3001\u3053\u308c\u307e\u3067\u306e\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u5b9f\u9a13\u3067\u5f97\u3089\u308c\u305f\u6700\u9069\u306a\u8a2d\u5b9a\u3092\u96c6\u3081\u3001\u9ad8\u89e3\u50cf\u5ea6\u306e\u753b\u50cf(1360\u00d71360 \u30d4\u30af\u30bb\u30eb) \u3092\u5bfe\u8c61\u306b\u3001\u30ce\u30fc\u30c9\u3042\u305f\u308a 2 \u3064\u306e MPI \u30e9\u30f3\u30af\u3068 33 \u306e OpenMP* \u30b9\u30ec\u30c3\u30c9\u3092\u7528\u3044\u3066\u3001\u5408\u8a08 100 \u306e\u30a8\u30dd\u30c3\u30af\u3092\u4f5c\u6210\u3057\u3001\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u3068 Open MPI \u306e\u6027\u80fd\u3092\u6bd4\u8f03\u3057\u307e\u3057\u305f\u3002\u4f7f\u7528\u3057\u305f\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u306e\u30b3\u30de\u30f3\u30c9\u30e9\u30a4\u30f3\u306f\u4ee5\u4e0b\u306e\u901a\u308a\u3067\u3059\u3002<\/p>\n

    \"\"<\/p>\n

    \u3053\u308c\u3089\u306e\u307b\u307c\u540c\u7b49\u306e\u30e9\u30f3\u30bf\u30a4\u30e0\u8a2d\u5b9a\u3092 Open MPI \u306b\u3082\u63a1\u7528\u3057\u307e\u3057\u305f\u3002<\/p>\n

    \"\"<\/p>\n

    \u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u306f\u3001\u30b7\u30f3\u30b0\u30eb\u30ce\u30fc\u30c9\u304a\u3088\u3073\u30de\u30eb\u30c1\u30ce\u30fc\u30c9\u306e\u30b7\u30ca\u30ea\u30aa\u3067 Open MPI \u3092\u4e0a\u56de\u308a\u307e\u3057\u305f\uff08\u56f38\uff09\u3002<\/p>\n

    \"\"<\/p>\n

    \u56f38. \u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u3068 Open MPI \u306e\u6027\u80fd\u6bd4\u8f03<\/p>\n

    1 \u79d2\u9593\u306e\u753b\u50cf\u6570\u306b\u57fa\u3065\u304f\u4e26\u5217\u30b9\u30b1\u30fc\u30e9\u30d3\u30ea\u30c6\u30a3\u30fc\u306f\u7dda\u5f62\u306b\u898b\u3048\u307e\u3059\u304c\u3001\u6642\u9593\u306b\u57fa\u3065\u304f\u30b9\u30b1\u30fc\u30e9\u30d3\u30ea\u30c6\u30a3\u30fc\u306f\u52a3\u7dda\u5f62\u3067\u3042\u308a\u3001\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306e\u975e\u52b9\u7387\u6027\u3092\u793a\u3057\u3066\u3044\u307e\u3059\u3002\u3055\u3089\u306b\u89e3\u6790\u3092\u9032\u3081\u308b\u3068\u3001\u8a55\u4fa1\u306e\u983b\u5ea6\u304c\u30ce\u30fc\u30c9\u7dcf\u6570\u306b\u5fdc\u3058\u3066\u6b63\u3057\u304f\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u3055\u308c\u3066\u3044\u306a\u3044\u3053\u3068\u304c\u308f\u304b\u308a\u307e\u3057\u305f\u3002\u30ce\u30fc\u30c9\u6570\u306e\u5897\u52a0\u306b\u4f34\u3044\u3001\u8a55\u4fa1\u306e\u983b\u5ea6\u3082\u5897\u52a0\u3057\u307e\u3057\u305f\u3002\u307e\u305f\u3001\u3053\u308c\u3089\u306e\u8a55\u4fa1\u306f MPI \u306e\u30de\u30b9\u30bf\u30fc\u30e9\u30f3\u30af\u304c\u884c\u3063\u3066\u3044\u305f\u305f\u3081\u3001\u30de\u30b9\u30bf\u30fc\u30e9\u30f3\u30af\u304c\u8a55\u4fa1\u3092\u7d42\u3048\u308b\u307e\u3067\u307b\u304b\u306e\u30e9\u30f3\u30af\u306f\u505c\u6ede\u3057\u3066\u3044\u307e\u3057\u305f\u3002\u3053\u306e\u554f\u984c\u3092\u89e3\u6c7a\u3059\u308b\u305f\u3081\u306b\u3001\u30ce\u30fc\u30c9\u6570\u306b\u95a2\u308f\u3089\u305a\u300115 \u56de\u306e\u5b66\u7fd2\u53cd\u5fa9\u3054\u3068\u306b\u8a55\u4fa1\u30b9\u30c6\u30c3\u30d7\u3092\u5b9f\u884c\u3057\u307e\u3057\u305f (\u56f39\u300110)\u3002\u3053\u306e\u5909\u66f4\u306b\u3088\u308b\u30e2\u30c7\u30eb\u306e\u7cbe\u5ea6\u3078\u306e\u5f71\u97ff\u306f\u3042\u308a\u307e\u305b\u3093\u3067\u3057\u305f\u3002<\/p>\n

    \"\"<\/p>\n

    \u56f39. \u6700\u9069\u5316\u524d (\u5de6) \u3068\u6700\u9069\u5316\u5f8c (\u53f3) \u306e\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u3068\u8a55\u4fa1\u306b\u8cbb\u3084\u3057\u305f\u7dcf\u6642\u9593\u306e\u5272\u5408<\/p>\n

    \"\"<\/p>\n

    \u56f310. \u7dda\u5f62\u30bf\u30a4\u30e0\u30fb\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u306b\u3088\u308b\u30de\u30eb\u30c1\u30ce\u30fc\u30c9\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9<\/p>\n

    \u307e\u3068\u3081<\/h2>\n

    \u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001S-GAN \u306e\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u6027\u80fd\u3092\u5411\u4e0a\u3055\u305b\u308b\u305f\u3081\u306b\u3001\u6bb5\u968e\u7684\u306a\u6700\u9069\u5316\u30a2\u30d7\u30ed\u30fc\u30c1\u3092\u7d39\u4ecb\u3057\u307e\u3057\u305f\u3002\u30e9\u30f3\u30bf\u30a4\u30e0\u304a\u3088\u3073\u30bd\u30fc\u30b9\u30fb\u30b3\u30fc\u30c9\u30d9\u30fc\u30b9\u306e\u6700\u9069\u5316\u3092\u884c\u3044\u3001\u30e1\u30e2\u30ea\u30fc\u306e\u554f\u984c\u3001\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u306e\u554f\u984c\u3001\u304a\u3088\u3073\u30b7\u30f3\u30b0\u30eb\u30ce\u30fc\u30c9\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u306e\u975e\u52b9\u7387\u6027\u3092\u89e3\u6c7a\u3057\u307e\u3057\u305f\u3002\u30de\u30eb\u30c1\u30ce\u30fc\u30c9\u30fb\u30af\u30e9\u30b9\u30bf\u30fc\u4e0a\u3067\u306e S-GAN \u30e2\u30c7\u30eb\u306e\u5206\u6563\u5b66\u7fd2\u3092\u5b9f\u88c5\u3059\u308b\u305f\u3081\u306bHorovod \u3092\u4f7f\u7528\u3057\u307e\u3057\u305f\u30022 \u7a2e\u985e\u306e MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc (\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u304a\u3088\u3073 Open MPI) \u3092\u6bd4\u8f03\u3057\u307e\u3057\u305f\u3002\"\"<\/p>\n

    \u56f311. \u30a4\u30f3\u30c6\u30eb\u00ae<\/sup> MPI \u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u3067\u5f97\u3089\u308c\u305f\u6700\u7d42\u7684\u306a\u30b9\u30d4\u30fc\u30c9\u30a2\u30c3\u30d7<\/p>\n

    \u56f311 \u306f\u3001\u4eca\u56de\u306e\u6700\u9069\u5316\u306b\u3088\u308b\u6027\u80fd\u5411\u4e0a\u3092\u307e\u3068\u3081\u305f\u3082\u306e\u3067\u3059\u3002\u30b7\u30f3\u30b0\u30eb\u30ce\u30fc\u30c9\u306e\u6700\u9069\u5316\u3067\u306f 2 \u500d\u3001\u30de\u30eb\u30c1\u30ce\u30fc\u30c9\u306e\u6700\u9069\u5316\u3067\u306f\u3055\u3089\u306b 8 \u500d\u306e\u30b9\u30d4\u30fc\u30c9\u30a2\u30c3\u30d7\u3092\u5b9f\u73fe\u3057\u3001\u30e2\u30c7\u30eb\u306e\u7cbe\u5ea6\u3092\u843d\u3068\u3059\u3053\u3068\u306a\u304f\u3001\u5168\u4f53\u306e\u89e3\u7b54\u307e\u3067\u306e\u6642\u9593\u3092 16 \u5206\u306e 1 \u306b\u77ed\u7e2e\u3057\u307e\u3057\u305f\u3002<\/p>\n

    \n

    \u5f53\u8a18\u4e8b\u3067\u3054\u7d39\u4ecb\u3057\u3066\u3044\u308b\u30a4\u30f3\u30c6\u30eb\u00ae<\/sup>\u30d9\u30fc\u30b9\u30fb\u30c4\u30fc\u30eb\u30ad\u30c3\u30c8\u306b\u95a2\u3059\u308b\u8a73\u7d30\u306f \u3053\u3061\u3089<\/a>\u304b\u3089\u3054\u53c2\u7167\u304f\u3060\u3055\u3044\u3002<\/p>\n

    \n

    \u53c2\u7167\u8a18\u4e8b:
    \n    Optimize Distributed AI Training Using Intel\u00ae<\/sup> oneAPI Toolkits<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

    \u30a4\u30f3\u30af\u30ea\u30e1\u30f3\u30bf\u30eb\u30fb\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3067\u5927\u5e45\u306a\u6027\u80fd\u5411\u4e0a\u3092\u5b9f […]<\/a><\/p>\n<\/div>","protected":false},"author":1,"featured_media":1128,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[12,44,40,43,38,31,14,21,19,34,41,35,42,37,36,39],"class_list":["post-1124","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-1","tag-ai","tag-bert","tag-gan","tag-gpt-3","tag-hpc","tag-mpi","tag-oneapi","tag-onedal","tag-onemkl","tag-python","tag-resnet","tag-tensorflow","tag-vgg","tag-vtune","tag-36","tag-39"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/posts\/1124","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/comments?post=1124"}],"version-history":[{"count":13,"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/posts\/1124\/revisions"}],"predecessor-version":[{"id":1899,"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/posts\/1124\/revisions\/1899"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/media\/1128"}],"wp:attachment":[{"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/media?parent=1124"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/categories?post=1124"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.xlsoft.com\/jp\/blog\/intel\/wp-json\/wp\/v2\/tags?post=1124"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}