{"id":779,"date":"2025-03-25T23:40:02","date_gmt":"2025-03-25T14:40:02","guid":{"rendered":"http:\/\/kyunam.com\/?p=779"},"modified":"2025-03-26T00:02:08","modified_gmt":"2025-03-25T15:02:08","slug":"gtc-2025-keynote-one-point-lesson-part-ii","status":"publish","type":"post","link":"http:\/\/kyunam.com\/?p=779","title":{"rendered":"GTC 2025 Keynote One Point Lesson (Part II)"},"content":{"rendered":"\n

\uc774\ubc88 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 GTC 2025 \ud0a4\ub178\ud2b8 \ubc1c\ud45c \ub0b4\uc6a9 \uc911 Hopper\uc758 \uc131\ub2a5 \ud5a5\uc0c1\uc744 \uc124\uba85\ud558\ub294 \uadf8\ub798\ud504\ub97c \ud574\uc11d\ud574 \ubcf4\uc790. \uc820\uc2a8 \ud669\uc774 \ubc1c\ud45c \uc911\uc5d0 “\uc6b0\ub9ac\ub294 \uc218\ud559\uc801\uc778 \ud45c\ud604\uc744 \ub300\uc911\uc801\uc778 \uc2a4\ud53c\uce58\uc5d0\uc11c \uc804\ub2ec\ud558\ub294 \uba87 \uc548 \ub418\ub294 \ud68c\uc0ac\ub2e4”\ub77c\uace0 \ub9d0\ud558\uae30\ub3c4 \ud588\ub2e4. \ub18d\ub2f4\ucc98\ub7fc \ub4e4\ub9ac\uc9c0\ub9cc, \uc2e4\uc81c\ub85c \ud574\ub2f9 \uadf8\ub798\ud504\ub97c \uc81c\ub300\ub85c \uc774\ud574\ud558\ub824\uba74 \uc0c1\ub2f9\ud55c \uc218\uc900\uc758 \ubc30\uacbd \uc9c0\uc2dd\uc774 \ud544\uc694\ud558\ub2e4. \uc774\ubc88 \uae00\uc5d0\uc11c\ub294 \uadf8 \uadf8\ub798\ud504\uac00 \ub2f4\uace0 \uc788\ub294 \ud575\uc2ec\uc801\uc778 \uc758\ubbf8\ub97c \uc54c\uc544\ubcf4\uc790. \uc6b0\uc120 \uadf8 \ubb38\uc81c(?)\uc758 \uadf8\ub798\ud504\ub97c \uc0b4\ud3b4 \ubcf4\uc790.<\/p>\n\n\n\n

\"\"
GTC 2025 Keynote\uc5d0\uc11c \uc18c\uac1c\ub41c \uc131\ub2a5 \ube44\uad50 (GTC 2025 Keynote \ub3d9\uc601\uc0c1\uc5d0\uc11c \ucea1\ucc98)<\/em><\/figcaption><\/figure>\n\n\n\n

\uc2dc\uc791\ubd80\ud130 \uc5b4\uc9c8 \uc5b4\uc9c8 \ud558\ub2e4. \uc6b0\uc120 X\ucd95\uc744 \uc0b4\ud3b4\ubcf4\uba74, \uc774\ub294 \uc2e4\uc81c AI \uc11c\ube44\uc2a4\ub97c \uc81c\uacf5\ud560 \ub54c \uc0ac\uc6a9\uc790\uac00 \uc5bc\ub9c8\ub098 \ube60\ub974\uac8c \ud1a0\ud070 \uc751\ub2f5\uc744 \ubc1b\uc744 \uc218 \uc788\ub294\uc9c0\ub97c \ub098\ud0c0\ub0b4\ub294 \uc9c0\ud45c\uc774\ub2e4. \uc774 \uac12\uc740 \ub2f9\uc5f0\ud788 \ub192\uc744\uc218\ub85d \uc88b\uc9c0\ub9cc, \uadf8\uc5d0 \ub530\ub978 \ubc18\ub300 \uae09\ubd80\uac00 \uc874\uc7ac\ud55c\ub2e4. \ubc14\ub85c Y\ucd95\uc774\ub2e4. Y\ucd95\uc740 \ub2e8\uc704 \uc804\ub825\ub2f9 \uc0dd\uc131\ud560 \uc218 \uc788\ub294 \ud1a0\ud070 \uc218, \uc989 \ub3d9\uc77c\ud55c \uc804\ub825 \uc18c\ube44 \ud558\uc5d0\uc11c \uc5bc\ub9c8\ub098 \ub9ce\uc740 TPS(Tokens per Sencond) \/ MW(Mega Watt)\ub97c \ub0bc \uc218 \uc788\ub294\uac00\ub97c \uc758\ubbf8\ud55c\ub2e4. \uc751\ub2f5 \uc18d\ub3c4\ub97c \ube60\ub974\uac8c \ud558\ub824\uba74 \ub354 \ub9ce\uc740 \uc5f0\uc0b0\uc744 \uc218\ud589\ud574\uc57c \ud558\ubbc0\ub85c, \ud544\uc5f0\uc801\uc73c\ub85c \uc804\ub825 \uc18c\ubaa8\uac00 \uc99d\uac00\ud558\uac8c \ub41c\ub2e4. \uadf8\ub807\uac8c \ub418\uba74 \ubd84\ubaa8\uac00 \ucee4\uc9c0\uace0, \uacb0\uacfc\uc801\uc73c\ub85c \uc804\ub825 \ub300\ube44 \ud1a0\ud070 \uc218\ub294 \uc904\uc5b4\ub4e4\uac8c \ub41c\ub2e4.<\/p>\n\n\n\n

\uc774\uac83\uc774 \uc774 \uadf8\ub798\ud504\uc758 \uae30\ubcf8\uc801\uc778 \ud615\uc0c1\uc774\ub2e4. \uadf8\ub798\ud504\uc758 \ubaa9\uc801\uc740 \uac04\ub2e8\ud558\ub2e4. \uac1c\uc778 \uc0ac\uc6a9\uc790\uc5d0\uac8c \ube60\ub974\uac8c \uc751\ub2f5\ud558\ub294 \uc18d\ub3c4(X\ucd95)\uc640, \uc2dc\uc2a4\ud15c \uc804\uccb4\uac00 \uc804\ub825 \ud6a8\uc728\uc801\uc73c\ub85c \uc5bc\ub9c8\ub098 \ub9ce\uc740 \ud1a0\ud070\uc744 \uc0dd\uc0b0\ud558\ub294\uc9c0(Y\ucd95) \uc774 \ub450 \uac00\uc9c0\ub97c \ub3d9\uc2dc\uc5d0 \uace0\ub824\ud558\ub294 \uac83\uc774\ub2e4. \uadf8\ub9ac\uace0 \uc774 \ub450 \uc9c0\ud45c\uac00 \ub9cc\ub098\ub294 \uadf8\ub798\ud504 \uc0c1\uc758 \ud55c \uc9c0\uc810\uc5d0\uc11c, X\uc640 Y\uac00 \ub9cc\ub4e4\uc5b4\ub0b4\ub294 \uba74\uc801\uc774 \uac00\uc7a5 \ub113\uc5b4\uc9c0\ub294, \uc989 \ucd5c\uc801\ud654\ub41c \uade0\ud615\uc810(Optimal Point)\uc744 \ucc3e\ub294 \ub370\uc5d0 \uc774 \uadf8\ub798\ud504\uac00 \uc0ac\uc6a9\ub41c\ub2e4. \uc820\uc2a8 \ud669\uc774 \ub9d0\ud55c \u201c\uc218\ud559\uc801\uc778 \uac1c\ub150\u201d\uc774 \ubc14\ub85c \uc774 \uc9c0\uc810\uc5d0 \ud574\ub2f9\ud55c\ub2e4. \uadf8\ub9ac\uace0 \uc774 \ucd5c\uc801 \uc870\ud569\uc744 \ucc3e\uc544\ub0b4\uae30 \uc704\ud574 \ud65c\uc6a9\ub41c \uac83\uc774 \ubc14\ub85c NVIDIA Dynamo\uc774\ub2e4. \uadf8\ub798\ud504\uc5d0\uc11c \ud655\uc778\ud560 \uc218 \uc788\ub4ef\uc774, Hopper Architecture, Blackwell architecture \ubaa8\ub450 NVIDIA Dynamo\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \ucd5c\uc801\ud654\uac00 \uc801\uc6a9\ub41c \ubaa8\uc2b5\uc744 \ubcf4\uc5ec\uc900\ub2e4.<\/p>\n\n\n\n

\"\"
GTC 2025 Keynote\uc5d0\uc11c \uc18c\uac1c\ub41c \uc131\ub2a5 \uac1c\ub150 (GTC 2025 Keynote \ub3d9\uc601\uc0c1\uc5d0\uc11c \ucea1\ucc98)<\/em><\/figcaption><\/figure>\n\n\n\n

\uadf8\ub807\ub2e4\uba74 Hopper architecture\uc5d0\uc11c Blackwell architecture\ub85c \ubc1c\uc804\ud558\uba74\uc11c \uadf8\ub798\ud504\uac00 \uc6b0\uc0c1\ud5a5 \ubc29\ud5a5\uc73c\ub85c \uc774\ub3d9\ud560 \uc218 \uc788\uc5c8\ub358 \uc774\uc720\ub294 \ubb34\uc5c7\uc77c\uae4c. \uc774 \uc9c8\ubb38\uc5d0 \ub300\ud55c \ud78c\ud2b8\ub294 \uc55e\uc11c \uc18c\uac1c\ud55c Inference Scale \uadf8\ub798\ud504\uc5d0\uc11c \ucc3e\uc744 \uc218 \uc788\ub2e4. \uc544\ub798 \uadf8\ub9bc\uc744 \ubcf4\uba74 X\ucd95\uacfc Y\ucd95\uc740 \ub3d9\uc77c\ud558\uac8c \uad6c\uc131\ub418\uc5b4 \uc788\uc73c\uba70, \uadf8\ub798\ud504\uc758 \uada4\uc801\uc774 \uc544\ub798\uc5d0\uc11c \uc704\ub85c \uc6b0\uc0c1\ud5a5\uc73c\ub85c \uc774\ub3d9\ud55c \uc6d0\uc778\uc744 \uba87 \uac00\uc9c0 \ud56d\ubaa9\uc73c\ub85c \uc124\uba85\ud558\uace0 \uc788\ub2e4. FLOPS\uc758 \uc99d\uac00, HBM \uba54\ubaa8\ub9ac \ub300\uc5ed\ud3ed \ud5a5\uc0c1, \uadf8\ub9ac\uace0 Software\uc640 Architecture\uc758 \ubcc0\ud654\uac00 \uadf8\uac83\uc774\ub2e4. \uc774 \uc911 \uc18c\ud504\ud2b8\uc6e8\uc5b4\ub294 NVIDIA Dynamo\ub97c \uc758\ubbf8\ud558\ub294 \uac83\uc73c\ub85c \ubcf4\uc774\uace0, architecture\ub294 \uc138\ub300\ubcc4 GPU \uad6c\uc870\uc778 Hopper, Grace Blackwell\uc744 \uc9c0\uce6d\ud558\ub294 \uac83\uc73c\ub85c \ud574\uc11d\ub41c\ub2e4. \uc131\ub2a5 \ud5a5\uc0c1\uc740 \uc5b8\uc81c\ub098 \uc5ec\ub7ec \uc694\uc778\uc774 \ubcf5\ud569\uc801\uc73c\ub85c \uc791\uc6a9\ud55c \uacb0\uacfc\uc774\uc9c0\ub9cc, \uc774\ubc88 \uacbd\uc6b0\ub294 \ud2b9\ud788 \ub354 \ubcf5\uc7a1\ud558\uac8c \uc5bd\ud600 \uc788\ub294 \uad6c\uc870\ub97c \ubcf4\uc778\ub2e4. \ub530\ub77c\uc11c Hopper\uc640 Blackwell\uc758 \ucc28\uc774\ub97c \uc774\ud574\ud558\ub294 \uac83\ub9cc\uc73c\ub85c\ub294 \ubd80\uc871\ud558\uba70, \uc774\ub4e4\uc774 NVL8, MVL72\uc640 \uac19\uc740 \ub2e8\uc77c \ub799 \ub2e8\uc704 \uc2dc\uc2a4\ud15c\uc73c\ub85c \uc5b4\ub5bb\uac8c \ud655\uc7a5\ub418\uc5c8\ub294\uc9c0\ub3c4 \ud568\uaed8 \uace0\ub824\ud574\uc57c \ud55c\ub2e4. \ub2e4\ub9cc \ud604\uc7ac \uc2dc\uc810\uc5d0\uc11c\ub294 Blackwell Architecture\uc758 White paper\uac00 Grace Blackwell Architecture\ub85c \ud1b5\ud569\ub418\uc5b4 \uc124\uba85\ub418\uace0 \uc788\uc5b4, Blackwell \ub2e8\ub3c5 Architecture\uc758 \uc131\ub2a5\uc744 \ubd84\ub9ac\ud574 Hopper Architecture\uc640 1\ub300 1\ub85c \ub300\uc751\ud574\uc11c \uac80\ud1a0\ud558\uae30\ub294 \uc5b4\ub835\ub2e4. \uac04\uc811\uc801\uc73c\ub85c RTX \uacc4\uc5f4 Blackwell Architecture\ub97c \ud1b5\ud574 \uc720\ucd94\ud560 \uc218\ub3c4 \uc788\uc9c0\ub9cc, \ub370\uc774\ud130\uc13c\ud130\uc6a9 Architecture\uc640 \ucee8\uc288\uba38\uc6a9 Architecture\ub294 \ubaa9\uc801\uacfc \uad6c\uc870\uac00 \ub2e4\ub974\uae30 \ub54c\ubb38\uc5d0 \uc815\ud655\ud55c \ube44\uad50\ub294 \uc5b4\ub835\ub2e4. (\uc774 \ubd80\ubd84\uc740 \uc791\uac00 \ubcf8\uc778\uc758 \uc815\ubcf4\uac00 \ubd80\uc871\ud560 \uc218 \uc788\uc73c\ub2c8 \uac10\uc548\ud558\uace0 \uc77d\uc5b4 \uc8fc\uae38 \ubc14\ub78c) \ub610\ud55c, GTC Keynote\uc5d0\uc11c \ubc1c\ud45c\ub41c \uadf8\ub798\ud504\uc5d0 \ud45c\uc2dc\ub41c Hopper\uac00 Grace Hopper\ub97c \uc758\ubbf8\ud558\ub294 \uac83\uc778\uc9c0, Hopper\ub97c \ub73b\ud558\ub294 \uac83\uc778\uc9c0 \uba85\ud655\ud558\uc9c0 \uc54a\uc544 \ube44\uad50 \uc9c0\uc810 \uc790\uccb4\ub3c4 \ub2e4\uc18c \uc560\ub9e4\ud55c \uce21\uba74\uc774 \uc788\ub2e4.<\/p>\n\n\n\n

\uc774\ub7ec\ud55c \uc774\uc720\ub85c, \uacf5\uc815\ud55c \ube44\uad50\ub97c \uc704\ud574 Grace Hopper\uc640 Grace Blackwell, \ud2b9\ud788 NVL\uc744 \ud1b5\ud574\uc11c \ub799 \ub2e8\uc704\ub85c \uad6c\ucd95\ub41c \ud658\uacbd \uae30\uc900\uc5d0\uc11c \ubd84\uc11d\ud558\uc600\ub2e4. \ucc38\uace0\ub85c Cerebras\uc5d0\uc11c\ub294 \uc774\ub7ec\ud55c \uad6c\uc131\uc744 appliance mode\ub77c\uace0 \ubd80\ub974\uc9c0\ub9cc, NVIDIA\ub294 \uc774\uc5d0 \ub300\ud55c \uba85\ud655\ud55c \uba85\uce6d\uc744 \uc0ac\uc6a9\ud558\uc9c0 \uc54a\ub294\ub2e4. \ub530\ub77c\uc11c \uc774 \uae00\uc5d0\uc11c\ub294 \ud3b8\uc758\uc0c1 \ud574\ub2f9 \uad6c\uc131\uc744 \ub2e8\uc77c \ub799 \uc2dc\uc2a4\ud15c, \ub610\ub294 \ub799 \ub2e8\uc704 \uc11c\ubc84 \ube44\uad50\ub77c\ub294 \uc784\uc758\uc758 \ud45c\ud604\uc73c\ub85c \uc0ac\uc6a9\ud558\uace0\uc790 \ud55c\ub2e4.<\/p>\n\n\n\n

\n
\n
\"\"
\ucd9c\ucc98: ‘NVIDIA GH200 Grace Hopper Superchip Architectue’ <\/em>White Paper<\/figcaption><\/figure>\n\n\n\n
\"\"
\ucd9c\ucc98: ‘NVIDIA H100 Tensor Core GPU Architecture’<\/em> White Paper<\/figcaption><\/figure>\n<\/div>\n\n\n\n
\n
\"\"
\ucd9c\ucc98: ‘NVIDIA Blackwell Architecture Technical Brief’<\/em> White Paper<\/figcaption><\/figure>\n<\/div>\n<\/div>\n\n\n\n

\uc2a4\ud399 \ubb38\uc11c\ub97c \uc77d\ub2e4 \ubcf4\uba74 \ub0b4\uc6a9\uc774 \ubc29\ub300\ud558\uace0 \ubcf5\uc7a1\ud574\uc11c \uc27d\uac8c \ud53c\ub85c\uac10\uc744 \ub290\ub07c\uac8c \ub41c\ub2e4. \ubaa8\ub4e0 \ud56d\ubaa9\uc744 \ub2e4 \ud655\uc778\ud560 \ud544\uc694\ub294 \uc5c6\uae30 \ub54c\ubb38\uc5d0, \uc774\ubc88\uc5d0\ub294 \uc67c\ucabd\uc5d0 \uc788\ub294 Table 3\ubc88\uacfc \uc624\ub978\ucabd\uc5d0 \uc788\ub294 Table 1\ubc88\uc5d0 \uc9d1\uc911\ud574 \ubcf4\uace0\uc790 \ud55c\ub2e4. \ub450 \ud45c\ub294 \uc11c\ub85c \ub2e4\ub978 white paper\uc5d0\uc11c \uac00\uc838\uc628 \uc790\ub8cc\uc774\ubbc0\ub85c, \ub0b4\uc6a9\uc744 \ud574\uc11d\ud560 \ub54c \ucd9c\ucc98\ub97c \uba85\ud655\ud788 \uad6c\ubd84\ud560 \ud544\uc694\uac00 \uc788\ub2e4. \uac00\uc7a5 \ud070 \ucc28\uc774\uc810 \uc911 \ud558\ub098\ub294 Blackwell Architecture\uc640 Hopper Architecture\uac00 \uc9c0\uc6d0\ud558\ub294 \ub370\uc774\ud130 \ud0c0\uc785\uc774 \ub2e4\ub974\ub2e4\ub294 \uc810\uc774\ub2e4. Blackwell\uc740 FP4\ub97c \ud3ec\ud568\ud55c FP4, FP8, FP16\uae4c\uc9c0 \ud3ed\ub113\uc740 \uc2e4\uc218 \uc5f0\uc0b0 \ud0c0\uc785\uc744 \uc9c0\uc6d0\ud558\uc9c0\ub9cc, Hopper\ub294 FP16\uae4c\uc9c0\ub9cc \uc9c0\uc6d0\ud55c\ub2e4. \uc989 Hopper\ub294 \ub2e8\uc815\ubc00\ub3c4(Single Precision) \uc774\ud558\uc758 \uc5f0\uc0b0\uc744 \ud558\ub4dc\uc6e8\uc5b4 \uc218\uc900\uc5d0\uc11c \uc9c1\uc811 \ucc98\ub9ac\ud558\uc9c0 \uc54a\ub294\ub2e4\ub294 \uac83\uc744 \uc758\ubbf8\ud55c\ub2e4. Hopper Architecture\uc5d0\uc11c\ub294 FP8\uc774\ub098 FP16 \uc5f0\uc0b0\uc744 \ud558\ub824\uba74, \uae30\uc874 FP32 \uc5f0\uc0b0 \uc720\ub2db\uc744 \ucabc\uac1c\uc5b4 \uc5d0\ubbac\ub808\uc774\uc158 \ubc29\uc2dd\uc73c\ub85c \ucc98\ub9ac\ud574\uc57c \ud558\ubbc0\ub85c, \uc5f0\uc0b0 \uc790\uccb4\ub294 \uac00\ub2a5\ud558\uc9c0\ub9cc \uc131\ub2a5 \uce21\uba74\uc5d0\uc11c\ub294 \ubd88\ub9ac\ud558\ub2e4. \ud558\ub4dc\uc6e8\uc5b4\uc801\uc73c\ub85c \uc9c1\uc811 \uc9c0\uc6d0\ud558\uc9c0 \uc54a\uae30 \ub54c\ubb38\uc5d0, \uc5f0\uc0b0 \uc18d\ub3c4\uac00 \ub290\ub9b4 \uc218\ubc16\uc5d0 \uc5c6\ub2e4.<\/p>\n\n\n\n

\uc774\ub7ec\ud55c \ucc28\uc774\ub294 white paper\uc758 \uc608\uc2dc\ub97c \ud1b5\ud574\uc11c\ub3c4 \ud655\uc778\ud560 \uc218 \uc788\ub2e4. \uc608\ub97c \ub4e4\uc5b4, FP8\uc744 FP16 \ucf54\uc5b4\ub85c \ub204\uc801(accumulation) \uc5f0\uc0b0\ud560 \ub54c\uc758 \uc131\ub2a5 \uc218\uce58\ub97c \ubcf4\uba74, Hopper\ub294 ALU \uae30\ubc18 \ucc98\ub9ac\ub85c \uc778\ud574 \ud6a8\uc728\uc774 \ub5a8\uc5b4\uc9c0\ub294 \ubc18\uba74, Blackwell\uc740 Tensor Core \uc218\uc900\uc5d0\uc11c \uc9c1\uc811 \uc5f0\uc0b0\uc744 \ucc98\ub9ac\ud560 \uc218 \uc788\uc5b4 \ud6e8\uc52c \ube60\ub978 \uc131\ub2a5\uc744 \ub0bc \uc218 \uc788\ub2e4. \uc815\ubc00\ub3c4 \uc218\uc900\uc5d0 \ub530\ub77c FP16, FP32, FP64 \uc5f0\uc0b0\uc774 \uc5b4\ub514\uc11c \ucc98\ub9ac\ub418\ub294\uc9c0\ub3c4 \uc911\uc694\ud55c \ucc28\uc774\uc810\uc774\ub2e4. Hopper\uc5d0\uc11c\ub294 \uc774\ub4e4 \uc5f0\uc0b0\uc774 \ubaa8\ub450 \uc77c\ubc18\uc801\uc778 ALU\uc5d0\uc11c \uc218\ud589\ub418\uc9c0\ub9cc, Blackwell\uc5d0\uc11c\ub294 Tensor Core\uc5d0\uc11c \uc9c1\uc811 \ucc98\ub9ac\ub418\ub3c4\ub85d \uc124\uacc4\ub418\uc5b4 \uc788\ub2e4. \uc774\ub294 \uace7 \ubc94\uc6a9 AI \uc5f0\uc0b0\ubfd0 \uc544\ub2c8\ub77c \uace0\uc815\ubc00 \uc218\uce58 \uacc4\uc0b0\uae4c\uc9c0\ub3c4 Tensor Core\ub97c \ud1b5\ud574 \uac00\uc18d\ud560 \uc218 \uc788\ub2e4\ub294 \ub73b\uc774\ub2e4. Blackwell\uc740 Tensor Core\uc5d0\uc11c Dense Matrix\uc640 Sparse Matrix\ub97c \uc791\uc740 \uc2e4\uc218\ud615 \ub370\uc774\ud130\uc5d0\uc11c\ub3c4 \ubaa8\ub450 \uc9c0\uc6d0\ud558\ub294\ub370, \ud2b9\ud788 Sparse Matrix\ub97c \ud558\ub4dc\uc6e8\uc5b4 \ub808\ubca8\uc5d0\uc11c \ucc98\ub9ac\ud55c\ub2e4\ub294 \uc810\uc774 \uc8fc\ubaa9\ud560 \ub9cc\ud558\ub2e4. \uc774 \uae30\ub2a5\uc740 \ub300\uaddc\ubaa8 \ubaa8\ub378\uc5d0\uc11c \ube48\ub3c4\uac00 \ub0ae\uc740 \uac00\uc911\uce58\ub97c \ud6a8\uc728\uc801\uc73c\ub85c \uc2a4\ud0b5\ud560 \uc218 \uc788\uc5b4, \uc5f0\uc0b0 \ud6a8\uc728\uc744 \ud68d\uae30\uc801\uc73c\ub85c \uac1c\uc120\ud560 \uc218 \uc788\ub294 \uc911\uc694\ud55c \uc694\uc18c\uac00 \ub41c\ub2e4.<\/p>\n\n\n\n

\ub610\ud55c Blackwell\uc740 8\ube44\ud2b8 \uc815\uc218(INT8)\ub3c4 \uc9c0\uc6d0\ud558\ub294\ub370, \uc774\ub294 \ud574\ub2f9 \ub370\uc774\ud130 \ud0c0\uc785\uc744 \ucc98\ub9ac\ud560 \uc218 \uc788\ub294 \uc804\uc6a9 \uc0b0\uc220 \uc7a5\uce58(ALU)\uac00 \ud0d1\uc7ac\ub418\uc5b4 \uc788\uc74c\uc744 \ub73b\ud55c\ub2e4. \uc774\ub7ec\ud55c \ud558\ub4dc\uc6e8\uc5b4\uc801 \uc9c0\uc6d0\uc740 \ubaa8\ub378 \uc591\uc790\ud654(quantization)\uc5d0 \uc720\ub9ac\ud558\uba70, \ub354 \uc791\uc740 \ub370\uc774\ud130 \ub2e8\uc704\ub97c \ube60\ub974\uac8c \ucc98\ub9ac\ud560 \uc218 \uc788\uc5b4 \ucd94\ub860 \uc18d\ub3c4 \ud5a5\uc0c1\uc5d0 \uae30\uc5ec\ud55c\ub2e4.<\/p>\n\n\n\n

White Paper\ub294 \ub2e4\uc591\ud55c \uc815\ubcf4\uac00 \ud63c\ud569\ub418\uc5b4 \uc788\uc73c\ub2c8 \uba87\uac00\uc9c0\ub9cc \ucd94\ub824\uc11c \ud655\uc778\ud574 \ubcf4\ub3c4\ub85d \ud558\uc790.<\/p>\n\n\n\n

\ube44\uad50 \ud56d\ubaa9<\/strong><\/td>Grace Hopper<\/td>Grace Blackwell<\/td><\/td><\/tr>
FP8 \/ FP4 \uc5f0\uc0b0<\/strong><\/td>4 PFLOPs
(FP16\uc73c\ub85c \uc5f0\uc0b0, 2\uac1c Hopper \ud569\uc0b0)<\/td>		40 PFLOPs
(2\uac1c Blackwell \ud569\uc0b0)<\/td>		10\ubc30 <\/td><\/tr>
HBM Bandwith<\/strong><\/td>4.8 TB\/s<\/td>16 TB\/s<\/td>\uc57d 4\ubc30<\/td><\/tr>
NVLink Bandwidth<\/strong><\/td>900 GB\/s<\/td>3.6 TB\/s<\/td> 4\ubc30<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

\uac1c\uc120\ub41c \ud56d\ubaa9\ub4e4\uc758 \ud6a8\uacfc\uc758 \uc870\ud569\uc73c\ub85c \uadf8\ub798\ud504\uc758 \uc704\uce58\uac00 \uc6b0\uc0c1\ud5a5\uc774 \ub418\uc5b4 \uc6b0\uc0c1\ub2e8\uc758 \uadf8\ub798\ud504\uac00 \uadf8\ub824\uc9c8 \uac83\uc774\ub2e4. \uc774\ub294 \uc544\uc5d0 Scale \ucee4\uc9c0\uace0 \ucc98\uc74c\uc5d0 \ub9d0\ud55c Dynamo\ub85c \uc120\ubd84\uc704\ub97c \ud0d0\uc0c9\ud558\uba74\uc11c X\ucd95 * Y\ucd95\uc758 \ub113\uc774\uac00 \uac00\uc7a5 \ud070 \uc810\uc744 \ud0d0\uc0c9 \ud558\ub294 \uac83\uc774\ub2e4. \uadf8\ub798\ud504\uc5d0 \ud45c\uc2dc\ud574\uc11c \ubcf4\uba74 \uc544\ub798\uc640 \uac19\ub2e4.<\/p>\n\n\n\n

\"\"
GTC 2025 Keynote\uc5d0\uc11c \uc18c\uac1c\ub41c \uc131\ub2a5 \ube44\uad50 \uadf8\ub798\ud504\uc5d0 \uc758\ubbf8 \uc124\uba85 \ucd94\uac00(GTC 2025 Keynote \ub3d9\uc601\uc0c1\uc5d0\uc11c \ucea1\ucc98)<\/em><\/figcaption><\/figure>\n\n\n\n

\uc5ec\uae30\uae4c\uc9c0 \uc815\ubcf4\ub97c \uc798 \uc18c\ud654\ud588\ub2e4\uba74 \uadf8\ub798\ud504\uc758 \uad6c\uc870\uc640 Performance Jumping Gap\uc758 \uad6c\uc870\ub97c \uc774\ud574\ud560 \uc218 \uc788\uc744 \uac83 \uac19\ub2e4. \uc774\uc81c \ub9c8\uc9c0\ub9c9\uc73c\ub85c NVIDIA Dynamo\uc640 \ucd5c\uc801\uc810 \ud0d0\uc0c9\uc5d0\uc11c \ub098\uc624\ub294 \uc218 \ub9ce\uc740 \uc870\ud569\uc758 \uc758\ubbf8\ub97c \uc54c\uc544 \ubcf4\uace0 \uae34 \uae00\uc744 \ub9c8\ubb34\ub9ac \ud558\ub3c4\ub85d \ud558\uaca0\ub2e4. \uc6b0\uc120 NVIDIA Dynamo\ub97c \uc124\uba85\ud55c Keynote \uc2ac\ub77c\uc774\ub4dc\ub97c \uc0b4\ud3b4 \ubcf4\uc790<\/p>\n\n\n\n

\"\"
GTC 2025 Keynote\uc5d0\uc11c \uc18c\uac1c\ub41c NVIDIA Dynamo(GTC 2025 Keynote \ub3d9\uc601\uc0c1\uc5d0\uc11c \ucea1\ucc98)<\/em><\/figcaption><\/figure>\n\n\n\n

NVIDIA Dynamo\ub294 \ubcc4\ub3c4\uc758 \ud3ec\uc2a4\ud305\uc5d0\uc11c \uc870\uae08 \uc790\uc138\ud788 \ub2e4\ub8e8\ub3c4\ub85d \ud558\uace0 \uc5ec\uae30\uc11c\ub294 \uac1c\ub150\ub9cc \uc18c\uac1c \ud558\ub3c4\ub85d \ud558\uaca0\ub2e4. NVIDIA Dynamo\ub294 LLM Model\uacfc Data Center GPU Architecture\uc5d0 \ub3c5\ub9bd\uc801\uc778 Software Framework\uc774\ub2e4. \uc544\ub798 \ub0b4\uc6a9\uc740 Github\uc758 NVIDIA Dynamo Repository\uc758 Document(https:\/\/github.com\/ai-dynamo\/dynamo\/blob\/main\/docs\/architecture.md<\/a>)\uc758 \uc77c\ubd80 \ubd84\uc774\ub2e4. \uadf8\ub807\ub2e4! \uc624\ud508 \uc18c\uc2a4\uc778\uac83\uc774\ub2e4!<\/p>\n\n\n\n