{"id":1331,"date":"2026-01-19T10:00:22","date_gmt":"2026-01-19T01:00:22","guid":{"rendered":"https:\/\/rtlearner.com\/?p=1331"},"modified":"2026-01-19T10:00:56","modified_gmt":"2026-01-19T01:00:56","slug":"ai-architecture-11-mobilenet-depthwise-separable","status":"publish","type":"post","link":"https:\/\/rtlearner.com\/en\/ai-architecture-11-mobilenet-depthwise-separable\/","title":{"rendered":"AI Architecture 11. Depthwise Separable Conv: The MobileNet Paradox"},"content":{"rendered":"\n<p>\uc9c0\ub09c <a href=\"https:\/\/rtlearner.com\/ai-architecture-9-convolution-operation-mapping\/\" data-type=\"post\" data-id=\"1312\">Conv \uc5f0\uc0b0\uc758 3\uac00\uc9c0 \ub9e4\ud551<\/a>\uc5d0\uc11c \uc6b0\ub9ac\ub294 \uc77c\ubc18\uc801\uc778 \ud569\uc131\uacf1(Standard Convolution)\uc744 \ud558\ub4dc\uc6e8\uc5b4\uc5d0\uc11c \ucc98\ub9ac\ud560 \ub54c, <strong>Im2Col<\/strong> \ubc29\uc2dd\uc744 \ud1b5\ud574 \uba54\ubaa8\ub9ac\ub97c \ud76c\uc0dd\ud558\uace0 \uc5f0\uc0b0 \uc18d\ub3c4(GEMM)\ub97c \uc5bb\ub294 \uc804\ub7b5\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n<p>2017\ub144, \uad6c\uae00\uc774 \ubc1c\ud45c\ud55c <strong>MobileNet<\/strong>\uc740 \ubaa8\ubc14\uc77c \ud658\uacbd\uc744 \uc704\ud55c \ud601\uba85\uc801\uc778 \ubaa8\ub378\uc774\uc5c8\uc2b5\ub2c8\ub2e4. \uae30\uc874 \ubaa8\ub378 \ub300\ube44 \uc5f0\uc0b0\ub7c9(FLOPs)\uacfc \ud30c\ub77c\ubbf8\ud130 \uc218\ub97c 1\/10 \uc218\uc900\uc73c\ub85c \uc904\uc774\uba74\uc11c\ub3c4 \uc900\uc218\ud55c \uc815\ud655\ub3c4\ub97c \ubcf4\uc5ec\uc8fc\uc5c8\uae30 \ub54c\ubb38\uc785\ub2c8\ub2e4. \uadf8 \ube44\uacb0\uc740 <strong>Depthwise Separable Convolution<\/strong>\uc774\ub77c\ub294 \ub3c5\ud2b9\ud55c \uad6c\uc870\uc5d0 \uc788\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n<p>\ud558\uc9c0\ub9cc \uc774 \uad6c\uc870\ub294 \ud558\ub4dc\uc6e8\uc5b4 \uc5d4\uc9c0\ub2c8\uc5b4\ub4e4\uc5d0\uac8c MobileNet\uc758 \uc5ed\uc124&#8221;\uc774\ub77c\uace0 \ubd88\ub9ac\ub294 \ud765\ubbf8\ub85c\uc6b4 \ud604\uc0c1\uc744 \uc548\uaca8\uc8fc\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>&#8220;\uc5f0\uc0b0\ub7c9\uc740 90%\uac00 \uc904\uc5c8\ub294\ub370, \uc65c \uc2e4\uc81c \uc2e4\ud589 \uc18d\ub3c4(Latency)\ub294 \uadf8\ub9cc\ud07c \ube68\ub77c\uc9c0\uc9c0 \uc54a\ub294\uac00?&#8221;<\/p>\n<\/blockquote>\n\n\n\n<p>\uc774\ubc88 \uae00\uc5d0\uc11c\ub294 \uc18c\ud504\ud2b8\uc6e8\uc5b4\uc801\uc73c\ub85c\ub294 \uc644\ubcbd\ud574 \ubcf4\uc774\ub294 \ubaa8\ub378\uc774, \ud558\ub4dc\uc6e8\uc5b4(NPU\/GPU) \ub0b4\ubd80\uc5d0\uc11c\ub294 \uc65c <strong>\uac00\ub3d9\ub960(Utilization)\uc758 \uae09\ub77d<\/strong>\uc774\ub77c\ub294 \ubd80\uc791\uc6a9\uc744 \ub0b3\ub294\uc9c0, \ud558\ub4dc\uc6e8\uc5b4 \uc5d4\uc9c0\ub2c8\uc5b4\uc758 \uad00\uc810\uc5d0\uc11c \uadf8 \ubb3c\ub9ac\uc801 \uc6d0\uc778\uc744 \ud30c\ud5e4\uccd0 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n<style>.kb-table-of-content-nav.kb-table-of-content-id1331_a487d9-5b .kb-table-of-content-wrap{padding-top:var(--global-kb-spacing-sm, 1.5rem);padding-right:var(--global-kb-spacing-sm, 1.5rem);padding-bottom:var(--global-kb-spacing-sm, 1.5rem);padding-left:var(--global-kb-spacing-sm, 1.5rem);box-shadow:0px 0px 14px 0px rgba(0, 0, 0, 0.2);}.kb-table-of-content-nav.kb-table-of-content-id1331_a487d9-5b .kb-table-of-contents-title-wrap{padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.kb-table-of-content-nav.kb-table-of-content-id1331_a487d9-5b .kb-table-of-contents-title{font-weight:regular;font-style:normal;}.kb-table-of-content-nav.kb-table-of-content-id1331_a487d9-5b .kb-table-of-content-wrap .kb-table-of-content-list{font-weight:regular;font-style:normal;margin-top:var(--global-kb-spacing-sm, 1.5rem);margin-right:0px;margin-bottom:0px;margin-left:0px;}@media all and (max-width: 767px){.kb-table-of-content-nav.kb-table-of-content-id1331_a487d9-5b .kb-table-of-contents-title{font-size:var(--global-kb-font-size-md, 1.25rem);}.kb-table-of-content-nav.kb-table-of-content-id1331_a487d9-5b .kb-table-of-content-wrap .kb-table-of-content-list{font-size:var(--global-kb-font-size-sm, 0.9rem);}}<\/style>\n\n<style>.kadence-column1331_b44b97-07 > .kt-inside-inner-col{box-shadow:0px 0px 14px 0px rgba(0, 0, 0, 0.2);}.kadence-column1331_b44b97-07 > .kt-inside-inner-col,.kadence-column1331_b44b97-07 > .kt-inside-inner-col:before{border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;}.kadence-column1331_b44b97-07 > .kt-inside-inner-col{column-gap:var(--global-kb-gap-sm, 1rem);}.kadence-column1331_b44b97-07 > .kt-inside-inner-col{flex-direction:column;}.kadence-column1331_b44b97-07 > .kt-inside-inner-col > .aligncenter{width:100%;}.kadence-column1331_b44b97-07 > .kt-inside-inner-col:before{opacity:0.3;}.kadence-column1331_b44b97-07{position:relative;}@media all and (max-width: 1024px){.kadence-column1331_b44b97-07 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}@media all and (max-width: 767px){.kadence-column1331_b44b97-07 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}<\/style>\n<div class=\"wp-block-kadence-column kadence-column1331_b44b97-07\"><div class=\"kt-inside-inner-col\">\n<p><strong>\uad00\ub828 \uae00<\/strong><\/p>\n\n\n\n<p>\u2705<a href=\"https:\/\/rtlearner.com\/ai-architecture-1-neuron-hardware-mac-analysis\/\" data-type=\"post\" data-id=\"1248\">AI Architecture 1. \uc778\uacf5 \ub274\ub7f0\uc758 \ud574\ubd80: silicon\uc5d0\uc11c Y=WX+B \uad6c\ud604<\/a><\/p>\n\n\n\n<p>\u2705<a href=\"https:\/\/rtlearner.com\/ai-architecture-2-activation-relu-vs-sigmoid\/\" data-type=\"post\" data-id=\"1255\">AI Architecture 2. \ud65c\uc131\ud654 \ud568\uc218\uc758 \ube44\uc6a9: ReLU vs Sigmoid<\/a><\/p>\n\n\n\n<p>\u2705<a href=\"https:\/\/rtlearner.com\/ai-architecture-3-matmul-simd-parallel-processing\/\" data-type=\"post\" data-id=\"1263\">AI Architecture 3. \ud589\ub82c\uacf1(MatMul)\uc758 \ubbf8\ud559: \ub525\ub7ec\ub2dd\uc774 GPU\/NPU\ub97c \uc120\ud0dd\ud55c \uc774\uc720<\/a><\/p>\n\n\n\n<p>\u2705<a href=\"https:\/\/rtlearner.com\/ai-architecture-4-training-vs-inference\/\" data-type=\"post\" data-id=\"1267\">AI Architecture 4. \ud559\uc2b5(Training) vs \ucd94\ub860(Inference)<\/a><\/p>\n<\/div><\/div>\n\n\n\n<h2 class=\"wp-block-heading\">1. Depthwise Separable Convolution<\/h2>\n\n\n\n<p>\uba3c\uc800, MobileNet\uc774 \uc5b4\ub5bb\uac8c \uc5f0\uc0b0\ub7c9\uc744 \uc904\uc600\ub294\uc9c0 \uac04\ub2e8\ud788 \uc9da\uace0 \ub118\uc5b4\uac11\uc2dc\ub2e4. \ud45c\uc900 Conv \uc5f0\uc0b0\uc744 \ub450 \ub2e8\uacc4\ub85c \ucabc\uac1c\ub294 \uac83\uc774 \ud575\uc2ec\uc785\ub2c8\ub2e4.<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>Depthwise Conv:<\/strong> \uac01 \uc785\ub825 \ucc44\ub110\ub9c8\ub2e4 \ubcc4\uac1c\uc758 \ud544\ud130\ub97c \uc801\uc6a9\ud569\ub2c8\ub2e4. (\ucc44\ub110 \uac04 \uc815\ubcf4 \uad50\ud658 \uc5c6\uc74c)<\/li>\n\n\n\n<li><strong>Pointwise Conv:<\/strong> 1 * 1 \ud06c\uae30\uc758 \ud544\ud130\ub85c \ucc44\ub110 \uac04\uc758 \uc815\ubcf4\ub97c \uc11e\uc5b4\uc90d\ub2c8\ub2e4.<\/li>\n<\/ol>\n\n\n\n<p>\uc785\ub825 \ub370\uc774\ud130\uac00 H * W * C, \ucee4\ub110 \ud06c\uae30\uac00 K * K\uc77c \ub54c, \uc5f0\uc0b0\ub7c9\uc740 \ub2e4\uc74c\uacfc \uac19\uc774 \uac10\uc18c\ud569\ub2c8\ub2e4.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Standard Conv:<\/strong><\/li>\n<\/ul>\n\n\n\n<div class=\"wp-block-math\"><math display=\"block\"><semantics><mrow><mi>H<\/mi><mo>\u22c5<\/mo><mi>W<\/mi><mo>\u22c5<\/mo><mi>C<\/mi><mo>\u22c5<\/mo><msup><mi>K<\/mi><mn>2<\/mn><\/msup><mo>\u22c5<\/mo><mi>N<\/mi><mtext>&nbsp;(N&nbsp;=&nbsp;number&nbsp;of&nbsp;output&nbsp;channel)<\/mtext><\/mrow><annotation encoding=\"application\/x-tex\">H \\cdot W \\cdot C \\cdot K^2 \\cdot N \\text{ (N = number of output channel)}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Depthwise Separable:<\/strong><\/li>\n<\/ul>\n\n\n\n<div class=\"wp-block-math\"><math display=\"block\"><semantics><mrow><mo form=\"prefix\" stretchy=\"false\">(<\/mo><mi>H<\/mi><mo>\u22c5<\/mo><mi>W<\/mi><mo>\u22c5<\/mo><mi>C<\/mi><mo>\u22c5<\/mo><msup><mi>K<\/mi><mn>2<\/mn><\/msup><mo form=\"postfix\" stretchy=\"false\">)<\/mo><mo>+<\/mo><mo form=\"prefix\" stretchy=\"false\">(<\/mo><mi>H<\/mi><mo>\u22c5<\/mo><mi>W<\/mi><mo>\u22c5<\/mo><mi>C<\/mi><mo>\u22c5<\/mo><mi>N<\/mi><mo form=\"postfix\" stretchy=\"false\">)<\/mo><\/mrow><annotation encoding=\"application\/x-tex\">(H \\cdot W \\cdot C \\cdot K^2) + (H \\cdot W \\cdot C \\cdot N)<\/annotation><\/semantics><\/math><\/div>\n\n\n\n<p>\uc218\uc2dd\uc801\uc73c\ub85c \uacc4\uc0b0\ud558\uba74 \uc57d <strong>8~9\ubc30\uc758 \uc5f0\uc0b0\ub7c9 \uac10\uc18c<\/strong> \ud6a8\uacfc\uac00 \uc788\uc2b5\ub2c8\ub2e4. \uc22b\uc790\ub9cc \ubcf4\uba74 \ud558\ub4dc\uc6e8\uc5b4\uac00 9\ubc30 \ub35c \uc77c\ud574\ub3c4 \ub418\ub2c8 9\ubc30 \ube68\ub77c\uc838\uc57c \ud560 \uac83 \uac19\uc2b5\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \ud604\uc2e4\uc740 \uadf8\ub807\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">2. \uc5ed\uc124\uc758 \uc6d0\uc778 1: \ub0ae\uc740 \uc5f0\uc0b0 \uac15\ub3c4 (Low Arithmetic Intensity)<\/h2>\n\n\n\n<p>\ud558\ub4dc\uc6e8\uc5b4 \uc131\ub2a5\uc744 \uacb0\uc815\ud558\ub294 \ud575\uc2ec \uc9c0\ud45c\ub294 Arithmetic Intensity (\uc5f0\uc0b0 \uac15\ub3c4)\uc785\ub2c8\ub2e4. \uc989, &#8220;\uba54\ubaa8\ub9ac\uc5d0\uc11c \ub370\uc774\ud130 1\ubc14\uc774\ud2b8\ub97c \uac00\uc838\uc654\uc744 \ub54c \uba87 \ubc88\uc758 \uc5f0\uc0b0\uc744 \uc218\ud589\ud558\ub294\uac00?&#8221;\uc785\ub2c8\ub2e4.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Standard Conv:<\/strong> \uc785\ub825 \ucc44\ub110 \ud558\ub098\uac00 N\uac1c\uc758 \ucd9c\ub825 \ucc44\ub110 \ud544\ud130\uc640 \ubaa8\ub450 \uc5f0\uc0b0\ub429\ub2c8\ub2e4. \ub370\uc774\ud130 \uc7ac\uc0ac\uc6a9(Reuse)\uc774 \uc544\uc8fc \ub192\uc2b5\ub2c8\ub2e4.<\/li>\n\n\n\n<li><strong>Depthwise Conv:<\/strong> \uc785\ub825 \ucc44\ub110 C\ub294 \uc624\uc9c1 \ucee4\ub110 \ucc44\ub110 C \ud558\uace0\ub9cc \uc5f0\uc0b0\ub429\ub2c8\ub2e4. <strong>Cross-channel Reuse\uac00 \uc804\ud600 \uc5c6\uc2b5\ub2c8\ub2e4.<\/strong><\/li>\n<\/ul>\n\n\n\n<p>Depthwise Conv\ub294 \ub370\uc774\ud130\ub97c \uba54\ubaa8\ub9ac\uc5d0\uc11c \ud798\ub4e4\uac8c \uac00\uc838\uc640\uc11c, \uace0\uc791 \uba87 \ubc88 \uacf1\ud558\uace0 \ub05d\ub0a9\ub2c8\ub2e4. \uc774\ub294 7\ubc88 \uae00\uc5d0\uc11c \ub2e4\ub8ec <strong>[MLP\uc640 \uba54\ubaa8\ub9ac \uc7a5\ubcbd]<\/strong> \ubb38\uc81c\uc640 \uc720\uc0ac\ud569\ub2c8\ub2e4. \uc5f0\uc0b0\uae30\uac00 \ubc14\uc058\uac8c \ub3cc\uc544\uac00\uae30\ub3c4 \uc804\uc5d0, \uba54\ubaa8\ub9ac \ub300\uc5ed\ud3ed\uc774 \ubcd1\ubaa9(Memory-Bound)\uc774 \ub418\uc5b4 \uc131\ub2a5 \uc800\ud558\ub97c \uc77c\uc73c\ud0b5\ub2c8\ub2e4.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">3. \uc5ed\uc124\uc758 \uc6d0\uc778 2: MAC \uac00\ub3d9\ub960(Utilization)\uc758 \ud30c\ud3b8\ud654<\/h2>\n\n\n\n<p>\ub300\ubd80\ubd84\uc758 \uace0\uc131\ub2a5 NPU\ub098 GPU\ub294 \uac70\ub300\ud55c \ud589\ub82c\uacf1(GEMM)\uc744 \ucc98\ub9ac\ud558\uae30 \uc704\ud574 \uc218\uc2ed, \uc218\ubc31 \uac1c\uc758 MAC \uc720\ub2db\uc744 \ud558\ub098\uc758 \uac70\ub300\ud55c \ubc30\uc5f4(Systolic Array \ub4f1)\uc774\ub098 \ub113\uc740 \ubca1\ud130 \ub808\uc9c0\uc2a4\ud130(SIMD)\ub85c \ubb36\uc5b4 \ub193\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n<p>\uc608\ub97c \ub4e4\uc5b4, <strong>64\uac1c\uc758 MAC\uc774 \ud55c \ubb36\uc74c<\/strong>\uc73c\ub85c \ub3d9\uc791\ud558\ub294 NPU\uac00 \uc788\ub2e4\uace0 \uac00\uc815\ud574 \ubd05\uc2dc\ub2e4.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Pointwise Conv (1 * 1, \ucc44\ub110 \uc11e\uae30):<\/strong> \ucc44\ub110 \uac04 \uc5f0\uc0b0\uc774\ubbc0\ub85c 64\uac1c MAC\uc774 \uaf49 \ucc28\uc11c(Dense) 100% \ud6a8\uc728\ub85c \ub3cc\uc544\uac11\ub2c8\ub2e4.<\/li>\n\n\n\n<li><strong>Depthwise Conv (\ucc44\ub110 \ubcc4 \ub3c5\ub9bd):<\/strong> \uac01 \ucc44\ub110\uc774 \ub3c5\ub9bd\uc801\uc785\ub2c8\ub2e4. \ub9cc\uc57d \ud558\ub4dc\uc6e8\uc5b4 \uc2a4\ucf00\uc904\ub7ec\uac00 \ucc44\ub110 \ub2e8\uc704\ub85c \ubcd1\ub82c\ud654\ub97c \uc2dc\ub3c4\ud558\ub294\ub370, \uba54\ubaa8\ub9ac \ub808\uc774\uc544\uc6c3\uc774\ub098 \ub370\uc774\ud130 \uc758\uc874\uc131 \ubb38\uc81c\ub85c \ud55c \ubc88\uc5d0 1\uac1c \ucc44\ub110\uc529\ub9cc \ucc98\ub9ac\ud574\uc57c \ud55c\ub2e4\uba74?\n<ul class=\"wp-block-list\">\n<li><strong>64\uac1c \uc911 1\uac1c\ub9cc \uc77c\ud558\uace0 63\uac1c\ub294 \ub189\ub2c8\ub2e4.<\/strong> (\uac00\ub3d9\ub960 1.5%)<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\uc774\uac83\uc774 <strong>MAC Starvation<\/strong> \ud604\uc0c1\uc785\ub2c8\ub2e4. Depthwise \uc5f0\uc0b0\uc740 \ud558\ub4dc\uc6e8\uc5b4 \uc785\uc7a5\uc5d0\uc11c \ub108\ubb34 \uc798\uac8c \ucabc\uac1c\uc9c4 \uc77c\uac10\uc774\ub77c, \ub300\uaddc\ubaa8 \ubcd1\ub82c \ucc98\ub9ac \uc7a5\uce58\uc758 \ud6a8\uc728\uc744 \ub5a8\uc5b4\ub728\ub9bd\ub2c8\ub2e4.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">4. \uba54\ubaa8\ub9ac \ub808\uc774\uc544\uc6c3\uc758 \ucda9\ub3cc: Channel-First vs Channel-Last<\/h2>\n\n\n\n<p>\ud558\ub4dc\uc6e8\uc5b4 \ud6a8\uc728\uc131\uc740 \ub370\uc774\ud130\uac00 \uba54\ubaa8\ub9ac\uc5d0 \uc5b4\ub5bb\uac8c \ub193\uc5ec\uc788\ub290\ub0d0(Layout)\uc5d0 \ub530\ub77c \uacb0\uc815\ub429\ub2c8\ub2e4.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>NCWH (Channel-First):<\/strong> \ucc44\ub110\uc774 \uba3c\uc800 \ub098\uc635\ub2c8\ub2e4.<\/li>\n\n\n\n<li><strong>NHWC (Channel-Last):<\/strong> \ud53d\uc140 \uc704\uce58\ubcc4\ub85c \ucc44\ub110\uc774 \ubb49\uccd0 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<\/ul>\n\n\n\n<p>\uc77c\ubc18\uc801\uc73c\ub85c \uace0\uc131\ub2a5 NPU\ub294 \ubca1\ud130 \uc5f0\uc0b0\uc744 \uc704\ud574 NHWC\ub97c \uc120\ud638\ud569\ub2c8\ub2e4. 1 * 1 Pointwise Conv\ub97c \ud560 \ub54c \ud2b9\uc815 \ud53d\uc140\uc758 \ubaa8\ub4e0 \ucc44\ub110\uc744 \ud55c \ubc88\uc5d0 \uac00\uc838\uc624\uae30 \uc88b\uae30 \ub54c\ubb38\uc785\ub2c8\ub2e4.<\/p>\n\n\n\n<p>\ud558\uc9c0\ub9cc Depthwise Conv\ub294 \uacf5\uac04\uc801(Spatial) \uc5f0\uc0b0\uc774\ubbc0\ub85c 3 * 3 \uc601\uc5ed\uc758 \ud53d\uc140\ub4e4\uc774 \ud544\uc694\ud569\ub2c8\ub2e4. NHWC \uad6c\uc870\uc5d0\uc11c\ub294 \uc774\uc6c3\ud55c \ud53d\uc140\ub4e4\uc774 \uba54\ubaa8\ub9ac\uc0c1\uc5d0\uc11c \uba40\ub9ac \ub5a8\uc5b4\uc838 \uc788\uc5b4, \uce90\uc2dc \ubbf8\uc2a4(Cache Miss)\ub97c \uc720\ubc1c\ud558\uac70\ub098 \ubcf5\uc7a1\ud55c \uc154\ud50c(Shuffle) \ub85c\uc9c1\uc744 \ud544\uc694\ub85c \ud569\ub2c8\ub2e4.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">5. \uacb0\ub860: FLOPs\ub294 \uc18d\ub3c4\uac00 \uc544\ub2c8\ub2e4<\/h2>\n\n\n\n<p>MobileNet\uc740 \ubd84\uba85 \ud6cc\ub96d\ud55c \ubaa8\ub378\uc785\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \uc5f0\uc0b0\ub7c9\uc774 \uc801\ub2e4(Low FLOPs)\ub294 \uac83\uc774 \ubc18\ub4dc\uc2dc \ud558\ub4dc\uc6e8\uc5b4\uc5d0\uc11c \ube60\ub974\ub2e4(Low Latency)\ub97c \uc758\ubbf8\ud558\uc9c0 \uc54a\ub294\ub2e4\ub294 \uac83\uc744 \uc99d\uba85\ud55c \ub300\ud45c\uc801\uc778 \uc0ac\ub840\uc774\uae30\ub3c4 \ud569\ub2c8\ub2e4.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Standard Conv:<\/strong> \uc5f0\uc0b0\ub7c9\uc740 \ub9ce\uc9c0\ub9cc, \ud558\ub4dc\uc6e8\uc5b4 \ud6a8\uc728(Utilization)\uc774 \ub192\uc544 <strong>\uac00\uc131\ube44 \uc88b\uc740 \ub178\ub3d9<\/strong>\uc744 \ud569\ub2c8\ub2e4.<\/li>\n\n\n\n<li><strong>Depthwise Conv:<\/strong> \uc5f0\uc0b0\ub7c9\uc740 \uc801\uc9c0\ub9cc, \uba54\ubaa8\ub9ac \ubcd1\ubaa9\uacfc \ud558\ub4dc\uc6e8\uc5b4 \uc720\ud734 \uc0c1\ud0dc(Idle)\ub97c \uc720\ubc1c\ud558\uc5ec <strong>\ube44\ud6a8\uc728\uc801\uc778 \ub178\ub3d9<\/strong>\uc744 \ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n\n\n\n<p>\ucd5c\uadfc\uc758 NPU \uc544\ud0a4\ud14d\ucc98\ub4e4\uc740 \uc774 \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 Depthwise \uc804\uc6a9 \uac00\uc18d \uc5d4\uc9c4\uc744 \ubcc4\ub3c4\ub85c \ud0d1\uc7ac\ud558\uac70\ub098, Pointwise Conv\uc640 \uc735\ud569(Fusion)\ud558\ub294 \ubc29\uc2dd\uc73c\ub85c \uc9c4\ud654\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. Hardware engieer\ub85c\uc11c \uc6b0\ub9ac\ub294 \ub2e8\uc21c\ud788 FLOPs \uc218\uce58\uc5d0 \ud604\ud639\ub418\uc9c0 \uc54a\uace0, \uc2e4\uc81c \ud558\ub4dc\uc6e8\uc5b4 \ud30c\uc774\ud504\ub77c\uc778\uc774 \uacaa\uc744 \ubd80\ud558\ub97c \uaff0\ub6ab\uc5b4 \ubcfc \uc218 \uc788\uc5b4\uc57c \ud569\ub2c8\ub2e4.<\/p>\n\n\n\n<p>\ub2e4\uc74c \uae00\uc5d0\uc11c\ub294 ResNet\uc774 \uac00\uc838\uc628 \ub525\ub7ec\ub2dd\uc758 \ud601\uba85 \ub4a4\uc5d0 \uc228\uaca8\uc9c4 \ud558\ub4dc\uc6e8\uc5b4\uc801 \uace8\uce6b\uac70\ub9ac, &#8220;ResNet\uacfc \ubcd1\ubaa9: Skip Connection\uc774 \ud558\ub4dc\uc6e8\uc5b4 \uba54\ubaa8\ub9ac \uad00\ub9ac\uc640 \ubc84\ud37c \uc2a4\ucf00\uc904\ub9c1\uc5d0 \uc8fc\ub294 \ubb38\uc81c&#8221;\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n<style>.kadence-column1331_b578a2-af > .kt-inside-inner-col{box-shadow:0px 0px 14px 0px rgba(0, 0, 0, 0.2);}.kadence-column1331_b578a2-af > .kt-inside-inner-col,.kadence-column1331_b578a2-af > .kt-inside-inner-col:before{border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;}.kadence-column1331_b578a2-af > .kt-inside-inner-col{column-gap:var(--global-kb-gap-sm, 1rem);}.kadence-column1331_b578a2-af > .kt-inside-inner-col{flex-direction:column;}.kadence-column1331_b578a2-af > .kt-inside-inner-col > .aligncenter{width:100%;}.kadence-column1331_b578a2-af > .kt-inside-inner-col:before{opacity:0.3;}.kadence-column1331_b578a2-af{position:relative;}@media all and (max-width: 1024px){.kadence-column1331_b578a2-af > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}@media all and (max-width: 767px){.kadence-column1331_b578a2-af > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}<\/style>\n<div class=\"wp-block-kadence-column kadence-column1331_b578a2-af\"><div class=\"kt-inside-inner-col\">\n<p><strong>\uad00\ub828 \uae00<\/strong><\/p>\n\n\n\n<p>\u2705<a href=\"https:\/\/rtlearner.com\/ai-architecture-1-neuron-hardware-mac-analysis\/\" data-type=\"post\" data-id=\"1248\">AI Architecture 1. \uc778\uacf5 \ub274\ub7f0\uc758 \ud574\ubd80: silicon\uc5d0\uc11c Y=WX+B \uad6c\ud604<\/a><\/p>\n\n\n\n<p>\u2705<a href=\"https:\/\/rtlearner.com\/ai-architecture-2-activation-relu-vs-sigmoid\/\" data-type=\"post\" data-id=\"1255\">AI Architecture 2. \ud65c\uc131\ud654 \ud568\uc218\uc758 \ube44\uc6a9: ReLU vs Sigmoid<\/a><\/p>\n\n\n\n<p>\u2705<a href=\"https:\/\/rtlearner.com\/ai-architecture-3-matmul-simd-parallel-processing\/\" data-type=\"post\" data-id=\"1263\">AI Architecture 3. \ud589\ub82c\uacf1(MatMul)\uc758 \ubbf8\ud559: \ub525\ub7ec\ub2dd\uc774 GPU\/NPU\ub97c \uc120\ud0dd\ud55c \uc774\uc720<\/a><\/p>\n\n\n\n<p>\u2705<a href=\"https:\/\/rtlearner.com\/ai-architecture-4-training-vs-inference\/\" data-type=\"post\" data-id=\"1267\">AI Architecture 4. \ud559\uc2b5(Training) vs \ucd94\ub860(Inference)<\/a><\/p>\n<\/div><\/div>\n\n\n\n<p>\ucc38\uace0: <em><a href=\"https:\/\/arxiv.org\/abs\/1704.04861\" target=\"_blank\" rel=\"noopener\">Efficient Convolutional Neural Networks for Mobile Vision Applications<\/a><\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>In the previous 3 Mappings of Conv Operations, we looked at a strategy to sacrifice memory and gain computational speed (GEMM) through the Im2Col method when processing standard convolutions in hardware.<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_kadence_starter_templates_imported_post":false,"_kad_post_transparent":"","_kad_post_title":"","_kad_post_layout":"","_kad_post_sidebar_id":"","_kad_post_content_style":"","_kad_post_vertical_padding":"","_kad_post_feature":"","_kad_post_feature_position":"","_kad_post_header":false,"_kad_post_footer":false,"_kad_post_classname":"","footnotes":""},"categories":[116],"tags":[117,118],"class_list":["post-1331","post","type-post","status-publish","format-standard","hentry","category-ai-and-hw-fundamentals","tag-ai","tag-architecture"],"_links":{"self":[{"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/posts\/1331","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/comments?post=1331"}],"version-history":[{"count":5,"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/posts\/1331\/revisions"}],"predecessor-version":[{"id":1369,"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/posts\/1331\/revisions\/1369"}],"wp:attachment":[{"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/media?parent=1331"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/categories?post=1331"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/rtlearner.com\/en\/wp-json\/wp\/v2\/tags?post=1331"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}<!-- This website is optimized by Airlift. Learn more: https://airlift.net. Template:. Learn more: https://airlift.net. Template: 69b92da9d36f73cd2808d6e8. Config Timestamp: 2026-03-17 10:32:09 UTC, Cached Timestamp: 2026-05-16 07:58:57 UTC -->