{"id":50247,"date":"2026-10-28T09:06:58","date_gmt":"2026-10-28T00:06:58","guid":{"rendered":"https:\/\/music.bear-blog.net\/?p=50247"},"modified":"2026-10-28T09:06:58","modified_gmt":"2026-10-28T00:06:58","slug":"dgx-spark-ling3-moe-inference-optimizations-20261028","status":"publish","type":"post","link":"https:\/\/music.bear-blog.net\/?p=50247","title":{"rendered":"DGX Spark\u306eAI\u63a8\u8ad6\u304c42.5%\u9ad8\u901f\u5316\uff01NVIDIA\u304c\u516c\u958b\u3057\u305f3\u3064\u306e\u6700\u9069\u5316\u3068\u300c185%\u300d\u306e\u843d\u3068\u3057\u7a74"},"content":{"rendered":"<p><strong>\u300cDGX Spark\u306e\u63a8\u8ad6\u6027\u80fd\u3092\u4e0a\u3052\u308b\u306b\u306f\u3001\u3082\u3063\u3068\u5927\u304d\u306aGPU\u304c\u5fc5\u8981\u306a\u306e\u304b\uff1f\u300d<\/strong>\u2014\u2014\u305d\u3046\u601d\u3063\u3066\u3044\u305f\u4eba\u306b\u3068\u3063\u3066\u3001\u4eca\u56de\u306eNVIDIA\u306e\u767a\u8868\u306f\u6c17\u306b\u306a\u308b\u30cb\u30e5\u30fc\u30b9\u3060\u3002<\/p>\n<p>2026\u5e7410\u670827\u65e5\u3001NVIDIA\u306f\u516c\u5f0f\u6280\u8853\u30d6\u30ed\u30b0\u3067\u3001<strong>DGX Spark\u4e0a\u306e\u30ed\u30fc\u30ab\u30ebLLM\u3092\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u6700\u9069\u5316\u3067\u9ad8\u901f\u5316\u3059\u308b3\u3064\u306e\u624b\u6cd5<\/strong>\u3092\u516c\u958b\u3057\u305f\u3002\u5bfe\u8c61\u30e2\u30c7\u30eb\u3092\u4f7f\u3063\u305f\u691c\u8a3c\u3067\u306f\u3001\u5b9f\u969b\u306e\u554f\u984c\u96c6\u3092\u51e6\u7406\u3057\u305f\u3068\u304d\u306e\u751f\u6210\u901f\u5ea6\u304c<strong>42.5%\u5411\u4e0a<\/strong>\u3002\u4e00\u65b9\u3001\u6761\u4ef6\u3092\u5358\u7d14\u5316\u3057\u305f\u30c6\u30b9\u30c8\u3067\u306f\u300c185.1%\u5411\u4e0a\u300d\u3068\u3044\u3046\u76ee\u3092\u5f15\u304f\u6570\u5b57\u3082\u51fa\u3066\u3044\u308b\u3002<\/p>\n<p>\u305f\u3060\u3057\u300cDGX Spark\u306a\u3089\u4f55\u3067\u3082185%\u901f\u304f\u306a\u308b\u300d\u3068\u53d7\u3051\u53d6\u308b\u306e\u306f\u65e9\u8a08\u3002\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001<strong>\u4f55\u3092\u5909\u3048\u305f\u3089\u901f\u304f\u306a\u3063\u305f\u306e\u304b\uff0f\u3069\u306e\u6570\u5b57\u3092\u4fe1\u7528\u3059\u3079\u304d\u304b\uff0f\u81ea\u5206\u306e\u74b0\u5883\u3067\u8a66\u3059\u3068\u304d\u306e\u6ce8\u610f\u70b9<\/strong>\u3092\u3001NVIDIA\u306e\u4e00\u6b21\u60c5\u5831\u3092\u3082\u3068\u306b\u6574\u7406\u3059\u308b\u3002<\/p>\n<h2>\u7d50\u8ad6\uff1a\u5b9f\u30bf\u30b9\u30af\u3067\u306f42.5%\u5411\u4e0a\u3002\u300c185%\u300d\u306f\u5408\u6210\u30c6\u30b9\u30c8\u306e\u6570\u5b57<\/h2>\n<p>\u4eca\u56de\u306e\u691c\u8a3c\u5bfe\u8c61\u306f\u3001Inclusion AI\u306eMoE\u30e2\u30c7\u30eb\u300c<strong>Ling-3.0-flash<\/strong>\u300d\u3002\u7dcf\u30d1\u30e9\u30e1\u30fc\u30bf1240\u5104\uff08124B\uff09\u306e\u3046\u3061\u30011\u30c8\u30fc\u30af\u30f3\u3042\u305f\u308a\u7d0451\u5104\uff085.1B\uff09\u304c\u6709\u52b9\u306b\u306a\u308b\u8a2d\u8a08\u3067\u3001DGX Spark\u4e0a\u3067<strong>\u540c\u6642\u51e6\u7406\u65701\uff08\u30d0\u30c3\u30c1\u30b5\u30a4\u30ba1\uff09<\/strong>\u3068\u3057\u3066\u8a55\u4fa1\u3055\u308c\u305f\u3002<\/p>\n<p>\u307e\u305a\u306f\u6700\u3082\u91cd\u8981\u306a\u6570\u5b57\u3092\u6bd4\u8f03\u3057\u3088\u3046\u3002NVIDIA\u306f\u7b97\u6570\u554f\u984c\u306e\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u300cGSM8K\u300d\u51681,319\u554f\u3067\u3001\u6700\u9069\u5316\u524d\u30683\u624b\u6cd5\u3092\u7d44\u307f\u5408\u308f\u305b\u305f\u69cb\u6210\u3092\u6bd4\u8f03\u3057\u3066\u3044\u308b\u3002<\/p>\n<figure class=\"wp-block-table\">\n<table>\n<thead>\n<tr>\n<th>\u6307\u6a19<\/th>\n<th>\u6700\u9069\u5316\u524d<\/th>\n<th>\u6700\u9069\u5316\u5f8c<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u751f\u6210\u901f\u5ea6\uff08Decode\uff09<\/td>\n<td>37.81 tokens\/\u79d2<\/td>\n<td><strong>53.89 tokens\/\u79d2<\/strong>\uff08+42.5%\uff09<\/td>\n<\/tr>\n<tr>\n<td>\u5e73\u5747\u5fdc\u7b54\u6642\u9593<\/td>\n<td>15.21\u79d2<\/td>\n<td><strong>11.49\u79d2<\/strong>\uff08-24.5%\uff09<\/td>\n<\/tr>\n<tr>\n<td>\u30c8\u30fc\u30af\u30f3\u9593\u6642\u9593\uff08TPOT\uff09<\/td>\n<td>26.45ms<\/td>\n<td>18.56ms\uff08-29.8%\uff09<\/td>\n<\/tr>\n<tr>\n<td>GSM8K\u6b63\u7b54\u7387<\/td>\n<td>96.89%<\/td>\n<td>96.82%\uff08\u7d04-0.08\u30dd\u30a4\u30f3\u30c8\uff09<\/td>\n<\/tr>\n<tr>\n<td>\u521d\u56de\u30c8\u30fc\u30af\u30f3\u6642\u9593\uff08TTFT\uff09<\/td>\n<td>0.207\u79d2<\/td>\n<td>0.214\u79d2<\/td>\n<\/tr>\n<\/tbody>\n<\/table><figcaption>\u51fa\u5178\uff1aNVIDIA\u516c\u5f0f\u6280\u8853\u30d6\u30ed\u30b0\uff082026\u5e7410\u670827\u65e5\uff09\u3002\u540c\u793e\u306e\u6e2c\u5b9a\u7d50\u679c\u3067\u3042\u308a\u3001\u5f53\u30d6\u30ed\u30b0\u72ec\u81ea\u306e\u5b9f\u6e2c\u3067\u306f\u306a\u3044\u3002<\/figcaption><\/figure>\n<p>\u3053\u3053\u304c\u5927\u5207\u3060\u3002<strong>\u30c8\u30fc\u30af\u30f3\u751f\u6210\u901f\u5ea6\u306f42.5%\u5897\u52a0\u3057\u305f\u304c\u3001\u56de\u7b54\u5168\u4f53\u306e\u5f85\u3061\u6642\u9593\u306f24.5%\u6e1b<\/strong>\u3002\u5358\u7d14\u306b\u300c\u30c1\u30e3\u30c3\u30c8\u304c42.5%\u65e9\u304f\u7d42\u308f\u308b\u300d\u3068\u8a00\u3044\u63db\u3048\u308b\u306e\u306f\u6b63\u78ba\u3067\u306f\u306a\u3044\u3002\u307e\u305f\u3001TTFT\u306f\u6539\u5584\u305b\u305a\u3001\u307b\u307c\u6a2a\u3070\u3044\u3060\u3063\u305f\u3002<\/p>\n<h2>\u4f55\u3092\u5909\u3048\u305f\u306e\u304b\uff1f \u9ad8\u901f\u5316\u3092\u751f\u3093\u30603\u3064\u306e\u5de5\u592b<\/h2>\n<h3>\u2460 MoE\u306e\u8a08\u7b97\u90e8\u5206\u3092\u300cHumming\u300d\u306b\u7f6e\u304d\u63db\u3048\u308b<\/h3>\n<p>MoE\uff08Mixture of Experts\uff09\u306f\u3001\u591a\u6570\u306e\u300c\u5c02\u9580\u5bb6\u300d\u304b\u3089\u5fc5\u8981\u306a\u3082\u306e\u3092\u9078\u3093\u3067\u8a08\u7b97\u3059\u308b\u4ed5\u7d44\u307f\u3060\u3002\u305d\u3053\u3067\u4f7f\u3046\u8a08\u7b97\u30ab\u30fc\u30cd\u30eb\u3092\u3001\u5f93\u6765\u306eFlashInfer MXFP4\u30d0\u30c3\u30af\u30a8\u30f3\u30c9\u304b\u3089<strong>Humming<\/strong>\u306b\u5909\u66f4\u3057\u305f\u3002<\/p>\n<p>Humming\u306f\u5b9f\u884c\u5bfe\u8c61\u306e\u884c\u5217\u5f62\u72b6\u3084GPU\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u5408\u308f\u305b\u3066\u30ab\u30fc\u30cd\u30eb\u3092\u30b3\u30f3\u30d1\u30a4\u30eb\u3059\u308b\u3002NVIDIA\u304c\u793a\u3057\u305f\u5408\u6210\u30c6\u30b9\u30c8\u3067\u306f\u3001\u3053\u308c\u3060\u3051\u3067<strong>36.09\u219239.45 tokens\/\u79d2\uff08+9.3%\uff09<\/strong>\u3068\u306a\u3063\u305f\u3002\u307e\u305a\u306f\u7121\u99c4\u306a\u8a08\u7b97\u3084\u30e1\u30e2\u30ea\u30a2\u30af\u30bb\u30b9\u3092\u6e1b\u3089\u3059\u3001\u6bd4\u8f03\u7684\u30a4\u30e1\u30fc\u30b8\u3057\u3084\u3059\u3044\u6700\u9069\u5316\u3060\u3002<\/p>\n<h3>\u2461 LM\u30d8\u30c3\u30c9\u3092FP8\u5316\u3057\u3066\u3001\u30e1\u30e2\u30ea\u304b\u3089\u8aad\u3080\u30c7\u30fc\u30bf\u3092\u6e1b\u3089\u3059<\/h3>\n<p>LLM\u306f\u6b21\u306e\u5358\u8a9e\u3092\u6c7a\u3081\u308b\u969b\u3001\u8a9e\u5f59\u5168\u4f53\u306e\u30b9\u30b3\u30a2\u3092\u8a08\u7b97\u3059\u308b\u300cLM\u30d8\u30c3\u30c9\u300d\u3092\u4f7f\u3046\u3002\u3053\u306e\u30e2\u30c7\u30eb\u3067\u306fLM\u30d8\u30c3\u30c9\u306e\u91cd\u307f\u304cBF16\u5f62\u5f0f\u3067\u4fdd\u6301\u3055\u308c\u3066\u3044\u305f\u3002<\/p>\n<p>\u3053\u3053\u3092<strong>FP8\u91cf\u5b50\u5316<\/strong>\u3059\u308c\u3070\u3001\u3053\u306e\u90e8\u5206\u306e\u91cd\u307f\u3092\u8aad\u307f\u8fbc\u3080\u30c7\u30fc\u30bf\u91cf\u3092\u7406\u8ad6\u4e0a\u534a\u5206\u306b\u3067\u304d\u308b\u3002DGX Spark\u306e\u3088\u3046\u306b\u3001\u63a8\u8ad6\u304c\u30e1\u30e2\u30ea\u5e2f\u57df\u306b\u5236\u7d04\u3055\u308c\u308b\u5834\u9762\u3067\u52b9\u3044\u3066\u304f\u308b\u3002<\/p>\n<p>Humming\u306b\u52a0\u3048\u3066LM\u30d8\u30c3\u30c9\u3092FP8\u5316\u3057\u305f\u5408\u6210\u30c6\u30b9\u30c8\u3067\u306f\u3001<strong>43.94 tokens\/\u79d2\uff08\u6700\u9069\u5316\u524d\u304b\u3089+21.7%\uff09<\/strong>\u3002\u306a\u304a\u3001NVIDIA\u304c\u793a\u3059<code>SGLANG_ENABLE_FP8_LM_HEAD=1<\/code>\u306fSGLang\u5074\u306e\u69cb\u6210\u306b\u95a2\u3059\u308b\u8a2d\u5b9a\u3067\u3001\u3059\u3079\u3066\u306e\u63a8\u8ad6\u30a8\u30f3\u30b8\u30f3\u306b\u305d\u306e\u307e\u307e\u901a\u7528\u3059\u308b\u5171\u901a\u30aa\u30d7\u30b7\u30e7\u30f3\u3067\u306f\u306a\u3044\u3002<\/p>\n<h3>\u2462 MTP\u6295\u6a5f\u7684\u30c7\u30b3\u30fc\u30c9\u3067\u3001\u8907\u6570\u30c8\u30fc\u30af\u30f3\u3092\u307e\u3068\u3081\u3066\u78ba\u304b\u3081\u308b<\/h3>\n<p>\u6700\u3082\u5927\u304d\u306a\u4f38\u3073\u3092\u751f\u3093\u3060\u306e\u304c<strong>MTP\uff08Multi-Token Prediction\uff09\u306b\u3088\u308b\u6295\u6a5f\u7684\u30c7\u30b3\u30fc\u30c9<\/strong>\u3060\u3002\u6b21\u306e1\u30c8\u30fc\u30af\u30f3\u3092\u9806\u756a\u306b\u751f\u6210\u3059\u308b\u4ee3\u308f\u308a\u306b\u3001\u5019\u88dc\u3092\u5148\u8aad\u307f\u3057\u3001\u307e\u3068\u3081\u3066\u691c\u8a3c\u3057\u3066\u63a1\u7528\u3059\u308b\u3002<\/p>\n<p>Ling-3.0-flash\u306b\u306fNEXTN\u3068\u3044\u3046\u4e88\u6e2c\u5c64\u304c\u542b\u307e\u308c\u3066\u304a\u308a\u3001\u8ffd\u52a0\u306e\u5225\u30e2\u30c7\u30eb\u3092\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u305b\u305a\u6d3b\u7528\u3067\u304d\u308b\u3002\u5408\u6210\u30c6\u30b9\u30c8\u3067\u306f3\u624b\u6cd5\u3092\u7d44\u307f\u5408\u308f\u305b\u3066<strong>36.09\u2192102.88 tokens\/\u79d2\uff08+185.1%\uff09<\/strong>\u306b\u5230\u9054\u3057\u305f\u3002<\/p>\n<p>\u305f\u3060\u3057\u3001\u3053\u306e\u30c6\u30b9\u30c8\u306f<strong>\u5165\u529b32K\u30c8\u30fc\u30af\u30f3\u3001\u51fa\u529b2K\u30c8\u30fc\u30af\u30f3\u3001\u30d0\u30c3\u30c1\u30b5\u30a4\u30ba1\u3001\u5358\u7d14\u5316\u3057\u305f\u751f\u6210\u6761\u4ef6<\/strong>\u3002\u5b9f\u5229\u7528\u3067\u5e38\u306b\u7d042.85\u500d\u306e\u901f\u5ea6\u304c\u5f97\u3089\u308c\u308b\u3068\u3044\u3046\u610f\u5473\u3067\u306f\u306a\u3044\u3002<\/p>\n<h2>\u306a\u305c\u300c185%\u300d\u304c\u5b9f\u30bf\u30b9\u30af\u3067\u306f\u300c42.5%\u300d\u306b\u306a\u3063\u305f\u306e\u304b<\/h2>\n<p>\u9375\u306b\u306a\u308b\u306e\u306f\u3001\u6295\u6a5f\u7684\u30c7\u30b3\u30fc\u30c9\u3067<strong>\u5148\u8aad\u307f\u3057\u305f\u30c8\u30fc\u30af\u30f3\u304c\u3069\u308c\u3060\u3051\u6b63\u3057\u304f\u63a1\u7528\u3055\u308c\u305f\u304b<\/strong>\u3060\u3002\u5408\u6210\u30c6\u30b9\u30c8\u3067\u306f1\u56de\u3042\u305f\u308a\u5e73\u57473.90\u30c8\u30fc\u30af\u30f3\u3092\u63a1\u7528\u3067\u304d\u305f\u304c\u3001GSM8K\u3067\u306f2.68\u30c8\u30fc\u30af\u30f3\u3060\u3063\u305f\u3002<\/p>\n<p>\u5b9f\u969b\u306e\u6587\u7ae0\u3084\u63a8\u8ad6\u306f\u5408\u6210\u30c6\u30b9\u30c8\u3088\u308a\u4e88\u6e2c\u3057\u306b\u304f\u3044\u3002\u3055\u3089\u306b\u6e29\u5ea6\u306a\u3069\u306e\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u8a2d\u5b9a\u3001\u56de\u7b54\u306e\u9577\u3055\u3082\u9055\u3046\u3002\u5148\u8aad\u307f\u304c\u5f53\u305f\u308a\u306b\u304f\u3051\u308c\u3070\u3001\u691c\u8a3c\u306e\u624b\u9593\u306b\u5bfe\u3059\u308b\u898b\u8fd4\u308a\u306f\u5c0f\u3055\u304f\u306a\u308b\u3002<\/p>\n<p>\u3064\u307e\u308a\u3001<strong>\u6027\u80fd\u8a55\u4fa1\u306f\u300c\u4e00\u756a\u6d3e\u624b\u306a\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u6570\u5b57\u300d\u3067\u306f\u306a\u304f\u300c\u81ea\u5206\u306e\u7528\u9014\u306b\u8fd1\u3044\u8ca0\u8377\u300d\u3067\u5224\u65ad\u3059\u308b\u5fc5\u8981\u304c\u3042\u308b<\/strong>\u3002\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0AI\u3001\u9577\u6587\u5206\u6790\u3001\u8907\u6570\u4eba\u306e\u540c\u6642\u30c1\u30e3\u30c3\u30c8\u3067\u306f\u3001\u7d50\u679c\u304c\u5909\u308f\u308b\u53ef\u80fd\u6027\u304c\u9ad8\u3044\u3002<\/p>\n<h2>\u8981\u6ce8\u610f\uff1a3\u3064\u3092\u540c\u6642\u306b\u6709\u52b9\u5316\u3059\u308b\u3068\u3001\u9006\u306b\u9045\u304f\u306a\u308b\u3053\u3068\u3082<\/h2>\n<p>\u4eca\u56de\u306e\u8a18\u4e8b\u3067\u3001\u3080\u3057\u308d\u6280\u8853\u8005\u304c\u898b\u9003\u305b\u306a\u3044\u306e\u306f\u3053\u3053\u3060\u3002NVIDIA\u306f<strong>FP8\u5316\u3057\u305fLM\u30d8\u30c3\u30c9\u3068MTP\u3092\u7d44\u307f\u5408\u308f\u305b\u308b\u969b\u306e\u5b9f\u88c5\u4e0a\u306e\u843d\u3068\u3057\u7a74<\/strong>\u3082\u8aac\u660e\u3057\u3066\u3044\u308b\u3002<\/p>\n<p>\u30c9\u30e9\u30d5\u30c8\u5074\u306e\u30e2\u30c7\u30eb\u3068\u672c\u4f53\u5074\u306e\u30e2\u30c7\u30eb\u3067\u3001\u5358\u306bLM\u30d8\u30c3\u30c9\u306e\u91cd\u307f\u3060\u3051\u3092\u5171\u6709\u3059\u308b\u3068\u6574\u5408\u304c\u53d6\u308c\u306a\u3044\u5834\u5408\u304c\u3042\u308b\u3002\u5143\u306eLing NEXTN\u5b9f\u88c5\u3067\u306f\u3001\u5fc5\u8981\u306a\u30e2\u30b8\u30e5\u30fc\u30eb\u5168\u4f53\u306e\u5171\u6709\u30a4\u30f3\u30bf\u30fc\u30d5\u30a7\u30fc\u30b9\u304c\u306a\u304f\u3001\u30c9\u30e9\u30d5\u30c8\u5074\u304cFP8\u91cd\u307f\u3092BF16\u3068\u3057\u3066\u6271\u3063\u3066\u3057\u307e\u3046\u30b1\u30fc\u30b9\u304c\u3042\u3063\u305f\u3002<\/p>\n<p>\u305d\u306e\u5834\u5408\u3001\u5148\u8aad\u307f\u30c8\u30fc\u30af\u30f3\u306e\u63a1\u7528\u6570\u306f\u5e73\u57471.00\u307e\u3067\u843d\u3061\u3001<strong>\u6700\u9069\u5316\u306a\u3057\u306e\u69cb\u6210\u3088\u308a\u3082\u4f4e\u901f<\/strong>\u306b\u306a\u308b\u3068\u3044\u3046\u3002\u89e3\u6c7a\u306b\u306f\u4e21\u8005\u304cLM\u30d8\u30c3\u30c9\u306e\u30e2\u30b8\u30e5\u30fc\u30eb\u5168\u4f53\u3092\u5171\u6709\u3059\u308b\u9069\u5207\u306a\u5b9f\u88c5\u304c\u5fc5\u8981\u3067\u3001NVIDIA\u306f\u516c\u5f0fCookbook\u306e\u624b\u9806\u3092\u6848\u5185\u3057\u3066\u3044\u308b\u3002<\/p>\n<p><strong>\u8a2d\u5b9a\u9805\u76ee\u3092ON\u306b\u3057\u305f\u3060\u3051\u3067\u300c\u9ad8\u901f\u5316\u3067\u304d\u305f\u300d\u3068\u5224\u65ad\u3057\u306a\u3044<\/strong>\u3053\u3068\u3002\u8d77\u52d5\u30ed\u30b0\u3084\u30d0\u30c3\u30af\u30a8\u30f3\u30c9\u306e\u9078\u629e\u3001MTP\u306e\u63a1\u7528\u7387\u3001\u5b9f\u6e2c\u306eTPOT\u3068TPS\u307e\u3067\u78ba\u8a8d\u3057\u3066\u521d\u3081\u3066\u52b9\u679c\u3092\u8a55\u4fa1\u3067\u304d\u308b\u3002<\/p>\n<h2>DGX Spark\u30e6\u30fc\u30b6\u30fc\u306f\u4f55\u3092\u8a66\u3059\u3079\u304d\uff1f<\/h2>\n<p>\u624b\u5143\u306eDGX Spark\u3092\u9ad8\u901f\u5316\u3057\u305f\u3044\u306a\u3089\u3001\u6700\u521d\u306b\u300c\u81ea\u5206\u304c\u4eca\u3001\u4f55\u306b\u5f85\u305f\u3055\u308c\u3066\u3044\u308b\u304b\u300d\u3092\u5207\u308a\u5206\u3051\u305f\u3044\u3002<\/p>\n<ul>\n<li><strong>1\u4eba\u3067\u9577\u6587\u3092\u751f\u6210\u3059\u308b\u7528\u9014\uff1a<\/strong>\u5358\u4e00\u30ea\u30af\u30a8\u30b9\u30c8\u306edecode TPS\u3001TPOT\u3001MTP\u63a1\u7528\u7387\u3092\u91cd\u8996\u3002\u4eca\u56de\u306e\u7814\u7a76\u306b\u8fd1\u3044\u6761\u4ef6\u3002<\/li>\n<li><strong>\u8907\u6570\u4eba\u304c\u540c\u6642\u306b\u4f7f\u3046\u7528\u9014\uff1a<\/strong>\u7dcfTPS\u3060\u3051\u3067\u306a\u304f\u3001\u5229\u7528\u8005\u5225TPS\u3001TTFT\u3001P95\u5f85\u3061\u6642\u9593\u3001GPU\u30e1\u30e2\u30ea\u4f7f\u7528\u91cf\u3092\u8a18\u9332\u3002<\/li>\n<li><strong>\u65e2\u5b58\u306eQwen\u306a\u3069\u3092\u4f7f\u3046\u7528\u9014\uff1a<\/strong>\u4eca\u56de\u306e\u7d50\u679c\u306fLing-3.0-flash\u3068\u6307\u5b9a\u306eSGLang\u69cb\u6210\u306b\u5bfe\u3059\u308b\u3082\u306e\u3002\u5225\u30e2\u30c7\u30eb\u3084llama.cpp\u3001vLLM\u306a\u3069\u3067\u540c\u3058\u6539\u5584\u7387\u3092\u4fdd\u8a3c\u3059\u308b\u3082\u306e\u3067\u306f\u306a\u3044\u3002<\/li>\n<li><strong>\u672c\u756a\u76f8\u5f53\u74b0\u5883\uff1a<\/strong>\u5909\u66f4\u524d\u306e\u8a2d\u5b9a\u3068\u6e2c\u5b9a\u7d50\u679c\u3092\u4fdd\u5b58\u3057\u30011\u9805\u76ee\u305a\u3064\u5909\u66f4\u3057\u3066\u52b9\u679c\u3092\u6bd4\u8f03\u3002\u7cbe\u5ea6\u30fb\u30a8\u30e9\u30fc\u7387\u30fb\u30e1\u30e2\u30ea\u5727\u8feb\u3082\u5408\u308f\u305b\u3066\u30c1\u30a7\u30c3\u30af\u3002<\/li>\n<\/ul>\n<p>\u5f53\u30d6\u30ed\u30b0\u3067\u3082\u4ee5\u524d\u3001<a href=\"https:\/\/music.bear-blog.net\/?p=49889\">DGX Spark\u4e0a\u3067Strata\u3068llama.cpp\u3092\u6bd4\u8f03\u3057\u305f\u5b9f\u6e2c\u8a18\u4e8b<\/a>\u3092\u516c\u958b\u3057\u3066\u3044\u308b\u3002\u3042\u3061\u3089\u306f<strong>\u63a8\u8ad6\u30a8\u30f3\u30b8\u30f3\u3068\u540c\u6642\u5b9f\u884c\u6761\u4ef6\u306e\u6bd4\u8f03<\/strong>\u3001\u4eca\u56de\u306f<strong>\u30e2\u30c7\u30eb\u30fb\u30ab\u30fc\u30cd\u30eb\u30fb\u30c7\u30b3\u30fc\u30c9\u65b9\u5f0f\u306e\u6700\u9069\u5316<\/strong>\u3068\u3044\u3046\u5225\u306e\u8996\u70b9\u3060\u3002\u4e21\u65b9\u3092\u4e00\u7dd2\u306b\u773a\u3081\u308b\u3068\u3001\u300c\u901f\u3044\u69cb\u6210\u300d\u304c\u5fc5\u305a\u3057\u30821\u3064\u3067\u306f\u306a\u3044\u3053\u3068\u304c\u898b\u3048\u3066\u304f\u308b\u3002<\/p>\n<h2>FAQ\uff1a\u4eca\u56de\u306e\u767a\u8868\u3067\u6c17\u306b\u306a\u308b3\u3064\u306e\u7591\u554f<\/h2>\n<h3>Q. DGX Spark\u3092\u30a2\u30c3\u30d7\u30c7\u30fc\u30c8\u3059\u308b\u3060\u3051\u306742.5%\u901f\u304f\u306a\u308b\uff1f<\/h3>\n<p><strong>\u3044\u3044\u3048\u3002<\/strong>\u3053\u308c\u306f\u7279\u5b9a\u30e2\u30c7\u30eb\u3068\u63a8\u8ad6\u30b9\u30bf\u30c3\u30af\u30673\u3064\u306e\u6700\u9069\u5316\u3092\u9069\u5207\u306b\u7d44\u307f\u5408\u308f\u305b\u305f\u691c\u8a3c\u7d50\u679c\u3002OS\u30a2\u30c3\u30d7\u30c7\u30fc\u30c8\u3060\u3051\u3067\u5168LLM\u306b\u9069\u7528\u3055\u308c\u308b\u6570\u5b57\u3067\u306f\u306a\u3044\u3002<\/p>\n<h3>Q. \u6700\u5927185%\u9ad8\u901f\u5316\u3068\u3044\u3046\u30cb\u30e5\u30fc\u30b9\u306f\u8a87\u5f35\uff1f<\/h3>\n<p>\u6570\u5b57\u305d\u306e\u3082\u306e\u306fNVIDIA\u306e<strong>\u5408\u6210\u30c6\u30b9\u30c8\u3067\u306e\u6e2c\u5b9a\u5024<\/strong>\u3060\u3002\u305f\u3060\u3057GSM8K\u5168\u554f\u3067\u306e\u6539\u5584\u7387\u3068\u306f\u6761\u4ef6\u304c\u7570\u306a\u308b\u305f\u3081\u3001\u5b9f\u5229\u7528\u306e\u4ee3\u8868\u5024\u3068\u3057\u3066185%\u3060\u3051\u3092\u5f37\u8abf\u3059\u308b\u306e\u306f\u4e0d\u9069\u5207\u3060\u3002<\/p>\n<h3>Q. \u7cbe\u5ea6\u3092\u843d\u3068\u3057\u3066\u901f\u304f\u3057\u305f\u306e\uff1f<\/h3>\n<p>GSM8K\u3067\u306f96.89%\u304b\u308996.82%\u3068\u5c0f\u3055\u306a\u5dee\u306b\u3068\u3069\u307e\u308a\u3001NVIDIA\u306f\u7d71\u8a08\u7684\u306b\u691c\u51fa\u3067\u304d\u308b\u7cbe\u5ea6\u5909\u5316\u306f\u306a\u304b\u3063\u305f\u3068\u3057\u3066\u3044\u308b\u3002\u305f\u3060\u3057\u3001<strong>\u65e5\u672c\u8a9e\u3001\u30d7\u30ed\u30b0\u30e9\u30df\u30f3\u30b0\u3001\u9577\u6587\u8981\u7d04\u306a\u3069\u3067\u306e\u7cbe\u5ea6\u7dad\u6301\u306f\u5225\u9014\u691c\u8a3c\u304c\u5fc5\u8981<\/strong>\u3060\u3002<\/p>\n<h2>\u307e\u3068\u3081\uff1aDGX Spark\u306f\u300c\u8cb7\u3044\u66ff\u3048\u300d\u4ee5\u5916\u3067\u3082\u307e\u3060\u901f\u304f\u3067\u304d\u308b<\/h2>\n<p>\u4eca\u56de\u306e\u6280\u8853\u30d6\u30ed\u30b0\u306e\u9762\u767d\u3055\u306f\u3001\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u3092\u5897\u3084\u3059\u306e\u3067\u306f\u306a\u304f\u3001<strong>\u30e1\u30e2\u30ea\u8ee2\u9001\u306e\u524a\u6e1b\u3068\u751f\u6210\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306e\u6539\u5584<\/strong>\u3067\u6570\u5b57\u3092\u4f38\u3070\u3057\u305f\u3053\u3068\u306b\u3042\u308b\u3002<\/p>\n<p>\u73fe\u5b9f\u7684\u306b\u53d7\u3051\u6b62\u3081\u305f\u3044\u6210\u679c\u306f\u3001<strong>Ling-3.0-flash\u306eGSM8K\u691c\u8a3c\u3067\u751f\u6210\u901f\u5ea6\u304c42.5%\u5411\u4e0a\u3057\u3001\u5e73\u5747\u5fdc\u7b54\u6642\u9593\u304c24.5%\u77ed\u7e2e\u3057\u305f<\/strong>\u3068\u3044\u3046\u70b9\u3002\u305d\u3057\u3066\u3082\u3046\u4e00\u3064\u306e\u6559\u8a13\u306f\u3001\u6700\u9069\u5316\u306b\u306f\u76f8\u4e92\u4f5c\u7528\u304c\u3042\u308a\u3001\u300c\u6a5f\u80fd\u3092\u5168\u90e8ON\u300d\u306b\u3059\u308c\u3070\u3088\u3044\u308f\u3051\u3067\u306f\u306a\u3044\u3001\u3068\u3044\u3046\u3053\u3068\u3060\u3002<\/p>\n<p>\u30ed\u30fc\u30ab\u30ebAI\u3092\u65e5\u5e38\u7684\u306b\u4f7f\u3046\u306a\u3089\u3001\u30e2\u30c7\u30eb\u540d\u3060\u3051\u3067\u306a\u304f<strong>\u30d0\u30c3\u30af\u30a8\u30f3\u30c9\u30fb\u91cf\u5b50\u5316\u30fb\u6295\u6a5f\u7684\u30c7\u30b3\u30fc\u30c9\u30fb\u540c\u6642\u5b9f\u884c\u6570<\/strong>\u3092\u542b\u3081\u3066\u69cb\u6210\u3092\u9078\u3076\u6642\u4ee3\u306b\u306a\u3063\u3066\u304d\u305f\u3002<\/p>\n<h2>\u4e00\u6b21\u60c5\u5831\u30fb\u95a2\u9023\u8cc7\u6599<\/h2>\n<ul>\n<li><a href=\"https:\/\/developer.nvidia.com\/blog\/3-ways-to-boost-moe-inference-on-nvidia-dgx-spark\/\" target=\"_blank\" rel=\"noopener noreferrer\">NVIDIA Technical Blog\uff1a3 Ways to Boost MoE Inference on NVIDIA DGX Spark\uff082026\u5e7410\u670827\u65e5\uff09<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/inclusionAI\/ling-cookbook\/blob\/main\/guide\/local-deploy\/ling-3.0-flash\/dgx-spark-sglang-ling-3-flash-mxfp4-humming.ipynb\" target=\"_blank\" rel=\"noopener noreferrer\">Inclusion AI\u516c\u5f0f\uff1aDGX Spark\u5411\u3051Ling 3.0 Flash\uff0fSGLang Cookbook<\/a><\/li>\n<li><a href=\"https:\/\/music.bear-blog.net\/?p=49889\">\u30d9\u30a2\u4e09\u90ce\u30d6\u30ed\u30b0\uff1aDGX Spark\u5b9f\u6e2c Strata vs llama.cpp<\/a><\/li>\n<\/ul>\n<p><small>\u203b2026\u5e7410\u670828\u65e5\u6642\u70b9\u3067\u516c\u958b\u6e08\u307f\u306e\u8cc7\u6599\u306b\u57fa\u3065\u304f\u89e3\u8aac\u3067\u3059\u3002\u6570\u5024\u306fNVIDIA\u516c\u8868\u306e\u691c\u8a3c\u7d50\u679c\u3067\u3042\u308a\u3001\u30d9\u30a2\u4e09\u90ce\u30d6\u30ed\u30b0\u72ec\u81ea\u306e\u8ffd\u8a66\u7d50\u679c\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u74b0\u5883\u30fb\u30e2\u30c7\u30eb\u30fb\u8ca0\u8377\u306b\u3088\u3063\u3066\u6027\u80fd\u306f\u7570\u306a\u308a\u307e\u3059\u3002<\/small><\/p>\n","protected":false},"excerpt":{"rendered":"<p>NVIDIA\u304c2026\u5e7410\u670827\u65e5\u306b\u516c\u958b\u3057\u305fDGX Spark\u306e\u63a8\u8ad6\u9ad8\u901f\u53163\u624b\u6cd5\u3092\u89e3\u8aac\u3002Ling-3.0-flash\u3067\u751f\u6210\u901f\u5ea642.5%\u5411\u4e0a\u3001\u4e00\u65b9185%\u306f\u5408\u6210\u30c6\u30b9\u30c8\u9650\u5b9a\u3002Humming\u30fbFP8\u30fbMTP\u306e\u52b9\u679c\u3068\u3001\u9006\u306b\u9045\u304f\u306a\u308b\u8a2d\u5b9a\u306e\u843d\u3068\u3057\u7a74\u3092\u6574\u7406\u3057\u307e\u3059\u3002<\/p>\n","protected":false},"author":2,"featured_media":50246,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"pochipp_cv_data":"","swell_btn_cv_data":"","footnotes":"","_wp_rev_ctl_limit":""},"categories":[360],"tags":[365],"class_list":["post-50247","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-360","tag-dgx-spark"],"_links":{"self":[{"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=\/wp\/v2\/posts\/50247","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=50247"}],"version-history":[{"count":1,"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=\/wp\/v2\/posts\/50247\/revisions"}],"predecessor-version":[{"id":50248,"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=\/wp\/v2\/posts\/50247\/revisions\/50248"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=\/wp\/v2\/media\/50246"}],"wp:attachment":[{"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=50247"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=50247"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/music.bear-blog.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=50247"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}