{"id":863,"date":"2025-07-25T08:29:10","date_gmt":"2025-07-24T23:29:10","guid":{"rendered":"http:\/\/kyunam.com\/?p=863"},"modified":"2025-07-25T08:29:10","modified_gmt":"2025-07-24T23:29:10","slug":"gpu%ec%9d%80-%ec%96%b4%eb%96%bb%ea%b2%8c-ai-%ec%8b%9c%eb%8c%80%ec%9d%98-%ed%95%b5%ec%8b%ac%ec%9d%b4-%eb%90%98%ec%97%88%eb%8a%94%ea%b0%80-npu-xpu","status":"publish","type":"post","link":"http:\/\/kyunam.com\/?p=863","title":{"rendered":"GPU\uc740 \uc5b4\ub5bb\uac8c AI \uc2dc\ub300\uc758 \ud575\uc2ec\uc774 \ub418\uc5c8\ub294\uac00? \u2013\u00a0NPU, XPU"},"content":{"rendered":"\n

AI \uac1c\ubc1c\uc5d0 \ub300\uaddc\ubaa8 GPU\ub294 \ud544\uc218\uc801\uc778 \uc694\uc18c\ub85c \uac04\uc8fc\ub41c\ub2e4. \ud55c\uad6d \uc815\ubd80\ub294 2025\ub144 \ucd08 \ubd80\uc871\ud55c GPU\ub97c \uacf5\uacf5 \uc778\ud504\ub77c \ud615\ud0dc\ub85c \uc81c\uacf5\ud560 \uacc4\ud68d\uc744 \ubc1c\ud45c\ud558\uc600\ub2e4. \uc774 \uc2dc\ub9ac\uc988\uc758 \ud575\uc2ec \uc8fc\uc81c\uc774\uae30\ub3c4 \ud558\ub2e4. \uadf8\ub7f0\ub370 \uad6d\ub0b4\uc678\uc801\uc73c\ub85c \ub2e4\uc591\ud55c \uc124\uacc4\ub97c \uae30\ubc18\uc73c\ub85c \ub2e4\uc591\ud55c Hardware Accelerator(Co-processor)\uac00 \ucd9c\uc2dc\ub418\uace0 \uc788\uace0, \uc131\ub2a5 \ud3c9\uac00 \uacb0\uacfc\ub97c \uacf5\ud45c\ud558\uba74\uc11c \uc790\uc2e0\ub4e4\uc758 \uc0c8\ub85c\uc6b4 \ud558\ub4dc\uc6e8\uc5b4\ub97c \ubf51\ub0b8\ub2e4. \ud558\uc9c0\ub9cc, \ud604\uc2e4\uc740 AI \ubaa8\ub378 \ud559\uc2b5\uc740 NVIDIA\uc758 GPU\uac00 \uac70\uc758 \ub3c5\uc810\uc801\uc73c\ub85c \uc0ac\uc6a9\ub418\uace0 \uc788\ub2e4. \uad6d\ub0b4 S\ubaa8\uc0ac\ub294 AMD\uc758 \ud558\ub4dc\uc6e8\uc5b4\ub97c \ub300\uaddc\ubaa8\ub85c \uad6c\ub9e4\ud574\uc11c \uc0ac\uc6a9\ud55c\ub2e4\ub294 \uae30\uc0ac\uac00 \uc788\uc5c8\uc9c0\ub9cc, \uadf8\uc678 \ub2e4\ub978 \ud68c\uc0ac\ub4e4\uc774 AMD\uc758 \ud558\ub4dc\uc6e8\uc5b4\ub97c AI \ud559\uc2b5\uc6a9\uc73c\ub85c \uc0ac\uc6a9\ud55c\ub2e4\ub294 \uc774\uc57c\uae30\ub294 \uc798 \ub4e4\ub9ac\uc9c0 \uc54a\ub294\ub2e4. \uadf8\ub7f0\ub370 \uc815\ubd80\uc5d0\uc11c\ub294 AI Chip \ud65c\uc131\ud654\ub97c AI \uad6d\ub0b4 \uae30\uc220 \uac1c\ubc1c\uc758 \ud558\ub098\uc758 \ucd95\uc73c\ub85c \uc0bc\ub294\ub4ef\ud55c\ub2e4. \uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 NPU\uac00 \uc65c GPU \ubcf4\ub2e4 \ud2b9\uc815 \uc5f0\uc0b0\uc5d0\uc11c \uc131\ub2a5\uc774 \ub354 \uc798 \ub098\uc624\ub294\uc9c0 \uc54c\uc544 \ubcf4\uace0 GPU \uc804\uc131\uc2dc\ub300\uc5d0 NPU\uc758 \ud65c\uc6a9\ubc95\uacfc \ud655\ub300 \ubc29\uc548\uc5d0 \ub300\ud574\uc11c \ub17c\uc758 \ud574\ubcf4\ub3c4\ub85d \ud558\uaca0\ub2e4. \uadf8\ub9ac\uace0 \uc88b\uc740 \uc131\ub2a5\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 NPU, TPU\ub4f1\uc774 \uc65c GPU\ubcf4\ub2e4 \uc0ac\uc6a9\uc774 \ub35c\ub418\uace0 \uc788\ub294\uc9c0\ub3c4 \uc54c\uc544 \ubcf4\uaca0\ub2e4. \uc6b0\uc120 NVIDIA\uc758 GPU\uc758 \uacf5\uace0\ud568\uc774 \uc5b4\ub5bb\uac8c \ubc1c\uc0dd\ud588\ub294\uc9c0 \uba3c\uc800 \uc54c\uc544 \ubcf4\uaca0\ub2e4.<\/p>\n\n\n\n

NVIDIA GPU\ub85c \uacf5\uace0\ud788 \ub9cc\ub4e4\uc5b4\uc9c4 \uae30\ub77c\uc131<\/h3>\n\n\n\n

\uc5ec\ub7ec \ub274\uc2a4 \ub610\ub294, \uc5ec\ub7ec \ud64d\ubcf4\uc5d0\uc11c GPU\uc758 \uc131\ub2a5\uc744 \ucd94\uc6d4\ud588\ub2e4\ub294 \uac83\uc744 \ub4e4\uc744 \uc218 \uc788\ub2e4. \uadf8\ub7f0\ub370 \uc5ec\uc804\ud788 NVIDIA\uc758 GPU\uac00 \uac00\uc7a5 \ub9ce\uc774 \ud65c\uc6a9\ub418\uace0 \uc788\ub2e4. \uc65c \uadf8\ub7f4\uae4c? NPU, TPU \uae30\uc5c5\uc758 \uc131\ub2a5\uc740 \uac70\uc9d3\uc774 \uc5c6\uc74c\uc5d0 \ubd84\uba85\ud558\uae30 \ub54c\ubb38\uc5d0 \ub354\ub354\uc6b1 \uc758\uc544\ud558\ub2e4. \uc65c \uadf8\ub7f4\uae4c? \ud06c\uac8c 2\uac00\uc9c0 \uc6d0\uc778\uc774 \uc788\ub2e4. \uccab \ubc88\uc9f8\ub294 NPU\ub294 AI\uc5d0 \ud2b9\ud654\ub418\uc5b4 \uc788\uace0, \ud2b9\uc815 \ubaa8\ub378\uc774\ub098 \ud2b9\uc815 \uc5f0\uc0b0\uc5d0 \ucd5c\uc801\ud654 \ub418\uc5b4 \uc788\ub2e4. \ub2e4\uc2dc \ubc1c\ud574\uc11c \ubc94\uc6a9\uc131 \uce21\uba74\uc5d0\uc11c \ubd80\uc871\ud568\uc774 \uc788\ub2e4. \ubc18\ub300\uae09\ubd80\ub85c, \uc774\ub85c \uc778\ud574\uc11c \ube44\uc6a9 \ubc0f \ud2b9\uc815 \ubaa9\uc801\uc5d0 \uc131\ub2a5\uc774 \ucd5c\ub300\ud654 \ub420 \uc218 \uc788\ub2e4. \ub450 \ubc88\uc9f8\ub294 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc0dd\ud0dc\uacc4(Eco)\/\ud234\uccb4\uc778(Toolchain) \uce21\uba74\uc774\ub2e4. \uc774\ubc88 \ub2e8\ub780\uc5d0\uc11c\ub294 \uc774 \ub450 \ubc88\uc9f8\uc5d0 \ub300\ud574\uc11c \uc774\uc57c\uae30 \ud574\ubcf4\uaca0\ub2e4.

\ucca8\ubd80\ud55c \uadf8\ub9bc\uc740 NVIDIA\uc758 CUDA Software Stack\uc5d0 \ud5c8\uae45\ud398\uc774\uc2a4\uac00 \uac00\uc7a5 \uc704\uc5d0 \uc788\ub294 \uc0dd\ud0dc\uacc4\ub97c \uadf8\ub9b0 \uadf8\ub9bc\uc774\ub2e4. Sora\ub85c 3D\ub85c \ub9cc\ub4e4\uc5b4\uc11c \uc77c\ubd80 \uae00\uc790\uac00 \uae68\uc838 \uc606\uc5d0 \uc6d0\ubcf8 \uadf8\ub9bc\ub3c4 \ud568\uaed8 \ubc30\uce58 \ud558\uc600\ub2e4. \uc6d0 \uadf8\ub9bc\uc5d0\ub294 \uad00\ub828\ub41c \ub0b4\uc6a9 \ucd9c\ucc98\ub3c4 \ud568\uaed8 \ud3ec\ud568\ub418\uc5b4 \uc788\ub2e4.<\/p>\n\n\n\n

\n
\n
\"\"<\/figure>\n<\/div><\/div>\n\n\n\n
\n
\"\"<\/figure>\n<\/div>\n<\/div>\n\n\n\n

\uc6b0\uc120 NVIDIA\uac00 \uc790\uccb4\uc801\uc73c\ub85c \uad6c\ucd95\ud55c CUDA, CUDA-X\ub97c \uc0b4\ud3b4 \ubcf4\uc790. CUDA\ub294 GPU\uc5d0 \ub370\uc774\ud130\ub97c \ubcf4\ub0b4\uace0 \uc5f0\uc0b0\uc6a9 \ucf54\ub4dc\ub97c \uc804\ub2ec\ud558\uae30 \uc704\ud55c \uac00\uc7a5 \uae30\ubcf8\uc801\uc778(Primitive) \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uae30\ubc18\uc73c\ub85c \uae30\uacc4\ud559\uc2b5, \ub525\ub7ec\ub2dd, HPC\ub97c \uc704\ud55c \ucd94\uac00\uc801\uc778 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub4e4\uc774 \ud3ec\uc9c4\ub418\uc5b4 \uc788\ub2e4. \uc774 \uc790\uccb4\uac00 \uc774\ubbf8 \ub108\ubb34 \ubc29\ub300\ud558\uace0, \uc774 \ubc29\ub300\ud568\uc740 1~2\ub144\ub9cc\uc5d0 \uad6c\ucd95\ub41c \uac83\uc774 \uc544\ub2c8\ub2e4. OpenCV\ub97c \ud45c\uc900\uc73c\ub85c \uac00\uc838\uac00\uc9c0 \uc54a\uace0 CUDA\ub97c \uc790\uccb4\uc801\uc73c\ub85c \uac1c\ubc1c\ud574\uc11c \uc9c4\ud589\ud55c \uc2ed\uc218\ub144\uc758 \ub178\ud558\uc6b0\uac00 \ub4e4\uc5b4\uac00 \uc788\ub2e4. \uc774\uac83\uc774 \uccab \ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd(Entry Barrier) \uc774\ub2e4. \uc798 \uae30\uc5b5\ud558\uace0 \ub450 \ubc88\uc9f8\ub85c \ub118\uc5b4 \uac00 \ubcf4\uc790.

\ub450 \ubc88\uc9f8\ub294 NVIDIA\uac00 \uc790\uccb4\uc801\uc73c\ub85c \uad6c\ucd95\ud558\uc9c0 \uc54a\uc740 CUDA\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \uac01\uc885 ML Framework(F\/W)\uc640 MF F\/W\uc5d0\uc11c \uc0ac\uc6a9\ud558\ub294 \uac01\uc885 \uc218\ud559, ML\/DL \ub77c\uc774\ube0c\ub7ec\ub9ac\uc774\ub2e4. \ud150\uc11c\ud50c\ub85c\uc6b0, \ud30c\uc774\ud1a0\uce58, ONNX \ub4f1\uc774 \uc5ec\uae30\uc5d0 \ud574\ub2f9\ub41c\ub2e4. \uadf8\uc678\uc5d0\ub3c4 \uac01\uc885 \uacfc\ud559\uae30\uc220\uc6a9 \uc0c1\uc6a9\/\uc624\ud508 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub4e4\uc774 \uc788\ub294\ub370, \uccab\ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc704\uc5d0 \uc5ce\ud600\uc9c4 \ub450 \ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc774\ub2e4. \ub9ce\uc740 \uc0ac\ub78c\ub4e4\uc774 \uc774 \uae30\uc220\uc7a5\ubcbd\uc5d0 \ub300\ud574\uc11c \uc798 \uc778\uc9c0\ud558\uc9c0 \ubabb\ud558\uace0 \uc788\uace0 \uc774\uc5b4\uc9c8 \uc138\ubc88\uc9f8 \uae30\uc220 \uc7a5\ubcbd\uc774 \uac00\ub2a5\ud55c \uac00\uad50 \uc5ed\ud560\uc744 \ud55c\ub2e4.

\ub9c8\uc9c0\ub9c9\uc758 \uae30\uc220\uc7a5\ubcbd\uc774 \ub9e4\uc6b0 \uc911\uc694\ud568!!!! \uccab \ubc88\uc9f8, \ub450 \ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc740 \ub208\uc5d0 \uc798 \ubcf4\uc774\uac70\ub098 \uc778\uc9c0\ud558\uae30 \uc27d\ub2e4. \uc65c\ub0d0\ud558\uba74 \uc9c1\uc811\uc801\uc778 \uc0b0\ucd9c\ubb3c\uc774 \uba85\uc2dc\uc801\uc73c\ub85c \uc874\uc7ac \ud558\uae30 \ub54c\ubb38\ub514\ub2e4. \ub9c8\uc9c0\ub9c9 \uae30\uc220\uc7a5\ubcbd\uc740 \uccab \ubc88\uc9f8, \ub450 \ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc744 \ud65c\uc6a9\ud55c \ud2b9\ud788 \ub450 \ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc744 \ud65c\uc6a9\ud574\uc11c \uc313\uc544 \uc62c\ub824\uc9c4 \ub178\ud558\uc6b0\uc774\ub2e4. \ub450 \ubc88\uc9f8 \uc7a5\ubcbd\uc744 \uae30\ubc18\uc73c\ub85c \uac01\uc885 ML Architecture(\uac01\uc885 CNN \uc544\ud0a4\ud14d\uccd0, Transformer)\uac00 \ub9cc\ub4e4\uc5b4 \uc9c0\uace0 ML\uc744 \uc11c\ube44\uc2a4\ub85c \uc81c\uacf5\ud560 \uc218 \uc788\ub294 \uc11c\ube59 \ud504\ub808\uc784\uc6cc\ud06c\ub4e4\uc774 \uc874\uc7ac\ud55c\ub2e4. vLLM, SGLang, TensorRT \ub4f1\uc774 \uc774\uc5d0 \ud574\ub2f9\ud55c\ub2e4. \ucd5c\uadfc\uc5d0\ub294 \uc774\ub7f0 \uac01 \uac1c\uc778\uc758, \uc5f0\uad6c \ub2e8\uccb4\uc758 \ub178\ud558\uc6b0\uac00 \ucee8\ud14c\uc774\ub108\uc5d0 \uad6c\uccb4\ud654 \ub418\uc5b4 \ud5c8\uae45\ud398\uc774\uc2a4\uc5d0\uc11c \uc81c\uacf5\ub41c\ub2e4. \ucd5c\uc2e0 Open Source LLM\uc744 vLLM\uc73c\ub85c \uc11c\ube59\ud558\ub294\ub370 \ud558\ub4dc\uc6e8\uc5b4\ub9cc \uc788\ub2e4\uba74 \uc218\uc2ed\ubd84 \uc548\uc5d0 \uc11c\ube44\uc2a4\ub97c \uc81c\uacf5\ud560 \uc218 \uc788\ub294\ub370 \uae30\uc220\uc7a5\ubcbd\uc778 \ub3d9\uc2dc\uc5d0 \ud3b8\ub9ac\uc131\uc744 \uc81c\uacf5\ud55c\ub2e4. \uc774\uac83\uc774 \uc138 \ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc785\ub2c8\ub2e4.

\uccab \ubc88\uc9f8, \ub450 \ubc88\uc9f8 \uadf8\ub9ac\uace0 \uc138\ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc740 \ud2b9\uc815 \uba87\uba85, \ud2b9\uc815 \uba87\uac1c\uc758 \uae30\uc5c5\uc774 \ub2e8\uae30\uac04\uc5d0 \ub9cc\ub4e0 \uc0dd\ud0dc\uacc4\uac00 \uc544\ub2c8\ub2e4. \uc218\uc2ed\ub144\uc5d0 Try and Error\ub97c \ud1b5\ud574\uc11c \uacf5\uace0\ud788 \uc313\uc5ec\uc9c4 \uc0dd\ud0dc\uacc4\uc774\ub2e4. \ub2e4\uc2dc \ucc98\uc74c\uc73c\ub85c \ub3cc\uc544\uac00\ubcf4\uc790. NPU\/TPU \uac1c\ubc1c \uae30\uc5c5\uc758 \uae30\uc220\ub825\uc740 \ub9e4\uc6b0 \ub6f0\uc5b4\ub0a8\uc744 \uc758\uc2ec\ud560 \uc5ec\uc9c0\uac00 \uc5c6\ub2e4. \ub9ce\uc740 \ub178\ub825\uc744 \uae30\uc6b8\uc5ec \uccab \ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc744 \uc77c\uc2dc\uc801\uc73c\ub85c \uadf9\ubcf5\ud560 \uc218 \uc788\ub2e4. \uadf8\ub7f0\ub370 \uc9c0\uae08 \ub9cc\ub4e4\uc5b4\uc9c4 \uc0dd\ud0dc\uacc4\ub97c \uba48\ucdb0\uc788\uc9c0 \uc54a\ub2e4. \uadf8\ub798\uc11c \uccab \ubc88\uc9f8 \uae30\uc220\uc7a5\ubcbd\uc774 \ub2e4\uc2dc \uacf5\uacf5\ud574\uc9c0\uace0 \uc774\ub97c \uae68\uae30 \uc704\ud55c \uc9c0\ud611\uc801\uc778 \ub178\ub825\uc5d0 \ud798\uc774 \ub354\ub354\uc6b1 \ub4e4\uc5b4 \uac04\ub2e4. \ub450 \ubc88\uc9f8, \uc138 \ubc88\uc9f8\uc758 \uae30\uc220\uc7a5\ubcbd\uc740 \uae68\ub294 \uac83\uc740 \uc544\ub9c8\ub3c4 \uc5c4\ub450\ub3c4 \ubabb\ub0bc \uac83\uc774\ub2e4.<\/p>\n\n\n\n

\uc774\ub807\uac8c \uacf5\uace0\ud55c \uae30\uc220\uc7a5\ubcbd\uc774 \uc788\ub294\ub370 NPU\ub97c \uc65c \uc18c\uac1c\ud558\ub0d0\uace0 \ubc18\ubb38\ud560 \uc218 \uc788\uc744 \uac83\uc774\ub2e4. \ud558\uc9c0\ub9cc \uc800 \uc7a5\ubcbd\uc744 \uc798 \uc774\uc6a9\ud558\uba74 NPU\/TPU\ub3c4 \uc774 \uc0dd\ud0dc\uacc4\uc5d0 \ub4e4\uc5b4 \uc62c \uc218 \uc788\uc744 \uac83\uc774\ub77c \uc0dd\uac01\ud558\ub2e4. \uc774 \ubd80\ubd84\uc740 \ub9c8\uc9c0\ub9c9\uc5d0 \uc54c\uc544 \ubcf4\ub3c4\ub85d \ud558\uaca0\ub2e4. \uc774\uc81c NPU \uae30\uc220\uc5d0 \ub300\ud574\uc11c \uc54c\uc544 \ubcf4\uc790.<\/p>\n\n\n\n

NPU – Neural Processing Unit<\/h3>\n\n\n\n

NPU\ub294 Neural Network \uc5f0\uc0b0\uc5d0 \ucd5c\uc801\ud654 \ub418\uc5b4 \uc788\ub294 \ud558\ub4dc\uc6e8\uc5b4 \uc774\ub2e4. GPU, CPU\uac00 \uac00\uc9c0\uace0 \uc788\ub294 \ubcd1\ub82c\ucc98\ub9ac \uad6c\uc870\ub97c \uba3c\uc800 \uc774\ud574\ud574\uc57c NPU\uc758 Neural Network\uc5d0 \ud2b9\ud654\ub418\uc5b4 \uc788\ub294 \uc5f0\uc0b0\uc744 \uc774\ud574 \ud560 \uc218 \uc788\ub2e4. \ub2e4\ub978 \ud3ec\uc2a4\ud305\uc758 GPU, CPU\uc758 \ubcd1\ub82c\ucc98\ub9ac \uae30\ubc95\uc744 \uc774\ud574\ud55c \ud6c4\uc5d0 \uc544\ub798 \ub0b4\uc6a9\uc744 \uc77d\uc73c\uba74 \uc870\uae08 \ub354 \ub300\ube44\ub418\ub294 \uc810\uc744 \uc54c \uc218 \uc788\ub2e4. (\uad00\ub828 \ub9c1\ud06c \ucd94\uac00 \uc608\uc815). \uac1c\ub150\uc801\uc73c\ub85c\ub294 GPU, CPU, NPU, TPU \ubaa8\ub450 \ub3d9\uc77c\ud55c \uc5f0\uc0b0\uc5d0 \ub300\ud574\uc11c \uc5ec\ub7ec\uac1c\uc758 \uc5f0\uc0b0\uc7a5\uce58\uac00 \ub3d9\uc77c\ud55c \ud615\ud0dc\uc758 \ub2e4\ub978 \uac12\uc744 \uac00\uc9c4 \ub9ce\uc740 \ub370\uc774\ud130\ub97c \ub3d9\uc2dc\uc5d0 \ucc98\ub9ac \ud55c\ub2e4\uace0 \uc0dd\uac01\ud558\uba74 \ub41c\ub2e4. NPU\/TPU\ub294 \uae30\ubcf8\uc801\uc778 \ubcd1\ub82c\ucc98\ub9ac\uc5d0 \ucd94\uac00\uc801\uc778 \uae30\ubc95\uc744 \ub354\ud55c \uac83\uc774\ub2e4. \uc774\ub54c \uc0ac\uc6a9\ub418\ub294 \uc8fc\uc694 \uae30\uc220\uc774 \uc2dc\uc2a4\ud1a8\ub9ad\uacfc \ud6a8\uc728\uc801\uc778 \uba54\ubaa8\ub9ac \uad00\ub9ac \uc785\ub2c8\ub2e4. \ud6a8\uc728\uc801\uc778 \uba54\ubaa8\ub9ac \uad00\ub9ac\ub294 AI \uc5f0\uc0b0\uc5d0 \ucd5c\uc801\ud654\ub41c \uba54\ubaa8\ub9ac \uad00\ub9ac\ub97c \uc758\ubbf8 \ud569\ub2c8\ub2e4. \ubc94\uc6a9\uc131\uc744 \ub2f4\ubcf4\ud558\uc9c0\ub294 \ubabb\ud569\ub2c8\ub2e4. \ub370\uc774\ud130 Quantazation\ub3c4 \ucd08\uae30\uc5d0\ub294 \ub9ce\uc774 \uc0ac\uc6a9\ub418\uc5c8\uc9c0\ub9cc, \ud604\uc7ac\ub294 ML \uc790\uccb4\uc5d0\uc11c \uc18c\ud504\ud2b8\uc6e8\uc5b4\uc801\uc73c\ub85c \ub9cc\uc774 \uc0ac\uc6a9\ub418\uc5b4\uc11c \ud604\uc7ac\ub294 NPU \ud6a8\uc6a9\uc131\uc740 \uc5c6\uc5b4 \ubcf4\uc778\ub2e4. \uadf8\ub9ac\uace0 \uc911\uc2ec\uc774 \ub418\ub294 \uae30\uc220\uc744 \ud65c\uc6a9\ud558\uace0 \uac01 AI \uce69\ub9c8\ub2e4 \uace0\uc720\uc758 \ud2b9\uc9d5\uc774 \ubc18\uc601\ub418\uc5b4 \uc788\ub2e4. \ud558\ub098\uc529 \uc54c\uc544 \ubcf4\uc790.<\/p>\n\n\n\n

\uc2dc\uc2a4\ud1a8\ub9ad(Systolic)<\/h4>\n\n\n\n

\uc2dc\uc2a4\ud1a8\ub9ad \uc544\ud0a4\ud14d\ucc98 \ub610\ub294 \uc2dc\uc2a4\ud1a8\ub9ad \uc5f0\uc0b0\uc774\ub77c\uace0 \ubd88\ub9ac\ub294 \uc2dc\uc2a4\ud1a8\ub9ad\uc740 \ud589\ub82c \uc5f0\uc0b0\uc5d0 \ucd5c\uc801\ud654 \ub418\uc5b4 \uc788\ub294 \ud558\ub4dc\uc6e8\uc5b4 \ub610\ub294 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc5f0\uc0b0 \uae30\ubc95\uc774\ub2e4. \ubcd1\ub82c \ucc98\ub9ac\uc5d0\uc11c \ub3d9\uc2dc \ucc98\ub9ac \ubcf4\ub2e4 \uc911\uc694\ud55c \ubd80\ubd84\uc774 \ub370\uc774\ud130\uac00 \uacc4\uc0b0\uc774 \ud544\uc694\ud55c \uacf3\uc5d0 \uac00\uc7a5 \uc54c\ub9de\uc740 \uc2dc\uae30\uc5d0 \ub3c4\ucc29\ud574 \uc788\ub294 \uac83\uc774\ub2e4. \uc774\ub97c \uc704\ud574\uc11c Pre-Patch\ub97c \ud560 \uc218\ub3c4 \uc788\uace0, Pipiline\uc744 \uc798 \uc124\uacc4\ud560 \uc218 \ub3c4 \uc788\uc9c0\ub9cc Host\uc5d0\uc11c Accelerator\ub85c \ub370\uc774\ud130 \ubcf5\uc0ac\uac00 \ube48\ubc88\ud560 \uacbd\uc6b0 \uc131\ub2a5 \uc800\ud558\uac00 \ubc1c\uc0dd\ud560 \uc218 \uc788\ub2e4. \uc2dc\uc2a4\ud1a8\ub9ad \uc544\ud0a4\ud14d\ucc98\ub294 \ub370\uc774\ud130 \uc5f0\uc0b0\uc2dc \ub370\uc774\ud130 \ubcf5\uc0ac\uac00 \ucd5c\uc18c\ud654 \ub418\ub3c4\ub85d \ud558\ub294 \ubc29\ubc95\uc774\ub2e4. \uc774\ub294 PE(Processing Element)\uac00 \ub3c5\ub9bd\ub41c NPU \uacc4\uc5f4\uc774 \ucc98\ub9ac\uac00 \uc720\ub9ac\ud558\uae30 \ub54c\ubb38\uc5d0 GPU\ubcf4\ub2e4 NPU \uacc4\uc5f4\uc5d0\uc11c \ub9ce\uc774 \uc0ac\uc6a9 \ub418\ub294 \ubc29\ubc95\uc774\uace0, \uc774\ub97c \ud1b5\ud574\uc11c Memory Locality\ub97c \ud5a5\uc0c1\ud560 \uc218 \uc788\ub2e4. \uad6c\uae00\uc758 TPU, Furiosa, GraphCore IPU\uc5d0\uc11c \ubaa8\ub450 \uc0ac\uc6a9\ud558\ub294 \ubc29\ubc95\ub2c8\ub2e4.<\/p>\n\n\n\n

\uc544\ub798 \uadf8\ub9bc\uc740 Matrix Multiply\ub97c \uadf8\ub9bc\uc73c\ub85c \ub098\ud0c0\ub0b8 \uadf8\ub9bc \uc774\ub2e4. \uadf8\ub9bc\uc744 \ud1b5\ud574\uc11c \uc2dc\uc2a4\ud1a8\ub9ad\uc5d0 \ub300\ud574\uc11c \uc54c\uc544 \ubcf4\uc790. A Matrix\uc640 B Matrix\ub97c \uacf1\ud574\uc11c C Marix\uc5d0 \uc800\uc7a5\ud55c\ub2e4\uace0 \ud560 \ub54c, A Matrix A00<\/sub>\uc774 \uc5f0\uc0b0\uc5d0 \uc0ac\uc6a9 \ub418\ub294 C Matrix \uc704\uce58\uc5d0 \ube68\uac04\uc0c9 \ub3d9\uadf8\ub77c\ubbf8\ub97c \ud45c\uc2dc \ud558\uc600\ub2e4. \uac19\uc740 \ubc29\ubc95\uc73c\ub85c B Matrix\uc758 B00<\/sub>\uc744 \ud30c\ub780\uc0c9 \ub3d9\uadf8\ub77c\ubbf8\ub85c \ud45c\uc2dc \ud558\uc600\ub2e4.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

C00<\/sub>\uc740 A00<\/sub>, A01<\/sub>, A02<\/sub>, A03<\/sub>\uac01\uac01\uacfc B00<\/sub>, B10<\/sub>, B20<\/sub>, B30<\/sub>\uac01\uac01\uc744 \uacf1\ud55c \uac12\uc744 \ub354\ud55c \uacb0\uacfc\ub2e4. C01<\/sub>\uc740 A00<\/sub>, A01<\/sub>, A02<\/sub>, A03<\/sub> \uac01\uac01\uacfc B10<\/sub>, B11<\/sub>, B12<\/sub>, B13<\/sub>\uac01\uac01\uc744 \uacf1\ud55c \uac12\uc744 \ub354\ud55c \uacb0\uacfc\ub2e4. \uc774\ub7f0 \uc2dd\uc73c\ub85c \uc0dd\uac01\ud574 \ubcf4\uba74 A00<\/sub>\ub294 C00<\/sub>, C01<\/sub>, C02<\/sub>, C03<\/sub> \uc5f0\uc0b0\uc5d0 \uc0ac\uc6a9\ub41c\ub2e4. B00<\/sub>\ub294 C00<\/sub>, C10<\/sub>, C20<\/sub>, C30<\/sub> \ub97c \uacc4\uc0b0\ud558\ub294 \uc5f0\uc0b0\uc5d0 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. A00<\/sub>, B00<\/sub>\ub4f1\uc744 \uacc4\uc0b0\uc2dc \ub9c8\ub2e4 Offloading\uc744 \ud558\uba74 \uc544\uc8fc \ub9ce\uc740 Cost\uac00 \ubc1c\uc0dd\ud560 \uac83\uc774\ub2e4. \uc2dc\uc2a4\ud1a8\ub9ad\uc740 \ud608\uad00\uc5d0 \ud608\uc561\uc774 \ud750\ub974\ub4ef\uc774 \uc774 A00<\/sub>, B00<\/sub>\ub97c PE\ub4e4\uc774 \uc5f0\uc0b0\ud560 \uc218 \uc788\ub3c4\ub85d \uccab\ubc88\uc9f8 \uc5f0\uc0b0\uc2dc\uc5d0\ub9cc Offloading\uc744 \ud558\uace0 PE\uac04 \ub370\uc774\ud130\ub97c \uc774\ub3d9\ud558\uba74\uc11c \uc0ac\uc6a9\ud558\ub294 \ubc29\ubc95\uc774\ub2e4. \ub2f9\uc5f0\ud788 PE\uc5d0 Local \uba54\ubaa8\ub9ac\uac00 \uc874\uc7ac \ud574\uc57c\ub9cc \ud55c\ub2e4. \ub9ce\uc740 NPU\ub4e4\uc774 Local Cache, Local Memoery\ub97c \uac00\uc9c0\uace0 \uc788\uae30 \ub54c\ubb38\uc5d0 Offloading\uc758 Overhead\ub97c \uadf9\ubcf5\ud560 \uc218 \uc788\ub294 \uac83\uc774\ub2e4. \uc544\ub798 \uadf8\ub9bc\uc774 \ubcf4\uc2dc\uba74 \uc774\ud574\uac00 \uc26c\uc6b8 \uac83\uc774\ub2e4. C Matrix\uc758 \uac01 \uc5d8\ub9ac\uba3c\ud2b8\uc758 \uc5f0\uc0b0\uc744 1\uac1c\uc758 PE\uac00 \ub2f4\ub2f9\ud55c\ub2e4\uace0 \uac00\uc815\ud55c\ub2e4.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\uc2dc\uc2a4\ud1a8\ub9ad \uc544\ud0a4\ud14d\ucc98\uc5d0\uc11c\ub294 \uc55e\uc11c \uc124\uba85\ud55c \uac83\ucc98\ub7fc A00<\/sub>, B00<\/sub>\uac00 \uac00\uc7a5 \uba3c\uc800 offloading \ub41c \ud6c4\uc5d0 \uacf1\ud574\uc9c4 \ud6c4\uc5d0 C00<\/sub>\uc5d0 \uc800\uc7a5\ub41c\ub2e4. A00<\/sub>, B00<\/sub>\ub294 \ub2e4\uc74c \uc5f0\uc0b0\uc744 \uc704\ud574 \uac01\uac01 C01<\/sub>, C10<\/sub>\uc73c\ub85c \uc774\ub3d9\ud55c\ub2e4. \uadf8\ub9ac\uace0 \ub2e4\uc74c \uc5f0\uc0b0\uc744 \uc704\ud574\uc11c \ub370\uc774\ud130\uac00 \ucd94\uac00\ub85c offloading \ub418\uace0, C00<\/sub> \uc758 \uc804\uccb4 \uc5f0\uc0b0\uc758 \ub2e4\uc74c \uacc4\uc0b0\uc744 \uc704\ud574\uc11c A01<\/sub>, B10<\/sub>\uc774 Offloading\ub418\uace0 C01<\/sub>\uc5d0\uc11c A00<\/sub>\uc640 \uacc4\uc0b0\ub420 B01<\/sub>\uc774 C10<\/sub>\uc5d0\uc11c B00<\/sub>\uc640 \uacc4\uc0b0\ub420 A10<\/sub> \uc774 Offloading \ub41c\ub2e4. \uc774\ub7f0 \uc5f0\uc0b0\uc744 \uc218\ud589\ud558\uba74 \uc5f0\uc0b0\uc758 \uc21c\uc11c\uac00 \uc67c\ucabd \uc0c1\ub2e8\uc5d0\uc11c \ubd80\ud130 \uc624\ub978\ucabd \ud558\ub2e8\uc73c\ub85c \uc5f0\uc0b0\uc774 \ud655\uc0b0\ub418\ub294 \uac83\uc744 \uc54c \uc218 \uc788\ub2e4. \uc0c8\ub86d\uac8c Offloading\ub41c \ub370\uc774\ud130\ub294 A Matrix \uac12\uc740 Row Index\ub97c \uc720\uc9c0\ud558\uba74\uc11c \uc624\ub978\ucabd\uc73c\ub85c B Matrix\uc73c \uac12\uc740 Column Index\ub97c \uc720\uc9c0 \ud558\uba74\uc11c \uc544\ub7ab\ucabd\uc73c\ub85c \uc774\ub3d9\ud558\uba74\uc11c \uc5f0\uc0b0\uc5d0 \uc0ac\uc6a9\ub41c\ub2e4.<\/p>\n\n\n\n

\ubc14\ub85c \uc704 \uc624\ub978\ucabd \uadf8\ub9bc\uc744 Cycle\uc744 X\ucd95\uc73c\ub85c \ub2e4\uc2dc \uadf8\ub9b0 \uadf8\ub9bc\uc744 GraphCore\uc758 IPC Whitepaper\uc5d0\uc11c\ub3c4 \uc18c\uac1c \ud558\uace0 \uc788\ub2e4.<\/p>\n\n\n

\n
\"\"
\ucd9c\ucc98: DELL DSS8440 GRAPHCORE IPU SERVER WhitePaper<\/figcaption><\/figure>\n<\/div>\n\n\n

\uadf8\ub9bc\uc5d0\uc11c \ubcf4\ub294 \uac83\ucc98\ub7fc \ud589\ub82c \uc5f0\uc0b0\uc5d0 \ucd5c\uc801\ud654 \ub418\uc5b4 \uc788\ub294 \ud558\ub4dc\uc6e8\uc5b4 \uad6c\uc870\uc774\ub2e4. \ud558\uc9c0\ub9cc NVIDIA GPU\ub294 \ucc98\uc74c\ubd80\ud130 \uc774 \uad6c\uc870\ub97c \uac00\uc9c0\uace0 \uc788\uc9c0 \uc54a\uc558\ub2e4. NVIDIA GPU\ub294 \ubcd1\ub82c\ucc98\ub9ac\uac00 \ud544\uc694\ud55c \ubc94\uc6a9 \ub300\uaddc\ubaa8 \uc5f0\uc0b0\uc5d0 \ucd5c\uc801\ud654 \ub418\uc5b4 \uc788\uae30 \ub54c\ubb38\uc774\ub2e4. <\/p>\n\n\n\n

\uba54\ubaa8\ub9ac \uacc4\uce35 \uad6c\uc870(Memory Hierarchy)<\/h4>\n\n\n\n

\uc2dc\uc2a4\ud1a8\ub9ad\uc774 \uc81c\ub300\ub85c \ub3d9\uc791\ud558\uae30 \uc704\ud574\uc11c \ud544\uc694\ud55c \uc120\ud589 \uc870\uac74\uc774 \uc788\ub2e4. \uc2dc\uc2a4\ud1a8\ub9ad\uc740 PE\uac00 \ud589\ub82c\uc758 \ud55c \uc694\uc18c \ucc98\ub7fc \ub3d9\uc791\ud558\uace0 \uacc4\uc0b0\uc744 \uc704\ud55c \ub370\uc774\ud130\uac00 \ud758\ub7ec \ub4e4\uc5b4 \uc624\ub294 \uad6c\uc870\uc774\ub2e4. \uc774\ub97c \uc704\ud574\uc11c\ub294 PE\uc5d0 \ub370\uc774\ud130\uac00 \uc801\uc808\ud558\uac8c \uacf5\uae09\uc774 \ub418\uc5b4\uc57c \ud55c\ub2e4. \uc801\uc808\ud558\uac8c\ub294 \uacf5\uae09\uc774 \ub41c\ub2e4\ub294 \uac83\uc740 \uacc4\uc0b0\uc5d0 \ud544\uc694\ud55c \uc2dc\uc810\uc5d0 \ub370\uc774\ud130\uac00 \uc900\ube44 \ub418\uc5b4 \uc788\uc5b4\uc57c \ud55c\ub2e4\ub294 \uac83\uc774\ub2e4. NPU, TPU \ub4f1\uc740 \ud3f0 \ub178\uc774\ub9cc \uc544\ud0a4\ud14d\uccd0\uc5d0\uc11c \ubcf4\uc870\uc5f0\uc0b0\uc7a5\uce58 \uc774\uae30 \ub54c\ubb38\uc5d0 \uba54\uc778 \uba54\ubaa8\ub9ac\uc5d0\uc11c \ub370\uc774\ud130\uac00 NPU\ub85c \uacf5\uae09\ub418\ub294\ub370 \uc0c1\ub2f9\ud788 \ub9ce\uc740 \uc2dc\uac04\uc774 \ud544\uc694\ud569\ub2c8\ub2e4. \ub2e8 \ud55c\ubc88\uc758 \uba54\ubaa8\ub9ac \uc804\ub2ec(\ub610\ub294 \ubcf5\uc0ac)\ub294 \ub208\uae5c\uc9dd\ud560 \uc2dc\uac04 \ubcf4\ub2e4 \uc9e7\uc9c0\ub9cc \ub370\uc774\ud130\uac00 \uc218\ubc31\ub9cc, \uc218\ubc31\uc5b5\ubc88 \ubcf5\uc0ac\uac00 \ubc18\ubcf5\uc774 \ub420 \ub54c \ubc1c\uc0dd\ud558\ub294 \uc131\ub2a5 \uc800\ud558\ub294 \ub208\uae5c\uc9dd\ud560 \uc2dc\uac04\uc774 \uc544\ub2cc \uc5c4\uccad\ub09c \uc2dc\uac04\uc744 \uba54\ubaa8\ub9ac \uc804\ub2ec\uc744 \uc704\ud574\uc11c \uc0ac\uc6a9\ud558\uac8c \ub41c\ub2e4. \uadf8\ub798\uc11c \uc774\ub7f0 \ubc18\ubcf5\ub41c \uba54\ubaa8\ub9ac \uc804\ub2ec\uc5d0\uc11c \uc624\ub294 \uc131\ub2a5 \uc800\ud558\ub97c \ucd5c\uc18c\ud654 \ud558\uae30 \uc704\ud574\uc11c \uc5ec\ub7ec\uac00\uc9c0 \uae30\ubc95\ub4e4\uc774 \uc608\uc804 \ubd80\ud130 \uc0ac\uc6a9\ub418\uace0 \uc788\uc5c8\ub2e4. \ud2b9\ud788 \uacc4\uc0b0\uc744 \uc704\ud55c \ub370\uc774\ud130\ub97c \uc0ac\uc804\uc5d0 \ubcf5\uc0ac\ud558\ub294 Pipelining, Offloading \ucd5c\uc801\ud654 \ub4f1\uc740 \uc798 \uc54c\ub824\uc9c4 \uae30\ubc95\uc774\ub2e4. \uc774\ub7f0 \uae30\ubcf8\uc801\uc778 \uae30\ubc95\uc740 \uac70\uc758 \ube60\uc9d0 \uc5c6\uc774 \uc801\uc6a9\ub418\uc5b4 \uc788\uc9c0\ub9cc, LLM \uc2dc\ub300, DL \uc2dc\ub300\uc5d0\ub294 \uc5ec\uc804\ud788 \uc131\ub2a5 \uc800\ud558\uac00 \uaf64 \uc874\uc7ac \ud55c\ub2e4.<\/p>\n\n\n

\n
\"\"
PE\uc640 Main memory\uac04 \ub370\uc774\ud130 \uad50\ud658 \ud1b5\ub85c PCIe (\ucd9c\ucc98: Sora\uc5d0\uc11c \uc0dd\uc131)<\/figcaption><\/figure>\n<\/div>\n\n\n

\ubc14\ub85c \uc774 \uc131\ub2a5 \uc800\ud558(Overhead)\ub97c \uc81c\uac70 \ud558\uae30 \uc704\ud574\uc11c NPU \uc124\uacc4\uc2dc \ub2e4\uce35 \uad6c\uc870\uc758 \uba54\ubaa8\ub9ac\ub97c \uac00\uc9c4\ub2e4. \uba54\ubaa8\ub9ac\uc758 \ub2e4\uce35 \uad6c\uc870\ub294 \uc77c\ubc18\uc801\uc73c\ub85c \uc0ac\uc6a9\ub418\ub294 \ucef4\ud4e8\ud130 \uad6c\uc870\uc5d0\uc11c \ucc3e\uc544 \ubcfc \uc218 \uc788\ub2e4. \ud754\ud788 \ub9d0\ud558\ub294 L1, L2, L3 Cache\uac00 \uba54\uc778 \uba54\ubaa8\ub9ac\uc5d0\uc11c CPU\ub85c \ub370\uc774\ud130\uac00 \uc804\ub2ec\ub418\ub294 \uc2dc\uac04\uc744 \ub2e8\ucd95\ud558\uae30 \uc704\ud574 Cache\ub97c \uccb4\uacc4\uc801\uc73c\ub85c \uac00\uc9c0\uace0 \uc788\uace0 \uc774\ub97c \ud65c\uc6a9\ud55c\ub2e4. \ubcf4\uc870\uc5f0\uc0b0\uc7a5\uce58\ub3c4 \uc720\uc0ac\ud55c \uad6c\uc870\ub97c \uac00\uc9c0\uace0 \uc788\ub2e4. \ud558\uc9c0\ub9cc PE\uac00 \ub9ce\uc740 \uad6c\uc870\uc5d0\uc11c\ub294 L1 \ub808\ubca8\uc758 Cache\ub97c \ubaa8\ub450 \uac00\uc9c0\ub294 \uac83\uc740 \ube44\uc6a9\uc801\uc778 \ubb38\uc81c\uac00 \uc788\uae30 \ub54c\ubb38\uc5d0 \ucd08\uae30\uc5d0\ub294 \uc798 \uc801\uc6a9\ub418\uc9c0 \uc54a\uc558\ub2e4. \uadf8 \ub300\uc2e0 L3, L2 \ub808\ubca8\uc758 Cache\uc758 \ud65c\uc6a9\uc774 \ud65c\ubc1c \ud588\uc2b5\ub2c8\ub2e4. \ucd08\uae30 \ubcf4\uc870\uc5f0\uc0b0\uc7a5\uce58\uc5d0\uc11c\ub294 \ud070 \ubb38\uc81c\uac00 \uc5c6\uc5c8\uc73c\ub098 Neural Architecture\uac00 \ucee4\uc9c0\uace0 \ud150\uc11c\uc5f0\uc0b0\uc774 \uae30\ud558\uae09\uc218\uc801\uc73c\ub85c \ub192\uc544 \uc9c0\uba74\uc11c L1\uc744 \ub3c4\uc785\ud558\uace0 \uc55e\uc5d0\uc11c \uc124\uba85\ud55c \uc2dc\uc2a4\ud1a8\ub9ad \uc544\ud0a4\ud14d\uccd0\uc758 \uc7a5\uc810\uc744 \uadf9\ub300\ud654 \ud558\uac8c \ub41c\ub2e4. GraphCore\uc758 IPU \ubc31\uc11c\ub97c \ubcf4\uba74 \ud2b9\uc9d5\uc774 \uac00\uc7a5 \ub3c4\ub4dc\ub77c \uc9c4\ub2e4.<\/p>\n\n\n

\n
\"\"
GraphCore IPU \uad6c\uc870 (\ucd9c\ucc98:  The GraphCore Software Stack: Built to Scale) <\/figcaption><\/figure>\n<\/div>\n\n\n

IPU Core(PE)\ub9c8\ub2e4 \ubc14\ub85c \uadfc\uc811\ud55c \uc704\uce58\uc5d0 \uba54\ubaa8\ub9ac\uac00 \uc704\uce58\ud574 \uc788\ub2e4. \uc2dc\uc2a4\ud1a8\ub9ad \uc544\ud06c\ud14d\ucc98\uc5d0\uc11c \uc804\ub2ec\ub41c \ub370\uc774\ud130\ub97c \uc800\uc7a5\ud558\ub294 \ube60\ub978 \uba54\ubaa8\ub9ac\ub85c \uc0ac\uc6a9\uc774 \ub41c\ub2e4. \uad6d\ub0b4\uc0ac\uc778 \ub9ac\ubca8\ub9ac\uc628\uc0ac\uc758 \ud558\ub4dc\uc6e8\uc5b4 \uad6c\uc870\ub97c \ubcf4\uba74 \uc870\uae08 \ucd94\uc0c1\uc801\uc774\uc9c0\ub9cc \uc774 \uba54\ubaa8\ub9ac \uad6c\uc870\ub97c \ud655\uc778\ud560 \uc218 \uc788\ub2e4. \uc544\ub798 \uadf8\ub9bc\uc740 \ub9ac\ubca8\ub9ac\uc628 \ud648\ud398\uc774\uc9c0\uc5d0\uc11c \ub2e4\uc6b4\ub85c\ub4dc \ubc1b\uc740 \ubc31\uc11c\uc5d0\uc11c \ucea1\ucc98 \ud558\uc600\ub2e4.<\/p>\n\n\n\n

\n
\n
\"\"

<\/em>\ub9ac\ubca8\ub9ac\uc628 NPU \uac1c\ub150\ub3c4(\ucd9c\ucc98: ATOM\u2122 Architecture \ubc31\uc11c)<\/figcaption><\/figure>\n<\/div>\n\n\n\n
\n
\"\"
\ub9ac\ubca8\ub9ac\uc628 \ub274\ub7f4\uc5d4\uc9c4 \uac1c\ub150\ub3c4 (\ucd9c\ucc98: ATOM\u2122 Architecture \ubc31\uc11c)<\/figcaption><\/figure>\n<\/div>\n<\/div>\n\n\n\n

\uc67c\ucabd \uadf8\ub9bc\uc744 \ubcf4\uba74 ATOM \ubcf4\uc870\uc5f0\uc0c1\uc7a5\uce58(\ubcf4\uc870\ub77c\uace0 \ud558\ub2c8 \ubd80\uc815\uc801\uc778 \uc5b5\uc591\uc774 \uc788\uc5b4 Co-processor\ub85c \uc774\ud6c4 \uc9c0\uce6d\ud558\ub3c4\ub85d \ud558\uaca0\uc74c)\uc758 \uad6c\uc870\uc774\ub2e4. Co-processor\uc640 Host\uac00 \uc5f0\uacb0\ub418\ub294 \uad6c\uc870\uc774\uace0 Nerual \uc5f0\uc0b0\uc744 \ubc30\uce58\ud558\ub294 \ub4f1\uc758 \uc5ed\ud560\uc744 \ud558\ub294 \uc720\ub2db\ub4e4\uacfc L2 Cache\uac00 \uc788\ub2e4. Co-Processor\ub0b4\uc5d0 \uc5ec\ub7ec\uac1c\uc758 Nerual Engine Cluster\uac00 \uba87\uac1c \uc874\uc7ac \ud558\uace0, Nuerual Engine Cluster\ub0b4\ubd80\uc5d0\ub294 L1 Cache\uc640 \uc5ec\ub7ec\uac1c\uc758 Neural Engine\uc774 \uc874\uc7ac \ud55c\ub2e4. <\/p>\n\n\n\n

\uc624\ub978\ucabd \uadf8\ub9bc\uc774 \uc5ec\ub7ec\uac1c\uc758 Neural Engine \ub0b4\ubd80\uc774\ub2e4. \uc5d4\uc9c4 \uc790\uccb4\uc5d0 L0 Cache\uac00 \uc874\uc7ac \ud558\uace0 Tensor \uc5f0\uc0b0 \uc720\ub2db\uc774 \uc874\uc7ac \ud55c\ub2e4. L0 Cache\ub294 \uc2dc\uc2a4\ud1a8\ub9ad \uc544\ud0a4\ud14d\ucc98\uc5d0 \ub3c4\uc6c0\uc744 \uc900\ub2e4. \uaf2d \ud150\uc11c \uc5f0\uc0b0\ubfd0 \uc544\ub2c8\ub77c\ub3c4 Co-Proceesor\uc758 L2 Memory\uc758 \uac12\ube44\uc2fc \uc811\uadfc\uc744 L1, L0 Cache\ub97c \ud1b5\ud574\uc11c \uc131\ub2a5 \ubd80\ud558\ub97c \uac10\uc18c \uc2dc\ud0a8\ub2e4. <\/p>\n\n\n\n

\uba54\ubaa8\ub9ac \uacc4\uce35 \uad6c\uc870\uc5d0\ub3c4 \ub2e8\uc810\uc740 \uc788\ub2e4. Cache miss\uac00 \ube48\ubc88\ud788 \uc77c\uc5b4\ub098\ub294 \uc54c\uace0\ub9ac\uc998\uc758 \uacbd\uc6b0 Cache\uac00 \uc5c6\ub294 \uacbd\uc6b0\ubcf4\ub2e4 \ub354 \ub192\uc740 \uc131\ub2a5 \ubd80\ud558\ub97c \uac00\uc9c8 \uc218 \uc788\ub2e4. \uadf8\ub798\uc11c Cache miss\uac00 \ucd5c\uc18c\ud654\uac00 \ub418\ub3c4\ub85d \ucf54\ub4dc\uac00 \ucd5c\uc801\ud654 \ub418\uc5b4 \uc788\uc5b4\uc57c \ud55c\ub2e4. \uadf8\ub9ac\uace0 \uc678\ubd80 \uba54\ubaa8\ub9ac\ub97c \uac00\uc838 \uc62c\ub54c \ud070 \uc131\ub2a5 \ubd80\ud558\uac00 \uc788\ub2e4. \uc774\ub54c \uc774\ub97c \uc904\uc5ec \uc904 \uc218 \uc788\ub294 \uad6c\uc870\uac00 DMA(Direct Memory Access)\uc774\ub2e4. DMA\ub294 \uc6d0\uaca9\uc758 \uba54\ubaa8\ub9ac\ub97c \uba54\ubaa8\ub9ac\ub97c \uad00\ub9ac\ud558\ub294 Processor \ub610\ub294 \uba54\ubaa8\ub9ac \uad00\ub9ac \ubaa8\ub4c8\uc744 \uac70\uce58\uc9c0 \uc54a\uace0 \ub9c8\uce58 \ud558\ub098\uc758 \uac00\uc0c1\uc758 \uba54\ubaa8\ub9ac \ucc98\ub7fc \uc811\uadfc\ud560 \uc218 \uc788\ub2e4. \uc774\ub97c \uc704\ud574\uc11c NoC \ub4f1\uc774 \ud558\ub4dc\uc6e8\uc5b4\uc801\uc73c\ub85c \uc9c1\uc811 \uc5f0\uacb0\ub418\uc5b4 \uc788\ub2e4. <\/p>\n\n\n\n

\uadf8\uc678 \uc131\ub2a5 \ucd5c\uc801\ud654 \ubc29\ubc95<\/h4>\n\n\n\n

\ub9ce\uc740 NPU\uac00 \ucd94\uac00\uc801\uc73c\ub85c \uac00\uc9c0\uace0 \uc788\ub294 \uad6c\uc131\uc694\uc18c\uac00 DL, LLM\uc5d0 \ud2b9\ud654\ub41c \uc5f0\uc0b0\uc744 \ud558\ub4dc\uc6e8\uc5b4\uc801\uc73c\ub85c \uad6c\ud604\ud55c \uacbd\uc6b0\ub3c4 \uc788\uc9c0\ub9cc, \ud1b5\uc0c1 \ud589\ub82c\uacf1\uc744 \ud558\ub4dc\uc6e8\uc5b4\ub85c \uc9c1\uc811 \uad6c\ud604\ud55c \uacbd\uc6b0\uac00 \uac00\uc7a5 \uc77c\ubc29\uc801\uc774\ub2e4. \ud558\uc9c0\ub9cc NVIDIA\uc758 Volta Architecture \ubd80\ud130 Tensor core\uc758 \uc9c0\uc6d0\uc73c\ub85c NPU\uc758 \ud2b9\ud654\ub41c \ud558\ub4dc\uc6e8\uc5b4 \uc5f0\uc0b0\uc774\ub77c\uace0 \ud560 \uc218\ub294 \uc5c6\ub2e4. \uadf8\ub9ac\uace0 DL, LLM\uc5d0 \ud2b9\ud654\ub41c \uc5f0\uc0b0\uc758 \uacbd\uc6b0 \ud2b9\uc815 Neural Architecture\uc5d0\ub9cc \ud2b9\ud654\ub420 \uc218 \uc788\uae30 \ub54c\ubb38\uc5d0 \uc7a5\uc810\uc774\ub77c\uace0\ub9cc \ud558\uae30 \uc5b4\ub835\ub2e4. \ub300\ud45c\uc801\uc778 \uacbd\uc6b0\uac00 \uad6d\ub0b4\uc758 Furiosa\uc758 NPU\uc774\ub2e4. 1\uc138\ub300\ub294 CNN \uc5f0\uc0b0\uc744 \uc911\uc2ec\uc73c\ub85c 2\uc138\ub300\ub294 LLM\uc744 \uc911\uc2ec\uc73c\ub85c \uc5f0\uc0b0 \ucd5c\uc801\ud654\ub97c \ud558\ub4dc\uc6e8\uc5b4\uc801\uc73c\ub85c \uc9c4\ud589\ud558\uc600\ub2e4.<\/p>\n\n\n\n

\uc804\ubd80\ub294 \uc544\ub2c8\uc9c0\ub9cc \uc77c\ubd80 NPU\ub294 \ud765\ubbf8\ub85c\uc6b4 \ud2b9\uc131\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. GPU\ub97c \ube44\ub86f\ud55c NPU\ub4e4\uc740 \ub3d9\uc77c\ud55c \ud615\ud0dc\uc758 \ub300\uaddc\ubaa8 \ub370\uc774\ud130\ub97c \ub3d9\uc77c\ud55c \uc5f0\uc0b0\uc73c\ub85c \ucc98\ub9ac\ud558\ub294 SIMD(Single Instruction Multiple Data)\uc744 \uadf9\ub300\ud654\ud558\ub294 \uad6c\uc870\uc785\ub2c8\ub2e4. SIMD\ub294 \ub300\uaddc\ubaa8 \ub370\uc774\ud130\ub97c \uc5f0\uc0b0\ub9cc \ubc18\ubcf5\ud560 \ub54c \ucd5c\uc120\uc758 \uacb0\uacfc\uac00 \ub098\uc624\ub294\ub370 \ub9cc\uc57d \ubd84\uae30\ubb38\uc774 \ud3ec\ud568\ub418\uc5b4 \uc788\uac70\ub098 \ub2e4\uc591\ud55c \uacbd\uc6b0\ub97c \uace0\ub824\ud574\uc57c \ud558\ub294 \uacbd\uc6b0\ub77c\uba74 \uc624\ud788\ub824 \uc131\ub2a5\uc774 \uc800\ud558\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7f0 \ub2e8\uc810\uc744 \uadf9\ubcf5\ud558\uae30 \uc704\ud574\uc11c GraphCore IPU, \ub9ac\ubca8\ub9ac\uc628\uc758 NPU\ub294 MIMD(Multiple Instructions Multiple Data) \ucf54\uc5b4\ub97c \ucd94\uac00\uc801\uc73c\ub85c \uac00\uc9c0\uace0 \uc788\uac70\ub098 \ucf54\uc5b4 \uc790\uccb4\uac00 SIMD, MIMD\ub97c \ubaa8\ub450 \uc9c0\uc6d0\ud558\ub294 \ud615\ud0dc\ub97c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc591\ud55c \ud615\ud0dc\uc758 \uacc4\uc0b0 \ubc29\uc2dd\uacfc \ub85c\uc9c1\uc744 \ucc98\ub9ac\ud558\ub354\ub77c\ub3c4 \ucd5c\ub300\uc758 \uc131\ub2a5\uc744 \ubc1d\ud718\ud558\uae30 \uc704\ud568\uc774\uc8e0.<\/p>\n\n\n\n

NPU\ub97c \uc704\ud55c \uc0dd\ud0dc\uacc4 \ub3d9\ud654, \ucc3d\uc870 \uc804\ub7b5<\/h3>\n\n\n\n

NPU \uc81c\uc870\uc0ac\ub4e4\uc758 \uc8fc\uc7a5\uacfc NPU\uc758 \uc5f0\uc0b0 \ub2a5\ub825\uc774\ub098 DL, LLM\uc5d0 \ucd5c\uc801\ud654\ub41c \uc124\uacc4\ub97c \ubcf4\uba74 GPU\uac00 \ubc14\ub85c NPU\ub85c \ub300\uccb4 \ub420 \uac83\ucc98\ub7fc \ubcf4\uc774\uc9c0\ub9cc \uc774 \uae00\uc758 \ucc98\uc74c\uc5d0\uc11c \ubd24\ub358 \uac83\ucc98\ub7fc \uacf5\uace0\ud788 \ub9cc\ub4e4\uc5b4\uc838 \uc788\ub294 GPU\uc758 \uc0dd\ud0dc\uacc4\uc5d0 \ub3d9\ud654 \ub418\uae30 \uc5b4\ub835\uace0 GPU\uc640 \uc720\uc0ac\ud55c \uc0dd\ud0dc\uacc4\ub97c \ub9cc\ub4e4\uae30 \uc5b4\ub835\ub2e4. \uac00\uc7a5 \uadfc\ubcf8\uc801\uc778 \uc774\uc720\ub294 \uac01 NPU\ub97c \uc6b4\uc6a9\ud558\uae30 \uc704\ud55c \ub4dc\ub77c\uc774\ubc84, \uc18c\ud504\ud2b8\uc6e8\uc5b4, \uc120\ud615 \ub77c\uc774\ube0c\ub7ec\ub9ac \ub4f1\uc740 \uac01 \ud558\ub4dc\uc6e8\uc5b4\uc5d0 \uc801\ud569\ud558\uac8c \ub9cc\ub4e4\uc5b4\uc838\uc57c \ud558\uae30 \ub54c\ubb38\uc774\ub2e4. \uac01 Co-Processor\uc758 \ub3d9\uc791 \ud750\ub984\uc774\ub098 \uad6c\uc131\uc774 \ub2e4\ub974\uae30 \ub54c\ubb38\uc5d0 \ub2f9\uc5f0\ud55c \uc77c\uc774\ub2e4. <\/p>\n\n\n\n

\ud558\ub4dc\uc6e8\uc5b4\uc758 \ud2b9\uc131\uc744 \uac00\uc7a5 \uc798 \ud65c\uc6a9\ud558\uae30 \uc704\ud574 \uc9c1\uc811\uc801\uc774\uace0 \uac00\uc7a5 \ud6a8\uacfc\uc801\uc778 \ubc29\ubc95\uc740 \uac01 Co-Processor\uc758 \ub4dc\ub77c\uc774\ubc84 \ubc0f SDK\ub97c \uc9c1\uc811 \ub9cc\ub4e4\uace0 \ud65c\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ud558\ub294 \uac83\uc774\ub2e4. SDK\ub294 \uc790\uccb4\uc801\uc73c\ub85c \ub9cc\ub4e4\ub54c 2\uac00\uc9c0 \uc811\uadfc\uc744 \ud560 \uc218 \uc788\ub2e4. \ud30c\uc774\uc36c\uc5d0\uc11c \uc0ac\uc6a9\uc774 \uac00\ub2a5\ud55c \ud328\ud0a4\uc9c0\ub97c \uc81c\uacf5\ud558\ub294 \ubc29\ubc95\uacfc \ud30c\uc774\uc36c\uc758 \ud655\uc7a5 \ud615\uc2dd\uc5d0 \ub9de\uac8c \uc778\ud130\ud398\uc774\uc2a4\ub97c \uad6c\ud604\ud558\uc5ec \ud655\uc7a5\ud558\ub294 \ubc29\ubc95\uc774\ub2e4. \ub450\uac00\uc9c0 \ubc29\ubc95 \ubaa8\ub450 \uc7a5\ub2e8\uc774 \uc788\ub2e4. \ubcc4\ub3c4\uc758 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub85c \uc81c\uacf5\ub420 \uacbd\uc6b0 C++ \ub610\ub294 Rust\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc131\ub2a5\uc5d0 \ucd5c\uc801\ud654\ub41c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc81c\uacf5\ud560 \uc218 \uc788\ub294 \uc7a5\uc810\uc774 \uc788\ub294 \ubc18\uba74, \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\uc744 \uc790\uccb4\uc801\uc73c\ub85c \ub9cc\ub4e4\uc5b4\uc57c \ud558\ub294 \ubd80\ub2f4\uc774 \uc788\ub2e4. \ub610\ud55c \ubcc4\ub3c4\ub85c \ub9cc\ub4e4\uc5b4\uc9c4 \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\uc758 \uacbd\uc6b0 \uc0c8\ub85c\uc6b4 \uc0ac\uc6a9\uc790\uc758 \uc720\uc785\uc5d0 \uc0c1\ub2f9\ud55c \uc2ec\ub9ac\uc801\uc778 \uc7a5\uc560\uac00 \ub420 \uc218 \uc788\ub2e4\ub294 \uc810\uc774\ub2e4. \ud30c\uc774\uc36c \ub610\ub294 \ud150\uc11c\ud50c\ub85c\uc6b0\uc758 \ud655\uc7a5 \ud615\uc2dd\uc744 \uc0ac\uc6a9\ud560 \uacbd\uc6b0 \uc774\uc804 \uc0ac\uc6a9\uc790\ub4e4\uc774 \ud558\ub4dc\uc6e8\uc5b4\ub97c \ubcc0\uacbd\ud558\ub294\ub370\uc11c \uc624\ub294 \uc2ec\ub9ac\uc801\uc778 \uc7a5\ubcbd\uc740 \uc81c\uac70 \ud560 \uc218 \uc788\uc9c0\ub9cc \ud574\ub2f9 ML Framework\uc5d0\uc11c \uc9c0\uc6d0\ud558\ub294 \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\uc744 \uadf8\ub300\ub85c \ub530\ub77c\uc57c \ud558\uae30 \ub54c\ubb38\uc5d0 \uacbd\uc6b0\uc5d0 \ub530\ub77c\uc11c \ucd5c\uc801\uc758 \uc131\ub2a5\uc744 \ud655\ubcf4\ud560 \uc218 \uc5c6\ub294 \ub2e8\uc810\uc774 \uc788\ub2e4. \uc774 \ub450\uac00\uc9c0\ub97c \ud45c\ub85c \uc815\ub9ac\ud574 \ubcf4\uba74 \ub2e4\uc74c\uacfc \uac19\ub2e4.<\/p>\n\n\n\n

\uc811\uadfc \ubc29\uc2dd<\/td>\uc7a5\uc810<\/td>\ub2e8\uc810<\/td><\/tr>
\ubcc4\ub3c4 \ub77c\uc774\ube0c\ub7ec\ub9ac (C++\/Rust \ud328\ud0a4\uc9c0)<\/strong><\/td>C++ \ub610\ub294 Rust\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc131\ub2a5\uc5d0 \ucd5c\uc801\ud654\ub41c \ub77c\uc774\ube0c\ub7ec\ub9ac \uc81c\uacf5 \uac00\ub2a5
\ud558\ub4dc\uc6e8\uc5b4\uc758 \ud2b9\uc131\uc744 \uac00\uc7a5 \uc798 \ud65c\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \uc9c1\uc811\uc801\uc778 \uc81c\uc5b4 \ubc0f \ucd5c\uc801\ud654 \uac00\ub2a5<\/td>		\ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\uc744 \uc790\uccb4\uc801\uc73c\ub85c \ub9cc\ub4e4\uc5b4\uc57c \ud558\ub294 \ubd80\ub2f4
\uc0c8\ub85c\uc6b4 \uc0ac\uc6a9\uc790\uc758 \uc720\uc785\uc5d0 \uc0c1\ub2f9\ud55c \uc2ec\ub9ac\uc801\uc778 \uc7a5\uc560 (\uc0c8\ub85c\uc6b4 \ud559\uc2b5 \uace1\uc120)<\/td><\/tr>
\ud30c\uc774\uc36c\/\ud150\uc11c\ud50c\ub85c\uc6b0 \ud655\uc7a5 \ud615\uc2dd<\/strong><\/td>\uae30\uc874 \uc0ac\uc6a9\uc790\ub4e4\uc758 \ud558\ub4dc\uc6e8\uc5b4 \ubcc0\uacbd\uc5d0 \ub300\ud55c \uc2ec\ub9ac\uc801\uc778 \uc7a5\ubcbd \uc81c\uac70 (\uce5c\uc219\ud55c \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378 \uc720\uc9c0)
\uae30\uc874 ML \ud504\ub808\uc784\uc6cc\ud06c\uc758 \ud48d\ubd80\ud55c \uae30\ub2a5 \ubc0f \uc0dd\ud0dc\uacc4 \ud65c\uc6a9 \uac00\ub2a5<\/td>		\ud574\ub2f9 ML \ud504\ub808\uc784\uc6cc\ud06c\uc5d0\uc11c \uc9c0\uc6d0\ud558\ub294 \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\uc744 \uadf8\ub300\ub85c \ub530\ub77c\uc57c \ud568
\uacbd\uc6b0\uc5d0 \ub530\ub77c \ud558\ub4dc\uc6e8\uc5b4\uc758 \ucd5c\uc801 \uc131\ub2a5\uc744 \ud655\ubcf4\ud558\uae30 \uc5b4\ub824\uc6b8 \uc218 \uc788\uc74c<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

\uc7a0\uc2dc \ud30c\uc774\uc36c\uc758 \ud558\ub4dc\uc6e8\uc5b4 \ud655\uc7a5 \ud615\uc2dd\uc5d0 \ub300\ud574\uc11c \uc0b4\ud3b4 \ubcf4\ub3c4\ub85d \ud558\uc790, \uc704 \uc774\uc57c\uae30\uc5d0 \ub300\ud55c \uc774\ud574\ubfd0 \uc544\ub2c8\ub77c NPU\uac00 \uc774\uc804 \uc0dd\ud0dc\uacc4\uc5d0 \uc798 \ub3d9\ud654\ub418\uae30 \uc704\ud55c \uc804\ub825 \uc5ed\uc2dc \ube44\uc2b7\ud55c \ubc29\uc2dd\uc774\uae30 \ub54c\ubb38\uc774\ub2e4. \ud30c\uc774\uc36c\uc740 GPU, NPU\ub4f1\uc758 \ud558\ub4dc\uc6e8\uc5b4 \ub9ac\uc18c\ub974\ub97c ProcessGroup\uc73c\ub85c \ubb36\uc5b4\uc11c allreduce, allgather\ub4f1\uc744 \ucd94\uc0c1\ud654 \ud558\uc5ec \uc0ac\uc6a9\ud55c\ub2e4. \uc55e\uc11c \uc2dc\uc2a4\ud1a8\ub9ad\uc744 \uc124\uba85\ud560 \ub54c \ub098\uc654\ub358 Processing Element(PE)\ub97c \uae30\ubcf8 \uac1c\ub150\uc73c\ub85c \ub123\uace0 \ub2e4\uc591\ud55c MPI \uae30\ubc18 \uc5f0\uc0b0\uc744 \ucd94\uc0c1\ud654 \ud55c \uac83\uc785\ub2c8\ub2e4. \ud30c\uc774\uc36c\uc758 \uacf5\uc2dd \ub808\ud30c\uc9c0\ud1a0\ub9ac\uc5d0 \uc62c\ub77c\uc628 \ud3ec\ud2b8(Port, Plugin)\uc740 NCCL, MPI, Gloo(CPU), UCX, oneAPI(Intel \ubd84\uc0b0\ucc98\ub9ac \ubc0f \uc5f0\uc0b0 \ud1b5\ud569 API)\uc774\ub2e4. \uad00\uacc4\ub3c4\ub97c \uadf8\ub9bc\uc73c\ub85c \ud45c\ud604\ud558\uba74 \uc544\ub798 \uadf8\ub9bc\uacfc \uac19\ub2e4.<\/p>\n\n\n

\n
\"\"
\ud30c\uc774\uc36c Procees Group \uacc4\uce35\ub3c4 (\ucd9c\ucc98: \uc790\uccb4)<\/figcaption><\/figure>\n<\/div>\n\n\n

\ud30c\uc774\uc36c, \ud150\uc11c\ud50c\ub85c\uc6b0 \ud655\uc7a5 \ud615\uc2dd\uc744 \ub530\ub97c \uacbd\uc6b0 \uc6d0\uce59\uc801\uc73c\ub85c\ub294 \uc2e0\uaddc \ud558\ub4dc\uc6e8\uc5b4\uc5d0\uc11c \ud559\uc2b5 \ubc0f \ucd94\ub860\uc5d0 \ubaa8\ub450 \uc0ac\uc6a9\ub420 \uc218 \uc788\ub2e4. \uc545\ub9c8\ub294 \ub514\ud14c\uc77c\uc5d0 \uc788\ub2e4\uace0 \ud55c\ub2e4. \uc2e4\uc81c\ub85c \ucd94\uc0c1\ud654\ub41c API\ub97c \uad6c\ud604\ud558\uace0 \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\uc744 \ub530\ub974\ub354\ub77c\ub3c4 \uc815\uc0c1\ub3d9\uc791\ud558\uc9c0 \uc54a\ub294 \uacbd\uc6b0\ub3c4 \ub9ce\uc744 \uac83\uc774\uba70, \ud2b9\ud788\ub098 \ud30c\uc774\uc36c\uc758 \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\ub85c \ub3d9\uc791 \ud588\uc744 \uacbd\uc6b0 \uae30\ub300\ud558\ub294 \uc131\ub2a5\uc5d0 \ud55c \ucc38 \ubabb \ubbf8\uce58\uc77c \uc218 \uc788\ub2e4. \uc774 \ubd80\ubd84\uc740 \uad6c\ud604 \ud6c4 \ucd5c\uc801\ud654\uc5d0 \uc5c4\uccad\ub09c \uacf5\uc744 \ub4e4\uc5ec\uc57c \ud558\ub294 \ubd80\ubd84\uc774\ub2e4. \uc774\ub7f0 \uc608\uc0c1\ub418\ub294 \uc5b4\ub824\uc6c0\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 \ud30c\uc774\uc36c, \ud150\uc11c\ud50c\ub85c\uc6b0\uc758 \ud655\uc7a5\uc744 \uc774\uc6a9\ud574\uc57c \ud558\ub294 \uc774\uc720\ub294 \uc55e\uc11c \uc774\uc57c\uae30\ud55c \uc0dd\ud0dc\uacc4\uc5d0 \ub3d9\ud654\ub418\uae30 \uc704\ud568\uc774\ub2e4. \uc55e\uc5d0\uc11c \uc774\uc57c\uae30\ud55c GPU\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \uc0dd\ud0dc\uacc4\uc758 \uc911\uac04 \uc9c0\uc810\uc5d0 \ud30c\uc774\uc36c\uacfc \ud150\uc11c\ud50c\ub85c\uc6b0\uac00 \uc788\ub2e4. \uc774 \ub450\uac00\uc9c0 ML Framework\ub97c \uae30\ubc18\uc73c\ub85c \ub3d9\uc791\ud558\ub294 \uc0dd\ud0dc\uacc4\ub294 \ub450 ML Framework\uc758 \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\uc744 \uc798 \uc218\uc6a9\ud560 \uacbd\uc6b0 \uc27d\uac8c \ud65c\uc6a9\ud560 \uc218 \uc788\ub2e4. \ubb3c\ub860, \ubcc4\ub3c4\uc758 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud55c \uc0dd\ud0dc\uacc4\uc758 \uacbd\uc6b0 \uac01\uae30 1\ub3001 \ub300\uc751\uc744 \ud574\uc57c \ud558\ub294 \uac83\uc5d0\ub294 \ubcc0\ud568\uc774 \uc5c6\ub2e4.<\/p>\n\n\n\n

LLM \uc0dd\ud0dc\uacc4\ub85c\uc758 NPU \ub3d9\ud654 \uc804\ub7b5<\/h4>\n\n\n\n

LLM\uc740 NLP\uc758 \ud2b9\uc218\ud55c \ud55c \ubd84\uc57c\uc774\ub2e4. \uadf8\ub9ac\uace0 NLP\ub97c DL\ub85c \ud480\uc5b4\ub0b8 \ubd84\uc57c\uc774\uae30 \ub54c\ubb38\uc5d0 DL\uc758 \ud55c \ubd84\uc57c\uc774\ub2e4. \ud558\uc9c0\ub9cc, \ud604\uc7ac AI, \ub354 \ub098\uc544\uac00 \uc804\uc138\uacc4 Tech\uc758 \uac00\uc7a5 \ud070 \ubd84\uc57c\uc774\uae30 \ub54c\ubb38\uc5d0 DL\uc5d0\uc11c\uc758 \uc804\ub7b5\uc5d0 \ud655\uc7a5\uc73c\ub85c \uc804\uac1c \ud560 \uc218 \ub3c4 \uc788\uace0, \ubcc4\ub3c4\uc758 LLM \uc804\ub7b5\uc744 \uc804\uac1c \ud560 \uc218 \ub3c4 \uc788\ub2e4. LLM \uc804\ub7b5\uc744 \ubcc4\ub3c4\ub85c \uc804\uac1c \ud55c\ub2e4\uace0 \ud574\uc11c LLM \uc911 \ud2b9\uc815 \ubaa8\ub378\uc774\ub098 \ubd84\uc57c\ub9cc\uc744 \ub300\uc0c1\uc73c\ub85c \ud55c\ub2e4\ub294 \uac83\uc740 \uc544\ub2c8\ub2e4. LLM \uad00\uc810\uc5d0\uc11c\ub3c4 \ub450 \uac00\uc9c0 \uc804\ub7b5\uc744 \uc0dd\uac01\ud560 \uc218 \uc788\ub2e4. \uc55e\uc11c \uc774\uc57c\uae30 \ud588\ub358 \ub450 \uac00\uc9c0 \uc804\ub7b5\uc911 \ub450 \ubc88\uc9f8 \uc804\ub7b5\uc744 \ud655\uc7a5\ud55c \uc804\ub7b5\uc774\ub2e4. \ub2e4\uc74c\uacfc \uac19\uc740 \ub450\uac00\uc9c0 \uc804\ub7b5\uc744 \uc0dd\uac01\ud560 \uc218 \uc788\ub2e4. \uadf8\ub9ac\uace0 \uc774 \ub450 \uc804\ub7b5\uc758 \uc804\uc81c\ub294 \ud63c\uc790\uc11c\uac00 \uc544\ub2c8\ub77c \uc624\ud508 \uc18c\uc2a4 \uc9c4\uc601\uc758 \ud45c\uc900\ud654\ub41c \uaddc\uc57d\uc744 \ub530\ub77c\uc11c \ud568\uaed8 \uc131\uc7a5\ud560 \uc218 \uc788\ub294 \ud658\uacbd\uc744 \uad6c\ucd95\ud558\ub294 \uac83\uc774\ub2e4.<\/p>\n\n\n\n

    \n
  1. vLLM\uc758 \ud50c\ub7ab\ud3fc(\ud558\ub4dc\uc6e8\uc5b4) PlugIn\uc744 \ud65c\uc6a9\ud55c LLM \uc0dd\ud0dc\uacc4 \ub3d9\ud654<\/li>\n\n\n\n
  2. OpenSource \uc9c4\ud615\uc758 \ud558\ub4dc\uc6e8\uc5b4 \ud655\uc0b0 \uc804\ub7b5\uc744 \ud65c\uc6a9\ud55c LLM \uc0dd\ud0dc\uacc4 \ub3d9\ud654 <\/li>\n<\/ol>\n\n\n\n
    1. vLLM\uc758 \ud50c\ub7ab\ud3fc(\ud558\ub4dc\uc6e8\uc5b4) PlugIn\uc744 \ud65c\uc6a9\ud55c LLM \uc0dd\ud0dc\uacc4 \ub3d9\ud654<\/h5>\n\n\n\n

    \uc5bc\ud54f \ubcf4\uba74 \ub450\uac00\uc9c0\uac00 \uac19\uc740 \uc804\ub7b5\uc778 \uac83 \uac19\uc9c0\ub9cc, \uc0b4\uc9dd \ub2e4\ub974\ub2e4. \uc6b0\uc120 vLLM\uc744 \ud65c\uc6a9\ud558\ub294 \uc804\ub7b5\uc740 \ud604\uc7ac LLM Serving\uc5d0 De-facto\uac00 \ub418\uc5b4\uac00\uace0 \uc788\ub294 vLLM\uc744 \ucda9\uc2e4\ud788 \uc9c0\uc6d0\ud558\ub294 \ubc29\ubc95\uc774\ub2e4. \uc774\ub294 \ud30c\uc774\uc36c, \ud150\uc11c\ud50c\ub85c\uc6b0\uc758 \ud504\ub85c\uadf8\ub798\ubc0d \ubaa8\ub378\uc744 \uc9c0\uc6d0\ud558\ub294 \uc804\ub825\uacfc \uc644\uc804\ud788 \ub3d9\uc77c\ud558\uc9c0\ub9cc \ub300\uc0c1\uc774 ML Framework\uac00 \uc544\ub2cc LLM Serving Framework\uc774\ub77c\ub294 \uac83\uc774\ub2e4. \ud30c\uc774\uc36c\uacfc \uc720\uc0ac\ud558\uac8c vLLM\uc740 \ud50c\ub7ab\ud3fc(\ud558\ub4dc\uc6e8\uc5b4) \ud655\uc7a5\uc744 \uc9c0\uc6d0\ud55c\ub2e4. \uadf8\ub9bc\uc73c\ub85c \uac1c\ub7b5\uc801\uc73c\ub85c \ud45c\uc2dc \ud558\uba74 \uc544\ub798\uc640 \uac19\ub2e4.<\/p>\n\n\n

    \n
    \"\"
    vLLM \ubd84\uc0b0 \uc11c\ube59\uc744 \uc704\ud55c \ud558\ub4dc\uc6e8\uc5b4 \ud50c\ub7ab\ud3fc \ud50c\ub7ec\uadf8\uc778 \uad6c\uc870 \ubc0f \ubd84\uc0b0 \ub77c\uc774\ube0c\ub7ec\ub9ac \uc5f0\uacc4 (\ucd9c\ucc98: \uc790\uccb4)<\/figcaption><\/figure>\n<\/div>\n\n\n

    \uc774 \uc804\ub7b5\uc744 \ud3bc\uce58\uace0 \uc788\ub294 \uac00\uc7a5 \ub300\ud45c\uc801\uc778 NPU \uc81c\uc870\uc0ac\uac00 \ud654\uc6e8\uc774\uc774\ub2e4. \ud654\uc6e8\uc774\ub294 \uc790\uc0ac\uc758 Ascend NPU\ub97c vLLM\uc758 Platform\uc73c\ub85c \uc9c0\uc6d0\ud558\ub824\uace0 \ud558\uace0 \uc788\ub2e4. \ucca0\uc800\ud558\uac8c vLLM\uc758 Platform Plugin \uc815\ucc45\uc744 \uadf8\ub300\ub85c \ub530\ub974\uace0 \uc788\ub2e4. <\/p>\n\n\n

    \n
    \"\"
    vLLM-ascend GitHub Repository – https:\/\/github.com\/vllm-project\/vllm-ascend<\/a><\/figcaption><\/figure>\n<\/div>\n\n\n

    \uad00\ub828\ub41c \ub0b4\uc6a9\uc740 \ub2e4\uc74c \ube14\ub85c\uadf8\uc5d0\uc11c \uc790\uc138\ud788 \ud655\uc778\ud574 \ubcfc \uc218 \uc788\ub2e4 – ‘Introducing vLLM Hardware Plugin, Best Practice from Ascend NPU’ (https:\/\/blog.vllm.ai\/2025\/05\/12\/hardware-plugin.html<\/a>). vLLM\uc758 PlugIn\uc73c\ub85c NPU\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uac8c \ud558\uba74 \uc5ec\ub7ec\uac00\uc9c0 \uc7a5\uc810\uc774 \uc788\ub2e4. NPU\ub294 vLLM\uc758 PlugIn\uc744 \ubb38\uc81c \uc5c6\uc774 \uc9c0\uc6d0\ud558\uace0, \uc790\uc0ac\uc758 \ud558\ub4dc\uc6e8\uc5b4 \ubc84\uc804\uc5d0 \ub530\ub978 \uc720\uc9c0\ubcf4\uc218\ub9cc \uc2e0\uacbd \uc4f0\uba74 \ub41c\ub2e4. vLLM\uc740 vLLM \uae30\ub2a5\uc774 \uc9c0\uc18d\uc801\uc73c\ub85c \uc5c5\ub370\uc774\ud2b8\uac00 \ub41c\ub2e4. \uc911\uc694\ud55c \ubd80\ubd84\uc774 \ubc14\ub85c \uc774 \ubd80\ubd84\uc774\ub2e4. vLLM\uacfc NPU\ub294 \uac01\uae30 \uc5c5\uadf8\ub808\uc774\ub4dc\ub97c \uc9c0\uc18d\ud558\uc9c0\ub9cc \uc0c1\ud638 \uc5f0\uad00\uc774 \ub418\ub294 Interface\ub9cc\uc744 \uc9c0\ud0a8\ub2e4\uba74 \uc815\uc0c1 \ub3d9\uc791\uc744 \ub2f4\ubcf4\ud560 \uc218 \uc788\ub294 \uac83\uc774\ub2e4. \uc774\ub294 \uac1d\uccb4\uc9c0\ud5a5 \uc124\uacc4 \uc6d0\uce59 SOLID \uc911 ISP(\uc778\ud130\ud398\uc774\uc2a4 \ubd84\ub9ac \uc6d0\uce59)\uc774\ub2e4. \ud3ec\uc2a4\ud305\uc5d0\uc11c \ubc18\ubcf5\ub418\uc5c8\ub358 \ub2e8\uc5b4\ub97c \uc0ac\uc6a9\ud558\uba74 \uc0dd\ud0dc\uacc4\uc5d0\uc11c \ub098 \uc790\uc2e0\uc740 \ubcf8\uc778\uc758 \uc5ed\ub7c9\uc744 \ucd5c\ub300\ud654 \ud558\uc5ec \uac00\uc7a5 \uc801\ud569\ud55c \uc601\uc5ed\uc5d0\uc11c\ub9cc \ub178\ub825\ud558\uace0, \ub2e4\ub978 \ubd80\ubd84\ub4e4\uc740 \ub2e4\ub978 \ud50c\ub808\uc774\uc5b4\ub4e4\uc774 \ub178\ub825\uc744 \ud574\uc11c \uc804\uccb4 \uc0dd\ud0dc\uacc4\uac00 \ubc1c\uc804\ud558\uac8c \ub418\ub294 \uac83\uc774\ub2e4. <\/p>\n\n\n\n

    2. OpenSource \uc9c4\ud615\uc758 \ud558\ub4dc\uc6e8\uc5b4 \ud655\uc0b0 \uc804\ub7b5\uc744 \ud65c\uc6a9\ud55c LLM \uc0dd\ud0dc\uacc4 \ub3d9\ud654 <\/h5>\n\n\n\n

    vLLM\uc774 \uc0ac\uc2e4\uc0c1 \ud45c\uc900\uc73c\ub85c \uc790\ub9ac \uc7a1\uc558\uc73c\ub098, Transformer based LLM Serving Framework\uc774\ub77c\ub294 \ud55c\uacc4\ub97c \uac00\uc9c0\uace0 \uc788\ub2e4. \uc989, \ud559\uc2b5\uc5d0 \uc0ac\uc6a9\ub420 \uc218 \uc5c6\uace0 \ub2e4\ub978 DL \ubaa8\ub378\uc744 \ud65c\uc6a9\ud560 \ub54c\uc5d0\ub294 \uc801\uc6a9\ud560 \uc218 \uc5c6\ub2e4\ub294 \uc774\uc57c\uae30 \uc774\ub2e4. \uadf8\ub807\uac8c \ub54c\ubb38\uc5d0 NPU \ud65c\uc6a9\uc744 LLM Serving \uc774\uc0c1\uc73c\ub85c \ud65c\uc6a9\ud558\uae30 \uc704\ud574\uc11c\ub294 \ub2e4\ub978 \ud655\uc0b0 \uc804\ub7b5\uc774 \ud544\uc694\ud558\ub2e4. <\/p>\n\n\n\n

    vLLM\uc740 \uc624\ud508\uc18c\uc2a4\uc774\ub2e4. \ub2e4\ub9cc, \uc801\uc6a9 \ubd84\uc57c\uac00 \ud55c\uc815\uc801\uc774\ub77c\ub294 \uc791\uc740 \ub2e8\uc810\uc774 \uc788\uc744 \ubfd0\uc774\ub2e4. \ubbf8\uad6d, \uc720\ub7fd\uc758 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uac1c\ubc1c\uc790\ub4e4\uc740 \ub298 \uac1c\ubc29\ub41c \ud45c\uc900, \ud2b9\uc815 \ubca4\ud130\uc758 \ub77d\uc778\uc5d0 \ub300\ud574\uc11c \uad49\uc7a5\ud788 \ubd80\uc815\uc801\uc774\ub2e4. NVIDIA GPU\uc5d0 \ub300\ud574\uc11c\ub3c4 \ube44\uc2b7\ud560 \uac83\uc774\ub2e4. \ube44\ub2e8 \ub300\ud55c\ubbfc\uad6d\uc5d0\uc11c\ub9cc NPU\ub97c \ub9cc\ub4e4\uace0 \uc788\uc9c0 \uc54a\uc744 \uac83\uc774\uace0, \ub9ce\uc740 \ud558\ub4dc\uc6e8\uc5b4 \uc5c5\uccb4\ub4e4\uc774 NPU, LPU, IPU \ub4f1\ub4f1 \ub2e4\uc591\ud55c \uc774\ub984\uc73c\ub85c NPU\ub97c \ub9cc\ub4e4\uc5b4\uc11c \ud310\ub9e4 \ud558\uace0 \uc788\ub2e4. \uc774 \uc218 \ub9ce\uc740 \ud558\ub4dc\uc6e8\uc5b4\ub97c \uac00\uc9c0\uace0 \ubaa8\ub378\uc744 \ucd94\ub860\uc5d0 \uc774\uc6a9\ud558\uace0\uc790 \ud560 \ub54c \ud638\ud658\uc131 \uc774\uc678\uc5d0\ub3c4 \uc778\ud130\ud504\ub9ac\ud130\ub85c \ub41c \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud5a5\uc0c1\ud558\uae30 \uc704\ud55c \ub178\ub825\ub3c4 \ud568\uaed8 \uc774\ub8e8\uc5b4\uc9c0\uace0 \uc788\ub2e4. \uac11\uc790\uae30 \uc65c \uc131\ub2a5 \uc774\uc57c\uae30\ub97c \ud558\ub294\uc9c0 \uc758\uc544\ud560\uc9c0\ub3c4 \ubaa8\ub974\uaca0\ub2e4. \ud558\uc9c0\ub9cc, \uc870\uae08\ub9cc \uc774\uc57c\uae30\ub97c \ub4e4\uc5b4 \ubcf4\ub3c4\ub85d \ud558\uc790. \uc778\ud130\ud504\ub9b0\ud130\ub85c \ub3d9\uc791\ud558\ub294 \ud30c\uc774\uc36c\uc73c\ub85c \ub9cc\ub4e4\uc5b4\uc9c4 \ubaa8\ub378\uc744 \ube60\ub974\uac8c \ub3d9\uc791\ud558\uae30 \uc704\ud574\uc11c XLA(Accelerated Linear Algebra)\ub97c \ud45c\uc900\ud654 \ud558\uace0 \uc0ac\uc6a9\ud558\uace0 \uc788\ub2e4. \uc131\ub2a5\uc5d0 \ucd5c\uc801\ud654 \ub418\ub3c4\ub85d \ucf54\ub4dc\ub97c \ubcc0\uacbd\ud55c\ub2e4. ML Framework\uc5d0\uc11c XLA\ub85c \ubc14\ub85c \ubcc0\uacbd\ud558\ub294 \uac83\uc740 \uac00\ub2a5\ud558\ub098 \ubcf5\uc7a1\ud558\uace0 \uc5b4\ub835\uae30 \ub54c\ubb38\uc5d0 \uc774\ub97c \ud55c \ubc88\ub354 \ucd94\uc0c1\ud654\ud55c StableHLO\uac00 \uc874\uc7ac \ud55c\ub2e4. ML Framework\uc5d0\uc11c XLA\ub97c \ucea1\uc290\ud654 \ud55c \ud45c\uc900 \uc5f0\uc0b0\uc758 \ubaa8\uc74c\uc774\ub2e4. \uadf8\ub9ac\uace0 XLA\ub97c \ud558\ub4dc\uc6e8\uc5b4\ub85c\ubd80\ud130 \uc885\uc18d\uc131\uc744 \uc81c\uac70 \ud558\uae30 \uc704\ud55c \uba54\uce74\ub2c8\uc998\uc774 PJRT(Protable JIT)\uc774\ub2e4. XLA\uc758 \ucd5c\uc801\ud654\ub41c \uc5f0\uc0b0\uc744 \ud558\ub4dc\uc6e8\uc5b4\uc5d0\uc11c \ub3d9\uc791\ud558\uae30 \uc704\ud55c \ud1b5\uc77c\ub418\uace0 \ucd94\uc0c1\ud654\ub41c \ub514\ubc14\uc774\uc2a4 API\ub97c \uc81c\uacf5\ud55c\ub2e4. \uc774\uc81c \uc5f0\uacb0\uace0\ub9ac\uac00 \ubcf4\uc77c \uac83\uc774\ub2e4!!!! PJRT\ub97c \uac01 NPU\uc5d0 \ub9de\uac8c \uad6c\ud604\ud558\uba74 XLA\ub97c \uc0ac\uc6a9\ud558\ub294 \ubaa8\ub378 \ucd94\ub860\/\uc11c\ube44\uc2a4\uc5d0\uc11c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub2e4\ub294 \uac83\uc774\ub2e4. <\/p>\n\n\n\n

    OpenXLA, StableHLO, PRJT\ub97c \uc5f0\uacb0\ud574\uc11c \uc0ac\uc6a9\ud558\uba74 \ucd5c\uc801\ud654\ub41c \ubaa8\ub378 \uad6c\ub3d9\uc744 \ud558\ub4dc\uc6e8\uc5b4\uc5d0 \ub3c5\ub9bd\uc801\uc73c\ub85c \ub3d9\uc791\ud558\ub294 \uc18c\ud504\ud2b8\uc6e8\uc5b4\ub97c \uad6c\ucd95\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubc14\uafb8\uc5b4\uc11c \ub9d0\ud558\uba74 OpenXLA, StableHLO, PRJT\ub97c \uc5f0\ub3d9\ud574\uc11c \uad6c\ucd95\ub41c \uc18c\ud504\ud2b8\uc6e8\uc5b4\ub294 \ub2e4\uc591\ud55c NPU\uc5d0\uc11c \uc9c0\uc6d0\ud560 \uc218 \uc788\ub2e4\ub294 \uac83\uc785\ub2c8\ub2e4. \uadf8\ub9ac\uace0 OpenXLA\ub294 ML Framework\uc5d0\uc11c \ubaa8\ub450 \ucc44\ud0dd \ud588\uae30 \ub54c\ubb38\uc5d0 \ubaa8\ub4e0 ML Framework\uc5d0\uc11c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. vLLM Platform PlugIn \ubc29\uc2dd\uacfc \ub9c8\ucc2c\uac00\uc9c0\ub85c \uc0dd\ud0dc\uacc4\uac00 \ud655\uc7a5\ub418\ub354\ub77c\ub3c4 \uc774\uc5d0 \ub530\ub77c\uc11c \ubcc0\uacbd\ub418\uc5b4\uc57c \ud560 \ub0b4\uc6a9\uc774 \uc801\uc740 \uc7a5\uc810\uc774 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\ub9cc, XLA\ub97c \uc0ac\uc6a9\ud560 \ub54c \uc758\ub3c4\uc640\ub294 \ubc18\ub300\ub85c \uc131\ub2a5\uc774 \uae09\uaca9\ud788 \ub5a8\uc5b4\uc9c0\ub294 \uacbd\uc6b0\ub3c4 \uc788\uc5b4\uc11c XLA\ub97c \uc704\ud55c \ucd5c\uc801\ud654 \ub2e8\uacc4\uac00 \ubc18\ub4dc\uc2dc \ud544\uc694\ud558\ub2e4. 1, 2 \ubc88 \uc804\ub7b5\uc744 \ube44\uad50 \ud558\uba74 \uc544\ub798 \uadf8\ub9bc\uacfc \uac19\ub2e4. <\/p>\n\n\n

    \n
    \"\"
    NPU\ub97c \uae30\uc874 \uc0dd\ud0dc\uacc4\uc5d0 \ub3d9\ud654\ub97c \uc704\ud55c \uc804\ub7b5 (\ucd9c\ucc98: Gemini Canvas \uc0dd\uc131) <\/figcaption><\/figure>\n<\/div>\n\n\n

    \ub9c8\uce58\uba70<\/h3>\n\n\n\n

    AI \ud558\ub4dc\uc6e8\uc5b4 \uc2dc\uc7a5\uc5d0\uc11c NPU\uac00 \uc131\uacf5\uc801\uc73c\ub85c \uc548\ucc29\ud558\uae30 \uc704\ud55c \ud575\uc2ec\uc740 \ub2e8\uc21c\ud788 \ud558\ub4dc\uc6e8\uc5b4 \uc131\ub2a5\uc758 \uc6b0\uc704\ub97c \ub118\uc5b4, \uad11\ubc94\uc704\ud558\uac8c \uad6c\ucd95\ub41c \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc0dd\ud0dc\uacc4\uc640\uc758 \uc870\ud654\ub85c\uc6b4 \ud1b5\ud569\uc5d0 \ub2ec\ub824 \uc788\ub2e4. \uc6d0\ubb38\uc774 \uc778\uc6a9\ud55c “\ube68\ub9ac \uac00\ub824\uba74 \ud63c\uc790 \uac00\uace0, \uba40\ub9ac \uac00\ub824\uba74 \ud568\uaed8 \uac00\ub77c”\ub294 \uaca9\uc5b8\ucc98\ub7fc <\/sup>, NPU \ubca4\ub354\ub4e4\uc740 \ub3c5\uc790\uc801\uc778 \uae30\uc220 \uac1c\ubc1c\uc5d0\ub9cc \ubab0\ub450\ud558\uae30\ubcf4\ub2e4\ub294, \uc218\uc2ed \ub144\uac04\uc758 \ub178\ub825\uc73c\ub85c \uacf5\uace0\ud788 \uc313\uc544 \uc62c\ub824\uc9c4 <\/sup> \uae30\uc874\uc758 \uac15\ub825\ud55c \uc624\ud508\uc18c\uc2a4 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc0dd\ud0dc\uacc4\uc5d0 \uc801\uadf9\uc801\uc73c\ub85c \ub3d9\ud654\ud558\uace0 \uae30\uc5ec\ud558\ub294 \uc804\ub7b5\uc744 \ucc44\ud0dd\ud574\uc57c \ud55c\ub2e4. \uc774\ub294 PyTorch, TensorFlow\uc640 \uac19\uc740 \uc8fc\ub958 ML \ud504\ub808\uc784\uc6cc\ud06c\uc758 \ud655\uc7a5 \ud615\uc2dd \uc9c0\uc6d0, VLLM\uacfc \uac19\uc740 LLM \uc11c\ube59 \ud504\ub808\uc784\uc6cc\ud06c\uc758 \ud50c\ub7ec\uadf8\uc778 \ucc38\uc5ec, \uadf8\ub9ac\uace0 OpenXLA\/PJRT\uc640 \uac19\uc740 \uc624\ud508 \ud45c\uc900 \uc774\ub2c8\uc154\ud2f0\ube0c\uc5d0 \ub300\ud55c \uc801\uadf9\uc801\uc778 \uae30\uc5ec\ub97c \ud1b5\ud574 \uc774\ub8e8\uc5b4\uc9c8 \uc218 \uc788\ub2e4.<\/sup> \uc774\ub7ec\ud55c \ub178\ub825\uc740 NPU\uac00 \uc2dc\uc7a5\uc5d0 \uc131\uacf5\uc801\uc73c\ub85c \uc9c4\uc785\ud558\uace0 \uc9c0\uc18d \uac00\ub2a5\ud55c \uc131\uc7a5\uc744 \uc774\ub8e8\ub294 \ub370 \ud544\uc218\uc801\uc778 \uc694\uc18c\uc774\ub2e4.  <\/p>\n\n\n\n

    \uc774\ub7ec\ud55c \ud611\ub825\uc801 \uc811\uadfc \ubc29\uc2dd\uc740 \ub2e8\uc21c\ud788 \uae30\uc220\uc801 \ud638\ud658\uc131\uc744 \ud655\ubcf4\ud558\ub294 \uac83\uc744 \ub118\uc5b4\uc120\ub2e4. \uc774\ub294 \uad11\ubc94\uc704\ud55c \uac1c\ubc1c\uc790 \ucee4\ubba4\ub2c8\ud2f0\uc758 \uc9c0\uc9c0\uc640 \ucc38\uc5ec\ub97c \uc774\ub04c\uc5b4\ub0b4\uace0, \uc7a5\uae30\uc801\uc778 \uc0b0\uc5c5 \ud45c\uc900\uc744 \ud615\uc131\ud558\uba70, \ud2b9\uc815 \ubca4\ub354\uc5d0 \ub300\ud55c \uc885\uc18d\uc131(lock-in)\uc744 \uc904\uc774\ub294 \ub370 \uae30\uc5ec\ud55c\ub2e4. NPU\uac00 \uae30\uc874 GPU \uc0dd\ud0dc\uacc4\uc5d0 \ud3b8\uc785\ub418\ub294 \uac83\uc740 \uadf8 \uc790\uccb4\uc758 \uc815\uccb4\uc131\uc744 \uc783\ub294 \uac83\uc774 \uc544\ub2c8\ub77c, \uc624\ud788\ub824 AI \ud558\ub4dc\uc6e8\uc5b4 \uc0dd\ud0dc\uacc4\ub97c \ub354\uc6b1 \ub2e4\uc591\ud558\uace0 \ud6a8\uc728\uc801\uc778 \ubc29\ud5a5\uc73c\ub85c \ud655\uc7a5\ud558\uba70 \ud568\uaed8 \uc131\uc7a5\ud558\ub294 \uc0c1\uc0dd\uc758 \uae38\uc774\ub2e4.<\/sup> \uad81\uadf9\uc801\uc73c\ub85c, \uc774\ub7ec\ud55c \uc804\ub7b5\uc801 \ud611\ub825\uacfc \uc624\ud508\uc18c\uc2a4 \uc0dd\ud0dc\uacc4\uc5d0 \ub300\ud55c \uc9c0\uc18d\uc801\uc778 \uae30\uc5ec\ub294 NPU\uac00 AI \ud558\ub4dc\uc6e8\uc5b4 \uc2dc\uc7a5\uc5d0\uc11c \uc9c0\uc18d \uac00\ub2a5\ud55c \uacbd\uc7c1 \uc6b0\uc704\ub97c \ud655\ubcf4\ud558\uace0, \ub098\uc544\uac00 \uc778\ub958\uc758 \uc0b6\uc744 \ubcc0\ud654\uc2dc\ud0ac AI \uae30\uc220\uc758 \ubc1c\uc804\uc5d0 \uae30\uc5ec\ud558\ub294 \ud575\uc2ec \ub3d9\ub825\uc774 \ub420 \uac83\uc774\ub2e4. \uc815\ubd80\uc758 AI \uce69 \ud65c\uc131\ud654 \uc815\ucc45\uacfc \ub9de\ubb3c\ub824 <\/sup>, NPU\ub294 \ubbf8\ub798 AI \uc778\ud504\ub77c\uc758 \uc911\uc694\ud55c \ucd95\uc73c\ub85c\uc11c \uadf8 \uc7a0\uc7ac\ub825\uc744 \ucd5c\ub300\ud55c \ubc1c\ud718\ud560 \uc218 \uc788\uc744 \uac83\uc73c\ub85c \uae30\ub300\ub41c\ub2e4.<\/p>\n\n\n\n

    <\/p>\n","protected":false},"excerpt":{"rendered":"

    AI \uac1c\ubc1c\uc5d0 \ub300\uaddc\ubaa8 GPU\ub294 \ud544\uc218\uc801\uc778 \uc694\uc18c\ub85c \uac04\uc8fc\ub41c\ub2e4. \ud55c\uad6d \uc815\ubd80\ub294 2025\ub144 \ucd08 \ubd80\uc871\ud55c GPU\ub97c \uacf5\uacf5 \uc778\ud504\ub77c \ud615\ud0dc\ub85c \uc81c\uacf5\ud560 \uacc4\ud68d\uc744 \ubc1c\ud45c\ud558\uc600\ub2e4. \uc774 \uc2dc\ub9ac\uc988\uc758 \ud575\uc2ec \uc8fc\uc81c\uc774\uae30\ub3c4 \ud558\ub2e4. \uadf8\ub7f0\ub370 \uad6d\ub0b4\uc678\uc801\uc73c\ub85c \ub2e4\uc591\ud55c \uc124\uacc4\ub97c \uae30\ubc18\uc73c\ub85c \ub2e4\uc591\ud55c\uc77d\uc5b4 \ubcfc\uae4c?GPU\uc740 \uc5b4\ub5bb\uac8c AI \uc2dc\ub300\uc758 \ud575\uc2ec\uc774 \ub418\uc5c8\ub294\uac00? \u2013\u00a0NPU, XPU<\/span><\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":889,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[224,72,63,1],"tags":[179,199,75,168,171,307,308,306],"class_list":["post-863","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-cloud","category-gpgpu","category-parallel-computing","category-uncategorized","tag-ai","tag-cloud","tag-gpgpu","tag-gpu","tag-npu","tag-openxla","tag-pjrt","tag-vllm"],"_links":{"self":[{"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/863","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=863"}],"version-history":[{"count":11,"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/863\/revisions"}],"predecessor-version":[{"id":895,"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/863\/revisions\/895"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/media\/889"}],"wp:attachment":[{"href":"http:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=863"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=863"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=863"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}