{"id":63834,"date":"2025-11-26T06:41:47","date_gmt":"2025-11-25T21:41:47","guid":{"rendered":"https:\/\/wordpress.realcollector.jp\/blog\/?p=63834"},"modified":"2025-11-26T06:41:47","modified_gmt":"2025-11-25T21:41:47","slug":"tpu","status":"publish","type":"post","link":"https:\/\/wordpress.realcollector.jp\/blog\/?p=63834","title":{"rendered":"tpu"},"content":{"rendered":"<p>Tensor Processing Unit\uff08TPU\uff09\u3068\u306f\u3001Google\u304c\u958b\u767a\u3057\u305f\u6a5f\u68b0\u5b66\u7fd2\u5411\u3051\u306e\u5c02\u7528\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30bf\uff08ASIC: Application-Specific Integrated Circuit\uff09\u3067\u3059\u3002TPU\u306f\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u306e\u5b66\u7fd2\u304a\u3088\u3073\u63a8\u8ad6\uff08\u30a4\u30f3\u30d5\u30a1\u30ec\u30f3\u30b9\uff09\u51e6\u7406\u3092\u9ad8\u901f\u5316\u3059\u308b\u3053\u3068\u3092\u76ee\u7684\u3068\u3057\u3066\u8a2d\u8a08\u3055\u308c\u3066\u304a\u308a\u3001\u7279\u306b\u5927\u898f\u6a21\u306a\u30c6\u30f3\u30bd\u30eb\u6f14\u7b97\u3092\u52b9\u7387\u826f\u304f\u5b9f\u884c\u3067\u304d\u308b\u3088\u3046\u6700\u9069\u5316\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u4ee5\u4e0b\u3067\u306f\u3001TPU\u306e\u6982\u8981\u30fb\u7279\u5fb4\u30fb\u5fdc\u7528\u4f8b\u306a\u3069\u3092\u65e5\u672c\u8a9e\u3067\u8a73\u3057\u304f\u89e3\u8aac\u3057\u307e\u3059\u3002<\/p>\n<p>TPU\u306e\u8a95\u751f\u306f2016\u5e74\u306b\u9061\u308a\u307e\u3059\u3002Google\u306f\u81ea\u793e\u306e\u30c7\u30fc\u30bf\u30bb\u30f3\u30bf\u30fc\u3067\u5229\u7528\u3059\u308b\u305f\u3081\u306b\u3001TensorFlow\u306e\u30e2\u30c7\u30eb\u8a08\u7b97\u3092\u3088\u308a\u9ad8\u901f\u304b\u3064\u7701\u96fb\u529b\u3067\u5b9f\u884c\u3059\u308b\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30bf\u306e\u958b\u767a\u3092\u6c7a\u5b9a\u3057\u307e\u3057\u305f\u3002\u305d\u306e\u7d50\u679c\u3068\u3057\u3066\u767a\u8868\u3055\u308c\u305f\u306e\u304c\u300cTPU v1\u300d\u3067\u3059\u3002\u4ee5\u964d\u3001TPU v2\u3001v3\u3001v4 \u3068\u3044\u3046\u4e16\u4ee3\u3092\u91cd\u306d\u3001\u9ad8\u901f\u5316\u6027\u80fd\u3084\u30e1\u30e2\u30ea\u5e2f\u57df\u306a\u3069\u304c\u5927\u5e45\u306b\u6539\u5584\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u7279\u306bTPU v3\u4ee5\u964d\u306f\u300cPod\u300d\u3068\u547c\u3070\u308c\u308b\u8907\u6570\u306eTPU\u30c1\u30c3\u30d7\u3092\u7d50\u5408\u3057\u3066\u6570\u767e\u5146\u5358\u4f4d\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u3064\u5927\u898f\u6a21\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u3092\u53ef\u80fd\u306b\u3057\u3066\u304a\u308a\u3001\u81ea\u7136\u8a00\u8a9e\u51e6\u7406\u3084\u753b\u50cf\u8a8d\u8b58\u3001\u5f37\u5316\u5b66\u7fd2\u306a\u3069\u306e\u6700\u5148\u7aef\u7814\u7a76\u306b\u3082\u5e83\u304f\u5229\u7528\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>TPU\u306fFPGA\u3084GPU\u3068\u6bd4\u8f03\u3057\u3066\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u5f37\u307f\u3092\u6301\u3063\u3066\u3044\u307e\u3059\u3002\u307e\u305a\u7b2c\u4e00\u306b\u3001\u884c\u5217\u4e57\u7b97\uff08\u30de\u30c8\u30ea\u30c3\u30af\u30b9\u4e57\u7b97\uff09\u3084\u30d9\u30af\u30c8\u30eb\u6f14\u7b97\u3092\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u30ec\u30d9\u30eb\u3067\u6700\u9069\u5316\u3057\u3066\u304a\u308a\u3001\u6f14\u7b97\u30e6\u30cb\u30c3\u30c8\u304c\u5927\u898f\u6a21\u306b\u4e26\u5217\u5316\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u306b\u3088\u308a\u3001\u540c\u3058\u6d88\u8cbb\u96fb\u529b\u3042\u305f\u308a\u306e\u6f14\u7b97\u6027\u80fd\uff08TOPS\/W\uff09\u304c\u975e\u5e38\u306b\u9ad8\u3044\u70b9\u304c\u7279\u5fb4\u3067\u3059\u3002\u6b21\u306b\u3001\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u306b\u6df1\u304f\u7d44\u307f\u8fbc\u307e\u308c\u305f\u30c7\u30fc\u30bf\u30d5\u30ed\u30fc\u30a8\u30f3\u30b8\u30f3\u304c\u3001\u30c6\u30f3\u30bd\u30eb\u9593\u306e\u30c7\u30fc\u30bf\u79fb\u52d5\u3092\u6700\u5c0f\u9650\u306b\u6291\u3048\u3001\u30ec\u30a4\u30c6\u30f3\u30b7\u3092\u4f4e\u6e1b\u3057\u307e\u3059\u3002\u307e\u305f\u3001\u7d71\u5408\u3055\u308c\u305f\u9ad8\u901f\u30aa\u30f3\u30c1\u30c3\u30d7\u30e1\u30e2\u30ea\uff08\u9ad8\u5e2f\u57df\u5e45\u30e1\u30e2\u30ea\uff1aHBM\uff09\u3092\u642d\u8f09\u3059\u308b\u3053\u3068\u3067\u3001\u5927\u898f\u6a21\u30e2\u30c7\u30eb\u3067\u3082\u30dc\u30c8\u30eb\u30cd\u30c3\u30af\u306b\u306a\u308a\u306b\u304f\u3044\u8a2d\u8a08\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u3055\u3089\u306b\u3001TPU\u306fGoogle Cloud Platform\uff08GCP\uff09\u4e0a\u3067\u30af\u30e9\u30a6\u30c9\u30b5\u30fc\u30d3\u30b9\u3068\u3057\u3066\u3082\u63d0\u4f9b\u3055\u308c\u3066\u304a\u308a\u3001\u7814\u7a76\u8005\u3084\u958b\u767a\u8005\u306f\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u3092\u501f\u308a\u3066\u624b\u8efd\u306b\u9ad8\u6027\u80fd\u306a\u6a5f\u68b0\u5b66\u7fd2\u57fa\u76e4\u3092\u5229\u7528\u3067\u304d\u307e\u3059\u3002TensorFlow\u3060\u3051\u3067\u306a\u304fPyTorch\u3084JAX\u306a\u3069\u306e\u4e3b\u8981\u306a\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u306b\u3082\u5bfe\u5fdc\u3057\u3001\u69d8\u3005\u306a\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u306b\u5fdc\u7528\u53ef\u80fd\u3067\u3059\u3002\u307e\u305f\u3001TPU\u306e\u304a\u304b\u3052\u3067\u5b66\u7fd2\u6642\u9593\u304c\u5f93\u6765\u6bd4\u3067\u6570\u500d\u304b\u3089\u6570\u5341\u500d\u306b\u77ed\u7e2e\u3055\u308c\u3001\u30e2\u30c7\u30eb\u958b\u767a\u306e\u30a4\u30c6\u30ec\u30fc\u30b7\u30e7\u30f3\u3092\u8fc5\u901f\u306b\u56de\u305b\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\u3002<\/p>\n<p>\u4e00\u65b9\u3067TPU\u306b\u306f\u3001FPGA\u3084GPU\u3068\u6bd4\u8f03\u3057\u3066\u30d5\u30ec\u30ad\u30b7\u30d3\u30ea\u30c6\u30a3\u304c\u82e5\u5e72\u52a3\u308b\u9762\u3082\u3042\u308a\u307e\u3059\u3002\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3084\u30ab\u30b9\u30bf\u30e0\u30aa\u30da\u30ec\u30fc\u30b7\u30e7\u30f3\u306e\u5bfe\u5fdc\u306b\u306f\u5c02\u7528\u30e9\u30a4\u30d6\u30e9\u30ea\u3084\u30b3\u30f3\u30d1\u30a4\u30e9\u306e\u6700\u9069\u5316\u304c\u5fc5\u8981\u3067\u3001\u5c0e\u5165\u521d\u671f\u306b\u306f\u5b66\u7fd2\u30b3\u30b9\u30c8\u3082\u767a\u751f\u3057\u307e\u3059\u3002\u3057\u304b\u3057\u5341\u5206\u306b\u6700\u9069\u5316\u304c\u9032\u3080\u3068\u3001\u5b66\u7fd2\u30fb\u63a8\u8ad6\u3068\u3082\u306b\u512a\u308c\u305f\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\u3092\u767a\u63ee\u3057\u3001\u7279\u306b\u5927\u898f\u6a21\u306a\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u30e2\u30c7\u30eb\u3092\u6271\u3046\u30b7\u30ca\u30ea\u30aa\u3067\u306f\u975e\u5e38\u306b\u9b45\u529b\u7684\u306a\u9078\u629e\u80a2\u3068\u306a\u308a\u307e\u3059\u3002<\/p>\n<p>\u307e\u3068\u3081\u308b\u3068\u3001TPU\u306fGoogle\u304c\u63d0\u4f9b\u3059\u308b\u6a5f\u68b0\u5b66\u7fd2\u5411\u3051\u5c02\u7528\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u3067\u3042\u308a\u3001\u9ad8\u3044\u4e26\u5217\u6f14\u7b97\u6027\u80fd\u30fb\u7701\u96fb\u529b\u6027\u30fb\u5927\u898f\u6a21\u30e2\u30c7\u30eb\u5bfe\u5fdc\u3092\u5f37\u307f\u3068\u3057\u3066\u3044\u307e\u3059\u3002\u7814\u7a76\u958b\u767a\u3084\u5546\u7528\u30b5\u30fc\u30d3\u30b9\u306b\u304a\u3044\u3066\u3001\u591a\u6570\u306e\u30c7\u30fc\u30bf\u30fb\u5927\u898f\u6a21\u30e2\u30c7\u30eb\u3092\u9ad8\u901f\u306b\u5b66\u7fd2\u30fb\u63a8\u8ad6\u3057\u305f\u3044\u5834\u5408\u306b\u9069\u3057\u305f\u30d7\u30e9\u30c3\u30c8\u30d5\u30a9\u30fc\u30e0\u3067\u3059\u3002<\/p>\n<p>\uff1cTPU\u306e\u4e3b\u306a\u7279\u5fb4\uff085\u9805\u76ee\u4ee5\u4e0a\uff09\uff1e 1. \u5927\u898f\u6a21\u4e26\u5217\u6f14\u7b97\u30e6\u30cb\u30c3\u30c8 2. \u9ad8\u5e2f\u57df\u5e45\u30aa\u30f3\u30c1\u30c3\u30d7\u30e1\u30e2\u30ea\uff08HBM\uff09 3. \u30ec\u30a4\u30c6\u30f3\u30b7\u3092\u6291\u3048\u308b\u30c7\u30fc\u30bf\u30d5\u30ed\u30fc\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3 4. TensorFlow \/ PyTorch \/ JAX\u306a\u3069\u4e3b\u8981\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u5bfe\u5fdc 5. Google Cloud TPU\u3068\u3057\u3066\u306e\u30af\u30e9\u30a6\u30c9\u5229\u7528\u53ef\u80fd 6. Pod\u69cb\u6210\u306b\u3088\u308b\u5927\u898f\u6a21\u30e2\u30c7\u30eb\u5b66\u7fd2\u30b5\u30dd\u30fc\u30c8 7. \u6d88\u8cbb\u96fb\u529b\u3042\u305f\u308a\u306e\u6f14\u7b97\u6027\u80fd\uff08TOPS\/W\uff09\u304c\u9ad8\u3044<\/p>\n<p>\uff1c\u53c2\u8003\u6587\u732e\u30fbURL\uff08\u65e5\u672c\u8a9e\uff0f\u82f1\u8a9e\uff09\uff1e 1. Google Cloud\u516c\u5f0f\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u300cCloud TPU \u306e\u6982\u8981\u300d <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/overview?hl=ja\">https:\/\/cloud.google.com\/tpu\/docs\/overview?hl=ja<\/a> 2. Google AI Blog\u300cCloud TPU v4 \u3092\u767a\u8868\u300d <a href=\"https:\/\/ai.googleblog.com\/2021\/07\/announcing-cloud-tpu-v4.html\">https:\/\/ai.googleblog.com\/2021\/07\/announcing-cloud-tpu-v4.html<\/a> 3. TensorFlow\u516c\u5f0f\u30b5\u30a4\u30c8\u300cTPU \u30ac\u30a4\u30c9\u300d <a href=\"https:\/\/www.tensorflow.org\/guide\/tpu?hl=ja\">https:\/\/www.tensorflow.org\/guide\/tpu?hl=ja<\/a> 4. Google Cloud Next \u201819 \u30bb\u30c3\u30b7\u30e7\u30f3\u300cTPU \u306e\u65b0\u4e16\u4ee3\u3068\u6d3b\u7528\u4e8b\u4f8b\u300d <a href=\"https:\/\/cloud.withgoogle.com\/next\/sf\/sessions\/cloud-tpu-new-generation\">https:\/\/cloud.withgoogle.com\/next\/sf\/sessions\/cloud-tpu-new-generation<\/a> 5. IEEE\u8ad6\u6587\u300cGoogle\u2019s Tensor Processing Unit: Architecture and Performance\u300d <a href=\"https:\/\/ieeexplore.ieee.org\/document\/8681280\">https:\/\/ieeexplore.ieee.org\/document\/8681280<\/a> 6. Qiita\u8a18\u4e8b\u300c\u306f\u3058\u3081\u3066\u306eCloud TPU\u5165\u9580\u300d <a href=\"https:\/\/qiita.com\/tags\/CloudTPU\">https:\/\/qiita.com\/tags\/CloudTPU<\/a><\/p>\n<p><a href=\"https:\/\/wordpress.realcollector.jp\/blog\/wp-content\/uploads\/2025\/11\/current_stablediffusion_api_response-320.png\" rel=\"attachment wp-att-63835\"><img loading=\"lazy\" decoding=\"async\" data-attachment-id=\"63835\" data-permalink=\"https:\/\/wordpress.realcollector.jp\/blog\/?attachment_id=63835\" data-orig-file=\"https:\/\/wordpress.realcollector.jp\/blog\/wp-content\/uploads\/2025\/11\/current_stablediffusion_api_response-320.png\" data-orig-size=\"512,512\" data-comments-opened=\"1\" data-image-meta=\"{&quot;aperture&quot;:&quot;0&quot;,&quot;credit&quot;:&quot;&quot;,&quot;camera&quot;:&quot;&quot;,&quot;caption&quot;:&quot;&quot;,&quot;created_timestamp&quot;:&quot;0&quot;,&quot;copyright&quot;:&quot;&quot;,&quot;focal_length&quot;:&quot;0&quot;,&quot;iso&quot;:&quot;0&quot;,&quot;shutter_speed&quot;:&quot;0&quot;,&quot;title&quot;:&quot;&quot;,&quot;orientation&quot;:&quot;0&quot;}\" data-image-title=\"current_stablediffusion_api_response\" data-image-description=\"\" data-image-caption=\"\" data-medium-file=\"https:\/\/wordpress.realcollector.jp\/blog\/wp-content\/uploads\/2025\/11\/current_stablediffusion_api_response-320-300x300.png\" data-large-file=\"https:\/\/wordpress.realcollector.jp\/blog\/wp-content\/uploads\/2025\/11\/current_stablediffusion_api_response-320.png\" src=\"https:\/\/wordpress.realcollector.jp\/blog\/wp-content\/uploads\/2025\/11\/current_stablediffusion_api_response-320.png\" alt=\"\" title=\"current_stablediffusion_api_response-320-png\" width=\"1024\" height=\"1024\" class=\"alignnone size-full wp-image-63835\" srcset=\"https:\/\/wordpress.realcollector.jp\/blog\/wp-content\/uploads\/2025\/11\/current_stablediffusion_api_response-320.png 512w, https:\/\/wordpress.realcollector.jp\/blog\/wp-content\/uploads\/2025\/11\/current_stablediffusion_api_response-320-300x300.png 300w, https:\/\/wordpress.realcollector.jp\/blog\/wp-content\/uploads\/2025\/11\/current_stablediffusion_api_response-320-150x150.png 150w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Tensor Processing Unit\uff08TPU\uff09\u3068\u306f\u3001Google\u304c\u958b\u767a\u3057\u305f\u6a5f\u68b0\u5b66\u7fd2\u5411\u3051\u306e\u5c02\u7528\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30bf\uff08ASIC: Application-Specific Integrated Circuit\uff09\u3067\u3059\u3002TPU [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[10942],"tags":[13796,13794,3847,13795,2296],"class_list":["post-63834","post","type-post","status-publish","format-standard","hentry","category-o4-mini","tag-ai","tag-tpu","tag-3847","tag-13795","tag-2296"],"jetpack_featured_media_url":"","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=\/wp\/v2\/posts\/63834","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=63834"}],"version-history":[{"count":1,"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=\/wp\/v2\/posts\/63834\/revisions"}],"predecessor-version":[{"id":63836,"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=\/wp\/v2\/posts\/63834\/revisions\/63836"}],"wp:attachment":[{"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=63834"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=63834"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/wordpress.realcollector.jp\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=63834"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}