\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8<\/title>\n<style>\nbody{margin:0;background:#fcfcfb;color:#0b0b0b;font-family:\"Noto Sans KR\",\"Apple SD Gothic Neo\",\"Malgun Gothic\",sans-serif;line-height:1.75;font-size:17px}\nmain{max-width:860px;margin:0 auto;padding:40px 20px 80px}\nh1{font-size:1.9em;line-height:1.3;margin:0 0 .6em}\nh2{font-size:1.3em;margin:2em 0 .6em;border-bottom:1px solid #e3e2dc;padding-bottom:.25em}\np{margin:0 0 1.1em}\nimg{display:block;width:100%;height:auto;border:1px solid #e3e2dc;border-radius:6px;margin:1.4em 0 .4em}\n.cap{font-size:.88em;color:#52514e;margin:0 0 1.8em}\nul{padding-left:1.3em} li{margin-bottom:.35em;word-break:break-all}\na{color:#2a78d6;text-decoration:none} a:hover{text-decoration:underline}\n.meta{color:#52514e;font-size:.9em;margin-bottom:2em}\n<\/style><\/head><body><main>\n<h1>\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8<\/h1>\n<p><img decoding=\"async\" alt=\"\ub300\ud45c \uc774\ubbf8\uc9c0\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/649AC5AF-26C9-4062-A34A-EF385763B59C.gif\"><\/p>\n<p class=\"cap\">\ub300\ud45c \uc774\ubbf8\uc9c0: Chiplet \ud328\ud0a4\uc9c0\uc758 Exploded View \uc77c\ub7ec\uc2a4\ud2b8(Higgsfield gpt_image_2_5 \uc0dd\uc131). \uc544\ub798\ubd80\ud130 Package Substrate, Silicon Interposer, \uadf8 \uc704\uc5d0 Compute Chiplet 4\uac1c, I\/O Die 2\uac1c, HBM Stack 8\uac1c\uac00 \ubd84\ub9ac\ub418\uc5b4 \ubc30\uce58\ub428. \ud2b9\uc815 \uc81c\ud488\uc758 Floorplan\uacfc \ubb34\uad00\ud568<\/p>\n<p>Microsoft\ub294 2026\ub144 10\uc6d4 12\uc77c OCP Global Summit\uc5d0\uc11c 3\uc138\ub300 \uc790\uccb4 Inference \uac00\uc18d\uae30 Maia 300\uc744 \uacf5\uac1c\ud558\uc600\ub2e4. 2027\ub144\ubd80\ud130 Azure\uc5d0 \ubc30\uce58\ub420 \uc774 \uc81c\ud488\uc740 Microsoft\uac00 \ucc98\uc74c\uc73c\ub85c Multi-Chiplet \uad6c\uc870\ub97c \ucc44\ud0dd\ud55c \uac00\uc18d\uae30\uc774\ub2e4. AMD\uac00 MI300X \uc774\ub798 MI455X\uc5d0 \uc774\ub974\uae30\uae4c\uc9c0 Compute\ub97c 8\uac1c Die\ub85c \ubd84\ud560\ud574 \uc628 \uac83\uacfc \ub2ec\ub9ac Microsoft\ub294 Compute Die\ub97c \ub2e8\uc77c\ub85c \uc720\uc9c0\ud558\uace0 I\/O\uc640 Memory Controller\ub9cc\uc744 \ubcc4\ub3c4 Die\ub85c \ubd84\ub9ac\ud558\uc600\ub2e4. \ub3d9\uc77c\ud55c Chiplet\uc774\ub77c\ub294 \uba85\uce6d \uc544\ub798 \ubd84\ud560\uc758 \ucd95\uc774 Compute\uc5d0\uc11c Memory\u00b7I\/O\ub85c \uc774\ub3d9\ud55c \uac83\uc774\ub2e4. \uad6d\ub0b4\uc5d0\uc11c\ub294 Rebellions\uac00 2025\ub144 Hot Chips\uc5d0\uc11c Compute Chiplet 4\uac1c\ub97c UCIe\ub85c \uc5f0\uacb0\ud55c REBEL-Quad\ub97c \uacf5\uac1c\ud558\uc600\uc73c\uba70, \uc774 \uc81c\ud488\uc740 AMD, Microsoft\uc640 \uad6c\ubd84\ub418\ub294 \uc81c3\uc758 \ubd84\ud560 \ubc29\uc2dd\uc5d0 \ud574\ub2f9\ud55c\ub2e4.<\/p>\n<h2>1. \uc138 \uac00\uc18d\uae30\uc758 Die \ubd84\ud560 \ubc29\uc2dd<\/h2>\n<p>AMD Instinct MI455X(CDNA 5, 2026\ub144 \ubc1c\ud45c)\ub294 TSMC N2 \uacf5\uc815\uc758 XCD(Accelerator Complex Die) 8\uac1c\ub97c N3 \uacf5\uc815\uc758 FCD(Fabric and Cache Die) 2\uac1c \uc704\uc5d0 Hybrid Bonding\uc73c\ub85c \uc218\uc9c1 \uc801\uce35\ud558\uace0, \uc5ec\uae30\uc5d0 N3 \uacf5\uc815\uc758 I\/O Die 2\uac1c\uc640 HBM4 12 Stack(432GB, 23.3TB\/s)\uc744 TSMC CoWoS-L\ub85c \uacb0\ud569\ud55c \uad6c\uc870\uc774\ub2e4. XCD \ud558\ub098\ub294 32\uac1c WGP(Workgroup Processor)\ub97c \ud3ec\ud568\ud558\uba70, FCD \ud558\ub098\ub294 96MB L2\uc640 HBM PHY, Infinity Fabric\uc744 \uc218\uc6a9\ud558\uc5ec XCD 4\uac1c\ub97c \ub2f4\ub2f9\ud55c\ub2e4. I\/O Die\ub294 \ud638\uc2a4\ud2b8 \uc5f0\uacb0, Scale-up, Scale-out NIC\ub97c \ub9e1\ub294\ub2e4. \uc804\uccb4 Transistor \uc218\ub294 3,200\uc5b5 \uac1c\uc774\uace0, Peak \uc5f0\uc0b0\uc740 MXFP4 40.3 PFLOPS, MXFP8 20.1 PFLOPS\uc774\ub2e4. \uc804 \uc138\ub300 MI355X(2025)\uac00 XCD 8\uac1c\ub97c IOD 2\uac1c \uc704\uc5d0 \uc801\uce35\ud558\uace0 Memory-side Infinity Cache 256MB\ub97c \ub450\uc5c8\ub358 \uac83\uacfc \ube44\uad50\ud558\uba74, MI455X\ub294 Infinity Cache\ub97c \uc81c\uac70\ud558\uace0 FCD\uc758 L2 192MB(96MB \u00d7 2)\ub85c \ub300\uccb4\ud558\uc600\uc73c\uba70 I\/O \uae30\ub2a5\uc744 \ubcc4\ub3c4 Die\ub85c \ubd84\ub9ac\ud558\uc600\ub2e4. \uc989 AMD\ub294 XCD 8\uac1c \uad6c\uc870\ub97c \uc138 \uc138\ub300\uc5d0 \uac78\uccd0 \uc720\uc9c0\ud558\uba74\uc11c \ud558\ubd80 Die\uc758 \uc5ed\ud560\uc744 Cache \uc911\uc2ec\uc73c\ub85c \uc7ac\ud3b8\ud558\uc600\ub2e4.<\/p>\n<p>Maia 300\uc740 TSMC N2P Compute Die 1\uac1c, I\/O Die 2\uac1c, Memory Die 4\uac1c\ub85c \uad6c\uc131\ub418\uba70 HBM4 8 Stack(288GB, 18TB\/s)\uc744 \ud0d1\uc7ac\ud55c\ub2e4. Compute Die \ub0b4\ubd80\ub294 12\uac1c Tile\ub85c \uad6c\uc131\ub41c Cluster 4\uac1c\uc640 \uc0c1\uc704 Network-on-Chip\uc73c\ub85c \uc774\ub8e8\uc5b4\uc9c0\uba70, \uac01 Tile\uc740 Tensor Unit, Vector Processor, DMA Engine, Local SRAM\uc744 \ud3ec\ud568\ud55c\ub2e4. \uc804 \uc138\ub300 Maia 200(2026\ub144 1\uc6d4 \ubc1c\ud45c)\uc740 N3P Monolithic Die\uc5d0 HBM3E 6 Stack(216GB, 7.1TB\/s)\uc744 \uacb0\ud569\ud55c \uad6c\uc870\uc600\uc73c\ubbc0\ub85c, Microsoft\ub294 \ud55c \uc138\ub300 \ub9cc\uc5d0 Monolithic\uc5d0\uc11c Chiplet\uc73c\ub85c \uc774\ud589\ud55c \uc148\uc774\ub2e4.<\/p>\n<p><img decoding=\"async\" alt=\"Azure Maia 300 \ud504\ub85c\ud1a0\ud0c0\uc785 \uce69\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/maia300_photo_microsoft.png\"><\/p>\n<p class=\"cap\">\uc0ac\uc9c4: Microsoft, Azure Maia 300 \ud504\ub85c\ud1a0\ud0c0\uc785 \uce69(OCP Global Summit 2026 \uacf5\uac1c \uc790\ub8cc). \ud328\ud0a4\uc9c0 \uc911\uc559 Die \ud45c\uba74\uc5d0 4\uac1c Cluster \uaca9\uc790\uac00 \uc2dd\ubcc4\ub428<\/p>\n<p>Rebellions Rebel100(\uac1c\ubc1c\uba85 REBEL-Quad)\uc740 Samsung SF4X \uacf5\uc815\uc758 320mm\u00b2 Compute Chiplet 4\uac1c\ub97c UCIe-Advanced\ub85c \uc5f0\uacb0\ud55c System-in-Package\uc774\ub2e4. Chiplet\ub9c8\ub2e4 HBM3E 36GB 1 Stack\uc774 \uadc0\uc18d\ub418\uc5b4 \ud328\ud0a4\uc9c0 \uc804\uccb4\ub85c 144GB\ub97c \uad6c\uc131\ud558\uba70, \ud328\ud0a4\uc9d5\uc740 TSMC CoWoS-S\uc774\uace0 \uc804\ub825 \uacfc\ub3c4 \uc751\ub2f5 \uc5b5\uc81c\ub97c \uc704\ud55c Integrated Silicon Capacitor(ISC) Die 4\uac1c\uac00 \ud568\uaed8 \ud0d1\uc7ac\ub41c\ub2e4. UCIe \ub9c1\ud06c\ub294 Lane\ub2f9 16Gbps, \uc9d1\uacc4 \uc57d 4TB\/s, Chiplet \uac04 \uc9c0\uc5f0 \uc57d 11ns\ub85c \ubc1c\ud45c\ub418\uc5c8\uc73c\uba70, 2D Mesh Network-on-Chip\uc774 UCIe\ub97c \uacbd\uc720\ud558\uc5ec 4\uac1c Chiplet\uc5d0 \uac78\uccd0 \ud655\uc7a5\ub418\ubbc0\ub85c \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uad00\uc810\uc5d0\uc11c\ub294 \ub2e8\uc77c \ud504\ub85c\uc138\uc11c\ub85c \uc778\uc2dd\ub41c\ub2e4\ub294 \uac83\uc774 Rebellions\uc758 \uc124\uba85\uc774\ub2e4. FP8 2 PFLOPS, TDP 600W\uc774\ub2e4.<\/p>\n<p>\uc138 \uad6c\uc870\ub97c \ubcd1\uce58\ud558\uba74 \ubd84\ud560 \ubc29\uc2dd\uc740 \uc138 \uac08\ub798\ub85c \uad6c\ubd84\ub41c\ub2e4. AMD\ub294 Compute(XCD), Cache\u00b7Fabric(FCD), I\/O(I\/O Die)\ub97c \uc0c1\uc774\ud55c \uacf5\uc815\uc758 Die\ub85c \ubd84\ub9ac\ud55c \uc774\uc885 \ubd84\ud560, Microsoft\ub294 Compute\ub97c \ub2e8\uc77c\ub85c \uc720\uc9c0\ud558\uace0 I\/O\uc640 Memory Controller\ub9cc \ubd84\ub9ac\ud55c \ubd84\ud560, Rebellions\ub294 IOD \uc5c6\uc774 \ub3d9\uc77c Compute Chiplet\uc744 \ubcf5\uc81c\ud558\uc5ec \uc9c1\uc811 \uc5f0\uacb0\ud55c \ub3d9\uc885 \ubd84\ud560\uc5d0 \ud574\ub2f9\ud55c\ub2e4.<\/p>\n<p><img decoding=\"async\" alt=\"\uadf8\ub9bc 1\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/IMG_1124-scaled.png\"><\/p>\n<p class=\"cap\">\uadf8\ub9bc 1. AMD MI455X(XCD 8\uac1c, FCD 2\uac1c, IOD 2\uac1c)\uc640 Microsoft Maia 300(Compute Die 1\uac1c, Memory Die 4\uac1c, I\/O Die 2\uac1c)\uc758 Chiplet \uad6c\uc870 \ube44\uad50. Rebellions Rebel100(\ub3d9\uc885 Compute Chiplet 4\uac1c, UCIe \uc5f0\uacb0)\uc758 \uad6c\uc870\ub294 1\uc808 \uc11c\uc220 \ucc38\uc870. \ud654\uc0b4\ud45c\ub294 LLM Inference \uc2dc \ub370\uc774\ud130 \uc774\ub3d9\uc744 \ub098\ud0c0\ub0b4\ub294 \uac1c\ub150\ub3c4\uc784 <\/p>\n<h2>2. \uc801\uce35 \ubc29\uc2dd\uc758 \ube44\uad50<\/h2>\n<p>AMD\ub294 XCD\ub97c FCD \uc704\uc5d0 \uc801\uce35\ud558\ub294 3D \uad6c\uc870\uc640 HBM\uc744 \uce21\uba74\uc5d0 \ubc30\uce58\ud558\ub294 2.5D CoWoS-L\uc744 \ubcd1\uc6a9\ud55c\ub2e4. XCD\uac00 FCD \uc704\uc5d0 \uc801\uce35\ub418\ub294 \uc774 \uad6c\uc870\uc5d0\uc11c\ub294 XCD\uc640 L2 \uc0ac\uc774\uc758 \uc218\uc9c1 \uc5f0\uacb0\uc774 \uc9e7\uc740 \ub300\uc2e0 \uc5f4 \ubc00\ub3c4\uac00 \uc0c1\uc2b9\ud558\uace0, \ub450 FCD \uc0ac\uc774\uc758 \ud1b5\uc2e0\uc740 Interposer\ub97c \uacbd\uc720\ud558\ub294 \uc57d 14TB\/s\uc758 Die-to-Die \ub9c1\ud06c\uc5d0 \uc758\uc874\ud55c\ub2e4. CoWoS-L\uc740 Interposer\ub97c \uc791\uc740 Local Silicon Interconnect \uc870\uac01\uacfc \uc720\uae30 \uc7ac\ubc30\uc120\uce35\uc73c\ub85c \uad6c\uc131\ud558\uc5ec \ub2e8\uc77c Silicon Interposer\uc758 Reticle \ud55c\uacc4\ub97c \ub118\ub294 \ud328\ud0a4\uc9c0 \uba74\uc801\uc744 \ud655\ubcf4\ud558\ub294 \ubc29\uc2dd\uc774\uba70, HBM4 12 Stack\uc744 \uc218\uc6a9\ud558\uae30 \uc704\ud55c \uc120\ud0dd\uc73c\ub85c \ud574\uc11d\ub41c\ub2e4. Rebellions\ub294 Compute Chiplet 4\uac1c\uc640 HBM\uc744 \ubaa8\ub450 CoWoS-S Interposer \uc704 \ub2e8\uc77c \uce35\uc5d0 \ubc30\uce58\ud558\uace0 Chiplet \uac04 \uc5f0\uacb0\uc744 Interposer \ubc30\uc120\uc758 UCIe-Advanced\ub85c \ucc98\ub9ac\ud558\ubbc0\ub85c, 3D \uc801\uce35 \uc5c6\uc774 2.5D\ub9cc\uc73c\ub85c Compute\ub97c \ubd84\ud560\ud55c \uc0ac\ub840\uc774\ub2e4. Maia 300\uc740 \uacf5\uac1c \uc790\ub8cc \uae30\uc900\uc73c\ub85c 3D \uc801\uce35 \uc5b8\uae09\uc774 \uc5c6\uc73c\uba70, Compute\u00b7I\/O\u00b7Memory Die\uac00 Interposer \uc704 \ub2e8\uc77c \uce35\uc5d0 \ub193\uc774\ub294 2.5D \uad6c\uc870\ub85c \ud30c\uc545\ub41c\ub2e4. Microsoft\ub294 “HBM Stack\uc744 Compute\uc5d0 \uadfc\uc811 \ubc30\uce58\ud558\uae30 \uc704\ud55c Advanced Packaging”\uc73c\ub85c\ub9cc \uc124\uba85\ud558\ubbc0\ub85c \uc138\ubd80 \uacf5\uc815\uc740 \ucd94\uac00 \uacf5\uac1c\ub97c \uc694\ud55c\ub2e4.<\/p>\n<p><img decoding=\"async\" alt=\"\uadf8\ub9bc 2\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/IMG_1125-scaled.png\"><\/p>\n<p class=\"cap\">\uadf8\ub9bc 2. \ub2e8\uba74 \ube44\uad50. AMD MI455X\ub294 XCD\ub97c FCD \uc704\uc5d0 \uc218\uc9c1 \uc801\uce35\ud558\uace0, Maia 300\uc740 \uacf5\uac1c \uc815\ubcf4 \uae30\uc900\uc73c\ub85c \ubaa8\ub4e0 Die\ub97c \ub3d9\uc77c \uce35\uc5d0 \ubc30\uce58\ud568<\/p>\n<h2>3. \uc218\uce58\ub85c \ubcf8 \uc124\uacc4 \ubc29\ud5a5<\/h2>\n<p>\uc138\ub300\ubcc4 \uc0ac\uc591\uc744 \ub3d9\uc77c \ucd95\uc5d0 \ubc30\uce58\ud558\uba74 \ub450 \ud68c\uc0ac\uc758 \ubc29\ud5a5\uc774 \uad6c\ubd84\ub41c\ub2e4. Microsoft\ub294 Maia 200\uc5d0\uc11c 300\uc73c\ub85c \uc774\ud589\ud558\uba70 FP8 \uc5f0\uc0b0\uc744 5.1\uc5d0\uc11c 6.9 PFLOPS\ub85c 1.35\ubc30 \uc99d\uac00\uc2dc\ud0a8 \ubc18\uba74 HBM \ub300\uc5ed\ud3ed\uc740 7.1\uc5d0\uc11c 18TB\/s\ub85c 2.5\ubc30 \uc99d\uac00\uc2dc\ucf30\ub2e4. AMD\ub294 MI355X\uc5d0\uc11c MI455X\ub85c \uc774\ud589\ud558\uba70 FP8 \uc5f0\uc0b0\uc744 5.0\uc5d0\uc11c \uc57d 20 PFLOPS\ub85c 4\ubc30, \ub300\uc5ed\ud3ed\uc744 8.0\uc5d0\uc11c 23.3TB\/s\ub85c 2.9\ubc30 \uc99d\uac00\uc2dc\ucf30\ub2e4. \uc5f0\uc0b0 1 PFLOPS\ub2f9 \ub300\uc5ed\ud3ed\uc73c\ub85c \ud658\uc0b0\ud558\uba74 Maia 300\uc740 2.61TB\/s, MI455X\ub294 1.17TB\/s\ub85c \uc0b0\uc815\ub41c\ub2e4. Maia 300\uc740 \ub300\uc5ed\ud3ed \uce21\uc73c\ub85c, MI455X\ub294 \uc5f0\uc0b0 \uce21\uc73c\ub85c \uade0\ud615\uc744 \uc774\ub3d9\uc2dc\ud0a8 \uac83\uc774\ub2e4.<\/p>\n<p><img decoding=\"async\" alt=\"\uadf8\ub9bc 3\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/IMG_1127-scaled.png\"><\/p>\n<p class=\"cap\">\uadf8\ub9bc 3. \uc138\ub300\ubcc4 \uc0ac\uc591 \ucd94\uc774. MI455X\uc758 FP8 \uc218\uce58\ub294 \ubc1c\ud45c\ub41c FP4 40 PFLOPS\uc758 \uc808\ubc18\uc73c\ub85c \uc0b0\uc815\ud568<\/p>\n<p>\uc774 \ucc28\uc774\ub294 \ub450 \uc81c\ud488\uc758 \uc5ed\ud560\uacfc \uc77c\uce58\ud55c\ub2e4. Maia 300\uc740 Microsoft \uc790\uc0ac \uc11c\ube44\uc2a4\uc758 Inference \uc804\uc6a9 \uac00\uc18d\uae30\uc774\uba70, Decode \ub2e8\uacc4\uc758 \ubcd1\ubaa9\uc740 \uc5f0\uc0b0\ubcf4\ub2e4 KV cache\uc640 Weight \uc801\uc7ac \ub300\uc5ed\ud3ed\uc5d0 \uc788\ub2e4. MI455X\ub294 \ud559\uc2b5\uacfc Inference\ub97c \ubaa8\ub450 \ub300\uc0c1\uc73c\ub85c \ud558\ub294 \ubc94\uc6a9 GPU\uc774\ubbc0\ub85c \uc5f0\uc0b0 \ubc00\ub3c4\uc758 \uc720\uc9c0\uac00 \uc694\uad6c\ub41c\ub2e4. \ub530\ub77c\uc11c Maia 300\uc758 Chiplet \ubd84\ud560\uc740 HBM4 8 Stack\uc744 \uc218\uc6a9\ud560 Beachfront(Die \uac00\uc7a5\uc790\ub9ac \uae38\uc774)\uc758 \ud655\ubcf4\uac00 \ubaa9\uc801\uc774\uace0, AMD\uc758 XCD \ubd84\ud560\uc740 Reticle \ud55c\uacc4\uc640 \uc218\uc728 \ubb38\uc81c\uc758 \ud574\uc18c\uac00 \ubaa9\uc801\uc774\ub2e4. \ubd84\ud560\uc758 \ub3d9\uae30\uac00 \uc0c1\uc774\ud558\ubbc0\ub85c \ubd84\ud560\uc758 \ucd95 \ub610\ud55c \uc0c1\uc774\ud558\ub2e4.<\/p>\n<h2>4. \uad6c\uc870\uc640 \ubb34\uad00\ud558\uac8c \uc794\uc874\ud558\ub294 NUMA<\/h2>\n<p>Compute\ub97c \ub2e8\uc77c Die\ub85c \uc720\uc9c0\ud558\ub354\ub77c\ub3c4 \uba54\ubaa8\ub9ac \uc811\uadfc\uc774 \uade0\uc77c\ud574\uc9c0\ub294 \uac83\uc740 \uc544\ub2c8\ub2e4. Microsoft\ub294 Maia 300\uc758 \uba54\ubaa8\ub9ac \ubaa8\ub378\uc744 \uba85\uc2dc\uc801 NUMA\ub85c \uaddc\uc815\ud558\uace0, \ube48\ubc88\ud788 \uc811\uadfc\ud558\ub294 \ub370\uc774\ud130\ub97c \ud574\ub2f9 Compute \uadfc\ucc98\uc5d0 \ubc30\uce58\ud558\ub294 \ucc45\uc784\uc744 Software-Defined Dataflow\uc5d0 \ubd80\uc5ec\ud55c\ub2e4. Memory Die 4\uac1c\uac00 \uac01\uac01 HBM4 2 Stack\uc744 \ub2f4\ub2f9\ud558\ub294 \uad6c\uc870\ub77c\uba74 Compute Die\uc758 Cluster\uc640 Memory Die \uc0ac\uc774\uc758 \uac70\ub9ac\uc5d0 \ub530\ub77c \uc9c0\uc5f0\uc774 \uc0c1\uc774\ud574\uc9c0\ub294 \uac83\uc740 \uad6c\uc870\uc0c1 \ud544\uc5f0\uc774\ub2e4.<\/p>\n<p>Rebel100\uc5d0\uc11c\ub294 NUMA\uac00 \uad6c\uc870\uc801\uc73c\ub85c \ub354\uc6b1 \ub69c\ub837\ud558\ub2e4. HBM\uc774 Chiplet\ubcc4\ub85c 1 Stack\uc529 \uadc0\uc18d\ub418\ubbc0\ub85c \ud0c0 Chiplet\uc5d0 \uacb0\ud569\ub41c HBM \uc811\uadfc\uc740 \ubc18\ub4dc\uc2dc UCIe \ub9c1\ud06c\ub97c \uacbd\uc720\ud558\uba70, \uc9d1\uacc4 4TB\/s\uc758 UCIe \ub300\uc5ed\ud3ed\uc744 4\uac1c Chiplet\uc774 \ubd84\ud560\ud558\uc5ec \uc0ac\uc6a9\ud55c\ub2e4. “\ub2e8\uc77c \ud504\ub85c\uc138\uc11c\ub85c \uc778\uc2dd\ub41c\ub2e4”\ub294 \uc124\uba85\uc740 \uc8fc\uc18c \uacf5\uac04\uacfc NoC \ud1a0\ud3f4\ub85c\uc9c0 \uad00\uc810\uc758 \uc11c\uc220\uc774\uace0, \uc9c0\uc5f0\uacfc \ub300\uc5ed\ud3ed\uc758 \ube44\uade0\uc77c\uc131\uc740 \uc794\uc874\ud55c\ub2e4. \ub530\ub77c\uc11c KV cache\uc640 Weight\ub97c \uac01 Chiplet\uc758 Local HBM\uc5d0 \ubd84\ud560 \ubc30\uce58\ud558\uace0 Chiplet \uacbd\uacc4\ub97c \ub118\ub294 \uc811\uadfc\uc744 \ucd5c\uc18c\ud654\ud558\ub294 \uac83\uc774 \uc131\ub2a5\uc758 \uc804\uc81c\uac00 \ub41c\ub2e4.<\/p>\n<p>AMD\uc5d0\uc11c\ub294 \uc774 \ubb38\uc81c\uac00 \ud558\ub4dc\uc6e8\uc5b4 \uc77c\uad00\uc131 \uc218\uc900\uc5d0\uc11c \ub4dc\ub7ec\ub098\uba70, \uc138\ub300\uc5d0 \ub530\ub77c \uadf8 \ub2e8\uc704\uac00 \ubcc0\uacbd\ub418\uc5c8\ub2e4. MI300X\uc640 MI355X\uc5d0\uc11c\ub294 XCD\ub9c8\ub2e4 L2\uac00 \ub3c5\ub9bd\uc801\uc73c\ub85c \uc874\uc7ac\ud558\uace0 XCD \uac04 L2\uac00 \ube44\uc77c\uad00\uc131(Non-Coherent)\uc774\uc5c8\uc73c\ubbc0\ub85c, \ud558\ub098\uc758 \uc791\uc5c5\uc744 \ud558\ub098\uc758 XCD \ub0b4\ubd80\uc5d0 \ud55c\uc815\ud558\uba74(XCD Confinement) Cache Maintenance \uc5c6\uc774 L2\ub97c \uc7ac\uc0ac\uc6a9\ud560 \uc218 \uc788\uc5c8\uc73c\ub098 XCD \uacbd\uacc4\ub97c \ub118\ub294 \uc21c\uac04 IOD \uacbd\uc720 \ube44\uc6a9\uacfc \uc77c\uad00\uc131 \ube44\uc6a9\uc774 \ub3d9\uc2dc\uc5d0 \ubc1c\uc0dd\ud558\uc600\ub2e4. MI455X\ub294 L2\ub97c XCD\uc5d0\uc11c FCD\ub85c \uc774\uc804\ud558\uc5ec XCD 4\uac1c\uac00 96MB L2 \ud558\ub098\ub97c \uacf5\uc720\ud558\ub3c4\ub85d \ud558\uc600\uace0, \uc774\ub85c\uc368 XCD \uac04 Atomic \uc5f0\uc0b0\uc774 \uac1c\uc120\ub418\uace0 MI300 \uacc4\uc5f4\uc5d0\uc11c \uc804\uc5ed \uac00\uc2dc\uc131 \ud655\ubcf4\uc5d0 \ud544\uc694\ud558\uc600\ub358 Kernel \uacbd\uacc4\uc758 Cache Flush\uac00 \uc81c\uac70\ub418\uc5c8\ub2e4. \ub2e4\ub9cc \ud55c FCD\uc758 WGP\ub294 \ub2e4\ub978 FCD\uc758 L2\uc5d0 \uc811\uadfc\ud560 \uc218 \uc5c6\uc73c\ubbc0\ub85c, \uc77c\uad00\uc131 \ub2e8\uc704\uac00 XCD 8\uac1c\uc5d0\uc11c FCD 2\uac1c\ub85c \ud655\ub300\ub418\uc5c8\uc744 \ubfd0 FCD \uacbd\uacc4\uc758 NUMA\ub294 \uc794\uc874\ud55c\ub2e4. \uac01 L2\uac00 \uc804\uc5ed \uc8fc\uc18c \uc804\uccb4\ub97c Cache\ud560 \uc218 \uc788\uc5b4 Die-to-Die \ud2b8\ub798\ud53d\uc774 \uc644\ud654\ub418\uc9c0\ub9cc, \uc791\uc5c5\uc744 FCD \ub2e8\uc704\ub85c \ubc30\uce58\ud558\ub294 \uac83\uc774 \uc5ec\uc804\ud788 \uc720\ub9ac\ud558\ub2e4. Maia 300\uc774 NUMA\ub97c \uc18c\ud504\ud2b8\uc6e8\uc5b4\uc5d0 \ub178\ucd9c\uc2dc\ud0a8 \uacb0\uc815\uc740, \ud558\ub4dc\uc6e8\uc5b4\uac00 Locality \ub2e8\uc704\ub97c \uc740\ud3d0\ud558\uc9c0 \uc54a\uace0 \ub4dc\ub7ec\ub0b4\ub294 \ud3b8\uc774 Inference \uac00\uc18d\uae30\uc5d0\uc11c\ub3c4 \uc720\ub9ac\ud558\ub2e4\ub294 \ud310\ub2e8\uc73c\ub85c \ud574\uc11d\ub41c\ub2e4. “\uc791\uc5c5\uc744 \ud558\ub4dc\uc6e8\uc5b4 Locality \ub2e8\uc704\uc5d0 \ud55c\uc815\ud558\uace0 \ub2e8\uc704 \uac04 \ud1b5\uc2e0\uc744 \uba85\uc2dc\uc801\uc73c\ub85c \uad00\ub9ac\ud55c\ub2e4”\ub294 \uc6d0\uce59\uc774 \uc774\uc885 Compute \ubd84\ud560\ud615(AMD), Memory\u00b7I\/O \ubd84\ud560\ud615(Microsoft), \ub3d9\uc885 Compute \ubd84\ud560\ud615(Rebellions) \uc138 \uad6c\uc870 \ubaa8\ub450\uc5d0 \uc801\uc6a9\ub418\ub294 \uac83\uc774\ub2e4.<\/p>\n<h2>5. \uc804\ub825\uacfc Rack \uc218\uc900\uc758 \ud568\uc758<\/h2>\n<p>Maia 300\uc758 TDP\ub294 950W\ub85c Maia 200(750W) \ub300\ube44 27% \uc0c1\uc2b9\ud558\uc600\uc73c\uba70, Rack\ub2f9 96\uac1c\ub97c 167kW\uc5d0 \uc218\uc6a9\ud55c\ub2e4. MI355X\ub294 1,400W\uc774\uace0 MI455X\ub294 \uacf5\uc2dd TDP\uac00 \ubbf8\uacf5\uac1c\uc774\ub098 Helios Rack \uc804\ub825\uc73c\ub85c\ubd80\ud130 2kW \uc774\uc0c1\uc73c\ub85c \ucd94\uc815\ub41c\ub2e4. \ub300\uc5ed\ud3ed \uc911\uc2ec \uc124\uacc4\uac00 \uc5f0\uc0b0 \uc911\uc2ec \uc124\uacc4\ubcf4\ub2e4 \uc804\ub825 \uc99d\uac00 \ud3ed\uc774 \uc791\ub2e4\ub294 \uc810\uc774 \uc218\uce58\ub85c \ud655\uc778\ub41c\ub2e4. GPU Cluster \uc6b4\uc601 \uad00\uc810\uc5d0\uc11c \ub450 \uac00\uc9c0 \ud568\uc758\uac00 \ub3c4\ucd9c\ub41c\ub2e4. \uccab\uc9f8, \uc790\uc6d0 \uad00\ub9ac \ub2e8\uc704\uac00 “GPU \ud55c \uc7a5”\uc5d0\uc11c Die \ub610\ub294 NUMA domain\uc73c\ub85c \uc138\ubd84\ud654\ub420 \uac00\ub2a5\uc131\uc758 \uc99d\ub300\uc774\ub2e4. MI300 \uacc4\uc5f4\uc758 DPX\u00b7QPX Partition \ubaa8\ub4dc\ub294 \uadf8 \ucd08\uae30 \ud615\ud0dc\uc774\uba70, MI455X\uc758 FCD 2\uac1c \uad6c\uc870, Maia 300\uc758 Cluster\u00b7Memory Die \uad6c\uc870, Rebel100\uc758 Chiplet \ub2e8\uc704 \uad6c\uc870 \ub610\ud55c \ub3d9\uc77c \ubc29\ud5a5\uc758 \ubd84\ud560\uc744 \ud5c8\uc6a9\ud560 \uc5ec\uc9c0\uac00 \uc788\ub2e4. \ub458\uc9f8, Inference \uc804\uc6a9 \uac00\uc18d\uae30\uc640 \ubc94\uc6a9 GPU\uc758 \uc0ac\uc591 \uaca9\ucc28\uac00 \ub300\uc5ed\ud3ed\/\uc5f0\uc0b0 \ube44\ub85c \uc815\ub7c9\ud654\ub418\ubbc0\ub85c, \uc6cc\ud06c\ub85c\ub4dc\ubcc4 \ubc30\uce58 \uae30\uc900\uc744 \uc774 \ube44\uc728\ub85c \uc218\ub9bd\ud558\ub294 \uac83\uc774 \uac00\ub2a5\ud574\uc9c4\ub2e4.<\/p>\n<h2>6. \uad6d\ub0b4 \uc0ac\ub840: \ub450 \uac08\ub798 \ubd84\ud560 \ubc29\uc2dd\uc758 \uad6d\ub0b4 \uc7ac\ud604<\/h2>\n<p>\uad6d\ub0b4 \ub450 \ud68c\uc0ac\ub294 Chiplet \ubd84\ud560\uc758 \uc0c1\uc774\ud55c \uac08\ub798\ub97c \uac01\uac01 \ucc44\ud0dd\ud558\uc600\ub2e4. Rebellions\ub294 \uc804\uc220\ud55c \ubc14\uc640 \uac19\uc774 Compute\ub97c \ub3d9\uc885 Chiplet 4\uac1c\ub85c \ubd84\ud560\ud558\ub294 \ubc29\uc2dd\uc774\ub2e4. FuriosaAI\uc758 \ud604\ud589 \uc81c\ud488 RNGD\ub294 TSMC 5nm \ub2e8\uc77c Die\uc5d0 HBM3 48GB(1.5TB\/s), SRAM 256MB, FP8 512 TFLOPS, TDP 180W\uc758 Monolithic \uc124\uacc4\uc774\uba70, 2026\ub144 5\uc6d4 Broadcom\uacfc \ubc1c\ud45c\ud55c 3\uc138\ub300 \uac00\uc18d\uae30\uc5d0\uc11c 2nm Compute Die\uc640 Scale-up Networking\uc6a9 \ubcc4\ub3c4 I\/O Die\ub97c Broadcom\uc758 Advanced Packaging\uc73c\ub85c \ud1b5\ud569\ud558\uace0 HBM4\/4E\ub97c \ucc44\ud0dd\ud560 \uc608\uc815\uc774\ub2e4(2028\ub144 \uc0c1\ubc18\uae30 \uc0d8\ud50c\ub9c1). \uc989 FuriosaAI\uc758 \ucc28\uc138\ub300\ub294 Maia 300\uacfc \ub3d9\uc77c\ud55c Compute \ud1b5\ud569\u00b7I\/O \ubd84\ub9ac\ud615\uc5d0 \ud574\ub2f9\ud55c\ub2e4. \uad6d\ub0b4\uc5d0\uc11c\ub3c4 Compute \ubd84\ud560\ud615\uacfc I\/O \ubd84\ub9ac\ud615\uc774 \ubcd1\uc874\ud558\uac8c \ub418\uba70, \uc5b4\ub290 \ubc29\uc2dd\uc774\ub4e0 Locality \uad00\ub9ac \ucc45\uc784\uc774 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc2a4\ud0dd\uc73c\ub85c \uc774\uc804\ub41c\ub2e4\ub294 \uc810\uc740 \ub3d9\uc77c\ud558\ub2e4.<\/p>\n<h2>7. \uacb0\ub860<\/h2>\n<p>Maia 300, AMD Instinct, Rebellions Rebel100\uc740 \ubaa8\ub450 Chiplet\uc774\ub098 \ubd84\ud560\uc758 \ucd95\uc774 \uc0c1\uc774\ud558\ub2e4. AMD\ub294 Compute\ub97c 8\uac1c XCD\ub85c \ubd84\ud560\ud558\uc5ec \uc218\uc728\uacfc \uba74\uc801 \ubb38\uc81c\ub97c \ud574\uc18c\ud558\uc600\uace0, MI455X\uc5d0\uc11c L2\ub97c FCD\ub85c \uc774\uc804\ud558\uc5ec \uc77c\uad00\uc131 \ub2e8\uc704\ub97c XCD 8\uac1c\uc5d0\uc11c FCD 2\uac1c\ub85c \ud655\ub300\ud558\uc600\uc73c\ub098 FCD \uacbd\uacc4\uc758 NUMA\ub294 \uc18c\ud504\ud2b8\uc6e8\uc5b4\uac00 \uad00\ub9ac\ud558\uc5ec\uc57c \ud55c\ub2e4. Microsoft\ub294 Compute Die\ub97c \ub2e8\uc77c\ub85c \uc720\uc9c0\ud558\uace0 Memory\u00b7I\/O\ub97c \ubd84\ub9ac\ud558\uc5ec HBM4 \ub300\uc5ed\ud3ed\uc744 2.5\ubc30 \ud655\ub300\ud558\uc600\uc73c\uba70, NUMA\ub97c \uc124\uacc4 \ucd08\uae30\ubd80\ud130 \uba85\uc2dc\uc801\uc73c\ub85c \ub178\ucd9c\ud558\uc600\ub2e4. Rebellions\ub294 IOD \uc5c6\uc774 \ub3d9\uc885 Compute Chiplet\uc744 UCIe\ub85c \uc9c1\uc811 \uc5f0\uacb0\ud558\uc5ec \ud328\ud0a4\uc9c0 \uc804\uccb4\ub97c \ub2e8\uc77c Mesh\ub85c \uacb0\ud569\ud558\uc600\uace0, \uadf8 \ub300\uc2e0 Chiplet\ubcc4\ub85c \uadc0\uc18d\ub41c HBM \uac04\uc758 \ube44\uade0\uc77c \uc811\uadfc\uc744 \uc18c\ud504\ud2b8\uc6e8\uc5b4\uac00 \uad00\ub9ac\ud558\uc5ec\uc57c \ud55c\ub2e4. \ud5a5\ud6c4 \uc5b4\ub290 \ubc29\uc2dd\uc774 \uc8fc\ub958\uac00 \ub420\uc9c0\ub294 Inference \ube44\uc911\uc774 \uacb0\uc815\ud560 \uac83\uc774\ub098, \uc138 \uad6c\uc870 \ubaa8\ub450\uc5d0\uc11c Locality\ub97c \uc18c\ud504\ud2b8\uc6e8\uc5b4\uac00 \uad00\ub9ac\ud558\uc5ec\uc57c \ud55c\ub2e4\ub294 \uc810\uc740 \uc774\ubbf8 \ud655\uc815\ub41c \uc0ac\uc2e4\uc774\ub2e4.<\/p>\n<h2>\uc2e4\ubb3c \uc0ac\uc9c4 \ucd9c\ucc98<\/h2>\n<p>\ubcf8\ubb38 \ub3c4\ud574\ub294 \ubaa8\ub450 \uacf5\uac1c \uc0ac\uc591\uc744 \ubc14\ud0d5\uc73c\ub85c \uc9c1\uc811 \uc791\uc131\ud55c \uac83\uc774\uba70, \uc2e4\ubb3c \uc0ac\uc9c4\uc740 \uc544\ub798 \ucd9c\ucc98\ub97c \ucc38\uc870\ud568.<\/p>\n<p>Microsoft Maia 300 \uacf5\uc2dd \uc774\ubbf8\uc9c0(Microsoft \uc81c\uacf5, news.microsoft.com\/maia-2026):<\/p>\n<ul>\n<li>\ud504\ub85c\ud1a0\ud0c0\uc785 \uce69 hero \uc774\ubbf8\uc9c0: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-prototype-chip-00_hero.png\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-prototype-chip-00_hero.png<\/a><\/li>\n<li>\uce69 render: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Azure-Maia-300-chip-render.png\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Azure-Maia-300-chip-render.png<\/a><\/li>\n<li>\uce69\u00b7\uc11c\ubc84\u00b7rack render: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Azure-Maia-300-Chip-to-Server-to-Rack-Render.png\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Azure-Maia-300-Chip-to-Server-to-Rack-Render.png<\/a><\/li>\n<li>\ubcf4\ub4dc \uc0ac\uc9c4: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-board-1.jpg\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-board-1.jpg<\/a><\/li>\n<li>\uc11c\ubc84 \uc0ac\uc9c4: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-server-1.jpg\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-server-1.jpg<\/a><\/li>\n<li>\uacf5\uc2dd \ubc1c\ud45c \ube14\ub85c\uadf8: <a href=\"https:\/\/blogs.microsoft.com\/blog\/2026\/10\/12\/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments\/\">https:\/\/blogs.microsoft.com\/blog\/2026\/10\/12\/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments\/<\/a><\/li>\n<\/ul>\n<p>Rebellions Rebel100 \/ REBEL-Quad:<\/p>\n<ul>\n<li>ServeTheHome Hot Chips 2025 \uae30\uc0ac(\ud328\ud0a4\uc9c0\u00b7\ubcf4\ub4dc \uc0ac\uc9c4 \uc218\ub85d): <a href=\"https:\/\/www.servethehome.com\/rebellions-rebel-quad-ucie-and-144gb-hbm3e-accelerator-at-hot-chips-2025\/\">https:\/\/www.servethehome.com\/rebellions-rebel-quad-ucie-and-144gb-hbm3e-accelerator-at-hot-chips-2025\/<\/a><\/li>\n<li>Rebellions \uacf5\uc2dd \ub274\uc2a4\ub8f8: <a href=\"https:\/\/rebellions.ai\/newsroom\/rebellions-debuts-rebel-quad-at-hot-chips-2025-breaking-ais-energy-tax-with-high-performance-chiplet-innovation\/\">https:\/\/rebellions.ai\/newsroom\/rebellions-debuts-rebel-quad-at-hot-chips-2025-breaking-ais-energy-tax-with-high-performance-chiplet-innovation\/<\/a><\/li>\n<li>Rebellions Rebel100 \uc81c\ud488 \ud398\uc774\uc9c0: <a href=\"https:\/\/rebellions.ai\/rebellions-product\/rebel100\/\">https:\/\/rebellions.ai\/rebellions-product\/rebel100\/<\/a><\/li>\n<\/ul>\n<p>AMD MI455X \ud328\ud0a4\uc9c0\u00b7Die \uc790\ub8cc:<\/p>\n<ul>\n<li>AMD Instinct MI400 Series \uc81c\ud488 \ud398\uc774\uc9c0: <a href=\"https:\/\/www.amd.com\/en\/products\/accelerators\/instinct\/mi400.html\">https:\/\/www.amd.com\/en\/products\/accelerators\/instinct\/mi400.html<\/a><\/li>\n<li>AMD CDNA 5 Architecture Whitepaper: <a href=\"https:\/\/www.amd.com\/content\/dam\/amd\/en\/documents\/products\/technologies\/cdna\/amd-cdna5-whitepaper.pdf\">https:\/\/www.amd.com\/content\/dam\/amd\/en\/documents\/products\/technologies\/cdna\/amd-cdna5-whitepaper.pdf<\/a><\/li>\n<li>ServeTheHome MI455X deep dive(\ud328\ud0a4\uc9c0 \uc0ac\uc9c4 \uc218\ub85d): <a href=\"https:\/\/www.servethehome.com\/amd-instinct-mi455x-deep-dive-cdna-5-marks-the-next-era-of-instinct\/\">https:\/\/www.servethehome.com\/amd-instinct-mi455x-deep-dive-cdna-5-marks-the-next-era-of-instinct\/<\/a><\/li>\n<li>StorageReview, MI455X and Helios: <a href=\"https:\/\/www.storagereview.com\/news\/amd-mi455x-and-helios-432gb-hbm4-72-gpu-racks-and-a-real-answer-to-vera-rubin\">https:\/\/www.storagereview.com\/news\/amd-mi455x-and-helios-432gb-hbm4-72-gpu-racks-and-a-real-answer-to-vera-rubin<\/a><\/li>\n<\/ul>\n<h2>\ucc38\uace0 \uc790\ub8cc<\/h2>\n<ul>\n<li>StorageReview, “Microsoft Maia 300 Puts 288GB of HBM4 at 18TB\/s on an N2P Chiplet”: <a href=\"https:\/\/www.storagereview.com\/news\/microsoft-maia-300-hbm4-n2p-chiplet-96-accelerator-rack\">https:\/\/www.storagereview.com\/news\/microsoft-maia-300-hbm4-n2p-chiplet-96-accelerator-rack<\/a><\/li>\n<li>IN Electronics, “Microsoft details Maia 300 chiplet accelerator architecture”: <a href=\"https:\/\/www.inelectronics.co.uk\/microsoft-details-maia-300-chiplet-accelerator-architecture\/\">https:\/\/www.inelectronics.co.uk\/microsoft-details-maia-300-chiplet-accelerator-architecture\/<\/a><\/li>\n<li>Glenn Lockwood, Maia 200 \uc815\ub9ac: <a href=\"https:\/\/glennklockwood.com\/garden\/processors\/Maia-200\">https:\/\/glennklockwood.com\/garden\/processors\/Maia-200<\/a><\/li>\n<li>HPCwire, Maia 300 \ubc1c\ud45c \ubcf4\ub3c4\uc790\ub8cc: <a href=\"https:\/\/hpcwire.com\/off-the-wire\/microsoft-announces-maia-300-its-next-gen-custom-ai-chip\">https:\/\/hpcwire.com\/off-the-wire\/microsoft-announces-maia-300-its-next-gen-custom-ai-chip<\/a><\/li>\n<li>ServeTheHome, AMD Instinct MI455X Deep Dive: <a href=\"https:\/\/www.servethehome.com\/amd-instinct-mi455x-deep-dive-cdna-5-marks-the-next-era-of-instinct\/\">https:\/\/www.servethehome.com\/amd-instinct-mi455x-deep-dive-cdna-5-marks-the-next-era-of-instinct\/<\/a><\/li>\n<li>AMD Instinct MI300X, MI355X, MI455X \uacf5\uac1c \uc0ac\uc591(AMD \uc81c\ud488 \ud398\uc774\uc9c0)<\/li>\n<li>Chips and Cheese, “AMD’s Instinct MI455X: Aiming for the Sun”: <a href=\"https:\/\/chipsandcheese.com\/p\/amds-instinct-mi455x-aiming-for-the\">https:\/\/chipsandcheese.com\/p\/amds-instinct-mi455x-aiming-for-the<\/a><\/li>\n<li>ServeTheHome, “Rebellions REBEL-Quad UCIe and 144GB HBM3E Accelerator at Hot Chips 2025”: <a href=\"https:\/\/www.servethehome.com\/rebellions-rebel-quad-ucie-and-144gb-hbm3e-accelerator-at-hot-chips-2025\/\">https:\/\/www.servethehome.com\/rebellions-rebel-quad-ucie-and-144gb-hbm3e-accelerator-at-hot-chips-2025\/<\/a><\/li>\n<li>abit.ee, “Rebellions’ Rebel100 AI accelerator: quad-chiplet UCIe design (ISSCC 2026)”: <a href=\"https:\/\/abit.ee\/en\/artificial-intelligence\/rebellions-rebel100-ucie-ai-accelerator-isscc-2026-nvidia-h200-hbm3e-multi-chiplet-en\">https:\/\/abit.ee\/en\/artificial-intelligence\/rebellions-rebel100-ucie-ai-accelerator-isscc-2026-nvidia-h200-hbm3e-multi-chiplet-en<\/a><\/li>\n<li>FuriosaAI, RNGD Specifications: <a href=\"https:\/\/furiosa.ai\/renegade-spec\">https:\/\/furiosa.ai\/renegade-spec<\/a><\/li>\n<li>FuriosaAI\u2013Broadcom 3\uc138\ub300 \uac00\uc18d\uae30 \ubc1c\ud45c(2026.05): <a href=\"https:\/\/finance.yahoo.com\/news\/furiosaai-partners-broadcom-build-next-232300613.html\">https:\/\/finance.yahoo.com\/news\/furiosaai-partners-broadcom-build-next-232300613.html<\/a><\/li>\n<\/ul>\n<\/main><\/body><\/html>\n","protected":false},"excerpt":{"rendered":"<p>\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8 \uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8 \ub300\ud45c \uc774\ubbf8\uc9c0: Chiplet \ud328\ud0a4\uc9c0\uc758 Exploded View \uc77c\ub7ec\uc2a4\ud2b8(Higgsfield gpt_image_2_5 \uc0dd\uc131). \uc544\ub798\ubd80\ud130 Package Substrate,<span class=\"more-button\"><a href=\"https:\/\/kyunam.com\/?p=1043\" class=\"more-link\">\uc77d\uc5b4 \ubcfc\uae4c?<span class=\"screen-reader-text\">\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8<\/span><\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":1059,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[72,63],"tags":[179,398,399,396,168,405,400,402,403,401,171,404,406,397],"class_list":["post-1043","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-gpgpu","category-parallel-computing","tag-ai","tag-amd","tag-azure","tag-chiplet","tag-gpu","tag-gpugpu","tag-maia300","tag-mi300x","tag-mi355x","tag-mi455x","tag-npu","tag-404","tag-406","tag-397"],"_links":{"self":[{"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1043","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1043"}],"version-history":[{"count":3,"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1043\/revisions"}],"predecessor-version":[{"id":1061,"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1043\/revisions\/1061"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/media\/1059"}],"wp:attachment":[{"href":"https:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1043"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1043"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1043"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}

{"id":1043,"date":"2026-10-19T09:28:14","date_gmt":"2026-10-19T00:28:14","guid":{"rendered":"http:\/\/kyunam.com\/?p=1043"},"modified":"2026-10-19T12:27:41","modified_gmt":"2026-10-19T03:27:41","slug":"%ec%b9%a9%eb%a0%9bchiplet-%ea%b5%ac%ec%a1%b0%eb%9e%80-%eb%ac%b4%ec%97%87%ec%9d%b8%ea%b0%80-amd-azure-%eb%a6%ac%eb%b2%a8%eb%a6%ac%ec%98%a8%ec%9c%bc%eb%a1%9c-%ec%82%b4%ed%8e%b4%eb%b3%b4%eb%8a%94","status":"publish","type":"post","link":"https:\/\/kyunam.com\/?p=1043","title":{"rendered":"\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8"},"content":{"rendered":"\n\n\n\n\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8<\/title>\n<style>\nbody{margin:0;background:#fcfcfb;color:#0b0b0b;font-family:\"Noto Sans KR\",\"Apple SD Gothic Neo\",\"Malgun Gothic\",sans-serif;line-height:1.75;font-size:17px}\nmain{max-width:860px;margin:0 auto;padding:40px 20px 80px}\nh1{font-size:1.9em;line-height:1.3;margin:0 0 .6em}\nh2{font-size:1.3em;margin:2em 0 .6em;border-bottom:1px solid #e3e2dc;padding-bottom:.25em}\np{margin:0 0 1.1em}\nimg{display:block;width:100%;height:auto;border:1px solid #e3e2dc;border-radius:6px;margin:1.4em 0 .4em}\n.cap{font-size:.88em;color:#52514e;margin:0 0 1.8em}\nul{padding-left:1.3em} li{margin-bottom:.35em;word-break:break-all}\na{color:#2a78d6;text-decoration:none} a:hover{text-decoration:underline}\n.meta{color:#52514e;font-size:.9em;margin-bottom:2em}\n<\/style><\/head><body><main>\n<h1>\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8<\/h1>\n<p><img decoding=\"async\" alt=\"\ub300\ud45c \uc774\ubbf8\uc9c0\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/649AC5AF-26C9-4062-A34A-EF385763B59C.gif\"><\/p>\n<p class=\"cap\">\ub300\ud45c \uc774\ubbf8\uc9c0: Chiplet \ud328\ud0a4\uc9c0\uc758 Exploded View \uc77c\ub7ec\uc2a4\ud2b8(Higgsfield gpt_image_2_5 \uc0dd\uc131). \uc544\ub798\ubd80\ud130 Package Substrate, Silicon Interposer, \uadf8 \uc704\uc5d0 Compute Chiplet 4\uac1c, I\/O Die 2\uac1c, HBM Stack 8\uac1c\uac00 \ubd84\ub9ac\ub418\uc5b4 \ubc30\uce58\ub428. \ud2b9\uc815 \uc81c\ud488\uc758 Floorplan\uacfc \ubb34\uad00\ud568<\/p>\n<p>Microsoft\ub294 2026\ub144 10\uc6d4 12\uc77c OCP Global Summit\uc5d0\uc11c 3\uc138\ub300 \uc790\uccb4 Inference \uac00\uc18d\uae30 Maia 300\uc744 \uacf5\uac1c\ud558\uc600\ub2e4. 2027\ub144\ubd80\ud130 Azure\uc5d0 \ubc30\uce58\ub420 \uc774 \uc81c\ud488\uc740 Microsoft\uac00 \ucc98\uc74c\uc73c\ub85c Multi-Chiplet \uad6c\uc870\ub97c \ucc44\ud0dd\ud55c \uac00\uc18d\uae30\uc774\ub2e4. AMD\uac00 MI300X \uc774\ub798 MI455X\uc5d0 \uc774\ub974\uae30\uae4c\uc9c0 Compute\ub97c 8\uac1c Die\ub85c \ubd84\ud560\ud574 \uc628 \uac83\uacfc \ub2ec\ub9ac Microsoft\ub294 Compute Die\ub97c \ub2e8\uc77c\ub85c \uc720\uc9c0\ud558\uace0 I\/O\uc640 Memory Controller\ub9cc\uc744 \ubcc4\ub3c4 Die\ub85c \ubd84\ub9ac\ud558\uc600\ub2e4. \ub3d9\uc77c\ud55c Chiplet\uc774\ub77c\ub294 \uba85\uce6d \uc544\ub798 \ubd84\ud560\uc758 \ucd95\uc774 Compute\uc5d0\uc11c Memory\u00b7I\/O\ub85c \uc774\ub3d9\ud55c \uac83\uc774\ub2e4. \uad6d\ub0b4\uc5d0\uc11c\ub294 Rebellions\uac00 2025\ub144 Hot Chips\uc5d0\uc11c Compute Chiplet 4\uac1c\ub97c UCIe\ub85c \uc5f0\uacb0\ud55c REBEL-Quad\ub97c \uacf5\uac1c\ud558\uc600\uc73c\uba70, \uc774 \uc81c\ud488\uc740 AMD, Microsoft\uc640 \uad6c\ubd84\ub418\ub294 \uc81c3\uc758 \ubd84\ud560 \ubc29\uc2dd\uc5d0 \ud574\ub2f9\ud55c\ub2e4.<\/p>\n<h2>1. \uc138 \uac00\uc18d\uae30\uc758 Die \ubd84\ud560 \ubc29\uc2dd<\/h2>\n<p>AMD Instinct MI455X(CDNA 5, 2026\ub144 \ubc1c\ud45c)\ub294 TSMC N2 \uacf5\uc815\uc758 XCD(Accelerator Complex Die) 8\uac1c\ub97c N3 \uacf5\uc815\uc758 FCD(Fabric and Cache Die) 2\uac1c \uc704\uc5d0 Hybrid Bonding\uc73c\ub85c \uc218\uc9c1 \uc801\uce35\ud558\uace0, \uc5ec\uae30\uc5d0 N3 \uacf5\uc815\uc758 I\/O Die 2\uac1c\uc640 HBM4 12 Stack(432GB, 23.3TB\/s)\uc744 TSMC CoWoS-L\ub85c \uacb0\ud569\ud55c \uad6c\uc870\uc774\ub2e4. XCD \ud558\ub098\ub294 32\uac1c WGP(Workgroup Processor)\ub97c \ud3ec\ud568\ud558\uba70, FCD \ud558\ub098\ub294 96MB L2\uc640 HBM PHY, Infinity Fabric\uc744 \uc218\uc6a9\ud558\uc5ec XCD 4\uac1c\ub97c \ub2f4\ub2f9\ud55c\ub2e4. I\/O Die\ub294 \ud638\uc2a4\ud2b8 \uc5f0\uacb0, Scale-up, Scale-out NIC\ub97c \ub9e1\ub294\ub2e4. \uc804\uccb4 Transistor \uc218\ub294 3,200\uc5b5 \uac1c\uc774\uace0, Peak \uc5f0\uc0b0\uc740 MXFP4 40.3 PFLOPS, MXFP8 20.1 PFLOPS\uc774\ub2e4. \uc804 \uc138\ub300 MI355X(2025)\uac00 XCD 8\uac1c\ub97c IOD 2\uac1c \uc704\uc5d0 \uc801\uce35\ud558\uace0 Memory-side Infinity Cache 256MB\ub97c \ub450\uc5c8\ub358 \uac83\uacfc \ube44\uad50\ud558\uba74, MI455X\ub294 Infinity Cache\ub97c \uc81c\uac70\ud558\uace0 FCD\uc758 L2 192MB(96MB \u00d7 2)\ub85c \ub300\uccb4\ud558\uc600\uc73c\uba70 I\/O \uae30\ub2a5\uc744 \ubcc4\ub3c4 Die\ub85c \ubd84\ub9ac\ud558\uc600\ub2e4. \uc989 AMD\ub294 XCD 8\uac1c \uad6c\uc870\ub97c \uc138 \uc138\ub300\uc5d0 \uac78\uccd0 \uc720\uc9c0\ud558\uba74\uc11c \ud558\ubd80 Die\uc758 \uc5ed\ud560\uc744 Cache \uc911\uc2ec\uc73c\ub85c \uc7ac\ud3b8\ud558\uc600\ub2e4.<\/p>\n<p>Maia 300\uc740 TSMC N2P Compute Die 1\uac1c, I\/O Die 2\uac1c, Memory Die 4\uac1c\ub85c \uad6c\uc131\ub418\uba70 HBM4 8 Stack(288GB, 18TB\/s)\uc744 \ud0d1\uc7ac\ud55c\ub2e4. Compute Die \ub0b4\ubd80\ub294 12\uac1c Tile\ub85c \uad6c\uc131\ub41c Cluster 4\uac1c\uc640 \uc0c1\uc704 Network-on-Chip\uc73c\ub85c \uc774\ub8e8\uc5b4\uc9c0\uba70, \uac01 Tile\uc740 Tensor Unit, Vector Processor, DMA Engine, Local SRAM\uc744 \ud3ec\ud568\ud55c\ub2e4. \uc804 \uc138\ub300 Maia 200(2026\ub144 1\uc6d4 \ubc1c\ud45c)\uc740 N3P Monolithic Die\uc5d0 HBM3E 6 Stack(216GB, 7.1TB\/s)\uc744 \uacb0\ud569\ud55c \uad6c\uc870\uc600\uc73c\ubbc0\ub85c, Microsoft\ub294 \ud55c \uc138\ub300 \ub9cc\uc5d0 Monolithic\uc5d0\uc11c Chiplet\uc73c\ub85c \uc774\ud589\ud55c \uc148\uc774\ub2e4.<\/p>\n<p><img decoding=\"async\" alt=\"Azure Maia 300 \ud504\ub85c\ud1a0\ud0c0\uc785 \uce69\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/maia300_photo_microsoft.png\"><\/p>\n<p class=\"cap\">\uc0ac\uc9c4: Microsoft, Azure Maia 300 \ud504\ub85c\ud1a0\ud0c0\uc785 \uce69(OCP Global Summit 2026 \uacf5\uac1c \uc790\ub8cc). \ud328\ud0a4\uc9c0 \uc911\uc559 Die \ud45c\uba74\uc5d0 4\uac1c Cluster \uaca9\uc790\uac00 \uc2dd\ubcc4\ub428<\/p>\n<p>Rebellions Rebel100(\uac1c\ubc1c\uba85 REBEL-Quad)\uc740 Samsung SF4X \uacf5\uc815\uc758 320mm\u00b2 Compute Chiplet 4\uac1c\ub97c UCIe-Advanced\ub85c \uc5f0\uacb0\ud55c System-in-Package\uc774\ub2e4. Chiplet\ub9c8\ub2e4 HBM3E 36GB 1 Stack\uc774 \uadc0\uc18d\ub418\uc5b4 \ud328\ud0a4\uc9c0 \uc804\uccb4\ub85c 144GB\ub97c \uad6c\uc131\ud558\uba70, \ud328\ud0a4\uc9d5\uc740 TSMC CoWoS-S\uc774\uace0 \uc804\ub825 \uacfc\ub3c4 \uc751\ub2f5 \uc5b5\uc81c\ub97c \uc704\ud55c Integrated Silicon Capacitor(ISC) Die 4\uac1c\uac00 \ud568\uaed8 \ud0d1\uc7ac\ub41c\ub2e4. UCIe \ub9c1\ud06c\ub294 Lane\ub2f9 16Gbps, \uc9d1\uacc4 \uc57d 4TB\/s, Chiplet \uac04 \uc9c0\uc5f0 \uc57d 11ns\ub85c \ubc1c\ud45c\ub418\uc5c8\uc73c\uba70, 2D Mesh Network-on-Chip\uc774 UCIe\ub97c \uacbd\uc720\ud558\uc5ec 4\uac1c Chiplet\uc5d0 \uac78\uccd0 \ud655\uc7a5\ub418\ubbc0\ub85c \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uad00\uc810\uc5d0\uc11c\ub294 \ub2e8\uc77c \ud504\ub85c\uc138\uc11c\ub85c \uc778\uc2dd\ub41c\ub2e4\ub294 \uac83\uc774 Rebellions\uc758 \uc124\uba85\uc774\ub2e4. FP8 2 PFLOPS, TDP 600W\uc774\ub2e4.<\/p>\n<p>\uc138 \uad6c\uc870\ub97c \ubcd1\uce58\ud558\uba74 \ubd84\ud560 \ubc29\uc2dd\uc740 \uc138 \uac08\ub798\ub85c \uad6c\ubd84\ub41c\ub2e4. AMD\ub294 Compute(XCD), Cache\u00b7Fabric(FCD), I\/O(I\/O Die)\ub97c \uc0c1\uc774\ud55c \uacf5\uc815\uc758 Die\ub85c \ubd84\ub9ac\ud55c \uc774\uc885 \ubd84\ud560, Microsoft\ub294 Compute\ub97c \ub2e8\uc77c\ub85c \uc720\uc9c0\ud558\uace0 I\/O\uc640 Memory Controller\ub9cc \ubd84\ub9ac\ud55c \ubd84\ud560, Rebellions\ub294 IOD \uc5c6\uc774 \ub3d9\uc77c Compute Chiplet\uc744 \ubcf5\uc81c\ud558\uc5ec \uc9c1\uc811 \uc5f0\uacb0\ud55c \ub3d9\uc885 \ubd84\ud560\uc5d0 \ud574\ub2f9\ud55c\ub2e4.<\/p>\n<p><img decoding=\"async\" alt=\"\uadf8\ub9bc 1\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/IMG_1124-scaled.png\"><\/p>\n<p class=\"cap\">\uadf8\ub9bc 1. AMD MI455X(XCD 8\uac1c, FCD 2\uac1c, IOD 2\uac1c)\uc640 Microsoft Maia 300(Compute Die 1\uac1c, Memory Die 4\uac1c, I\/O Die 2\uac1c)\uc758 Chiplet \uad6c\uc870 \ube44\uad50. Rebellions Rebel100(\ub3d9\uc885 Compute Chiplet 4\uac1c, UCIe \uc5f0\uacb0)\uc758 \uad6c\uc870\ub294 1\uc808 \uc11c\uc220 \ucc38\uc870. \ud654\uc0b4\ud45c\ub294 LLM Inference \uc2dc \ub370\uc774\ud130 \uc774\ub3d9\uc744 \ub098\ud0c0\ub0b4\ub294 \uac1c\ub150\ub3c4\uc784 <\/p>\n<h2>2. \uc801\uce35 \ubc29\uc2dd\uc758 \ube44\uad50<\/h2>\n<p>AMD\ub294 XCD\ub97c FCD \uc704\uc5d0 \uc801\uce35\ud558\ub294 3D \uad6c\uc870\uc640 HBM\uc744 \uce21\uba74\uc5d0 \ubc30\uce58\ud558\ub294 2.5D CoWoS-L\uc744 \ubcd1\uc6a9\ud55c\ub2e4. XCD\uac00 FCD \uc704\uc5d0 \uc801\uce35\ub418\ub294 \uc774 \uad6c\uc870\uc5d0\uc11c\ub294 XCD\uc640 L2 \uc0ac\uc774\uc758 \uc218\uc9c1 \uc5f0\uacb0\uc774 \uc9e7\uc740 \ub300\uc2e0 \uc5f4 \ubc00\ub3c4\uac00 \uc0c1\uc2b9\ud558\uace0, \ub450 FCD \uc0ac\uc774\uc758 \ud1b5\uc2e0\uc740 Interposer\ub97c \uacbd\uc720\ud558\ub294 \uc57d 14TB\/s\uc758 Die-to-Die \ub9c1\ud06c\uc5d0 \uc758\uc874\ud55c\ub2e4. CoWoS-L\uc740 Interposer\ub97c \uc791\uc740 Local Silicon Interconnect \uc870\uac01\uacfc \uc720\uae30 \uc7ac\ubc30\uc120\uce35\uc73c\ub85c \uad6c\uc131\ud558\uc5ec \ub2e8\uc77c Silicon Interposer\uc758 Reticle \ud55c\uacc4\ub97c \ub118\ub294 \ud328\ud0a4\uc9c0 \uba74\uc801\uc744 \ud655\ubcf4\ud558\ub294 \ubc29\uc2dd\uc774\uba70, HBM4 12 Stack\uc744 \uc218\uc6a9\ud558\uae30 \uc704\ud55c \uc120\ud0dd\uc73c\ub85c \ud574\uc11d\ub41c\ub2e4. Rebellions\ub294 Compute Chiplet 4\uac1c\uc640 HBM\uc744 \ubaa8\ub450 CoWoS-S Interposer \uc704 \ub2e8\uc77c \uce35\uc5d0 \ubc30\uce58\ud558\uace0 Chiplet \uac04 \uc5f0\uacb0\uc744 Interposer \ubc30\uc120\uc758 UCIe-Advanced\ub85c \ucc98\ub9ac\ud558\ubbc0\ub85c, 3D \uc801\uce35 \uc5c6\uc774 2.5D\ub9cc\uc73c\ub85c Compute\ub97c \ubd84\ud560\ud55c \uc0ac\ub840\uc774\ub2e4. Maia 300\uc740 \uacf5\uac1c \uc790\ub8cc \uae30\uc900\uc73c\ub85c 3D \uc801\uce35 \uc5b8\uae09\uc774 \uc5c6\uc73c\uba70, Compute\u00b7I\/O\u00b7Memory Die\uac00 Interposer \uc704 \ub2e8\uc77c \uce35\uc5d0 \ub193\uc774\ub294 2.5D \uad6c\uc870\ub85c \ud30c\uc545\ub41c\ub2e4. Microsoft\ub294 “HBM Stack\uc744 Compute\uc5d0 \uadfc\uc811 \ubc30\uce58\ud558\uae30 \uc704\ud55c Advanced Packaging”\uc73c\ub85c\ub9cc \uc124\uba85\ud558\ubbc0\ub85c \uc138\ubd80 \uacf5\uc815\uc740 \ucd94\uac00 \uacf5\uac1c\ub97c \uc694\ud55c\ub2e4.<\/p>\n<p><img decoding=\"async\" alt=\"\uadf8\ub9bc 2\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/IMG_1125-scaled.png\"><\/p>\n<p class=\"cap\">\uadf8\ub9bc 2. \ub2e8\uba74 \ube44\uad50. AMD MI455X\ub294 XCD\ub97c FCD \uc704\uc5d0 \uc218\uc9c1 \uc801\uce35\ud558\uace0, Maia 300\uc740 \uacf5\uac1c \uc815\ubcf4 \uae30\uc900\uc73c\ub85c \ubaa8\ub4e0 Die\ub97c \ub3d9\uc77c \uce35\uc5d0 \ubc30\uce58\ud568<\/p>\n<h2>3. \uc218\uce58\ub85c \ubcf8 \uc124\uacc4 \ubc29\ud5a5<\/h2>\n<p>\uc138\ub300\ubcc4 \uc0ac\uc591\uc744 \ub3d9\uc77c \ucd95\uc5d0 \ubc30\uce58\ud558\uba74 \ub450 \ud68c\uc0ac\uc758 \ubc29\ud5a5\uc774 \uad6c\ubd84\ub41c\ub2e4. Microsoft\ub294 Maia 200\uc5d0\uc11c 300\uc73c\ub85c \uc774\ud589\ud558\uba70 FP8 \uc5f0\uc0b0\uc744 5.1\uc5d0\uc11c 6.9 PFLOPS\ub85c 1.35\ubc30 \uc99d\uac00\uc2dc\ud0a8 \ubc18\uba74 HBM \ub300\uc5ed\ud3ed\uc740 7.1\uc5d0\uc11c 18TB\/s\ub85c 2.5\ubc30 \uc99d\uac00\uc2dc\ucf30\ub2e4. AMD\ub294 MI355X\uc5d0\uc11c MI455X\ub85c \uc774\ud589\ud558\uba70 FP8 \uc5f0\uc0b0\uc744 5.0\uc5d0\uc11c \uc57d 20 PFLOPS\ub85c 4\ubc30, \ub300\uc5ed\ud3ed\uc744 8.0\uc5d0\uc11c 23.3TB\/s\ub85c 2.9\ubc30 \uc99d\uac00\uc2dc\ucf30\ub2e4. \uc5f0\uc0b0 1 PFLOPS\ub2f9 \ub300\uc5ed\ud3ed\uc73c\ub85c \ud658\uc0b0\ud558\uba74 Maia 300\uc740 2.61TB\/s, MI455X\ub294 1.17TB\/s\ub85c \uc0b0\uc815\ub41c\ub2e4. Maia 300\uc740 \ub300\uc5ed\ud3ed \uce21\uc73c\ub85c, MI455X\ub294 \uc5f0\uc0b0 \uce21\uc73c\ub85c \uade0\ud615\uc744 \uc774\ub3d9\uc2dc\ud0a8 \uac83\uc774\ub2e4.<\/p>\n<p><img decoding=\"async\" alt=\"\uadf8\ub9bc 3\" src=\"https:\/\/kyunam.com\/wp-content\/uploads\/2026\/10\/IMG_1127-scaled.png\"><\/p>\n<p class=\"cap\">\uadf8\ub9bc 3. \uc138\ub300\ubcc4 \uc0ac\uc591 \ucd94\uc774. MI455X\uc758 FP8 \uc218\uce58\ub294 \ubc1c\ud45c\ub41c FP4 40 PFLOPS\uc758 \uc808\ubc18\uc73c\ub85c \uc0b0\uc815\ud568<\/p>\n<p>\uc774 \ucc28\uc774\ub294 \ub450 \uc81c\ud488\uc758 \uc5ed\ud560\uacfc \uc77c\uce58\ud55c\ub2e4. Maia 300\uc740 Microsoft \uc790\uc0ac \uc11c\ube44\uc2a4\uc758 Inference \uc804\uc6a9 \uac00\uc18d\uae30\uc774\uba70, Decode \ub2e8\uacc4\uc758 \ubcd1\ubaa9\uc740 \uc5f0\uc0b0\ubcf4\ub2e4 KV cache\uc640 Weight \uc801\uc7ac \ub300\uc5ed\ud3ed\uc5d0 \uc788\ub2e4. MI455X\ub294 \ud559\uc2b5\uacfc Inference\ub97c \ubaa8\ub450 \ub300\uc0c1\uc73c\ub85c \ud558\ub294 \ubc94\uc6a9 GPU\uc774\ubbc0\ub85c \uc5f0\uc0b0 \ubc00\ub3c4\uc758 \uc720\uc9c0\uac00 \uc694\uad6c\ub41c\ub2e4. \ub530\ub77c\uc11c Maia 300\uc758 Chiplet \ubd84\ud560\uc740 HBM4 8 Stack\uc744 \uc218\uc6a9\ud560 Beachfront(Die \uac00\uc7a5\uc790\ub9ac \uae38\uc774)\uc758 \ud655\ubcf4\uac00 \ubaa9\uc801\uc774\uace0, AMD\uc758 XCD \ubd84\ud560\uc740 Reticle \ud55c\uacc4\uc640 \uc218\uc728 \ubb38\uc81c\uc758 \ud574\uc18c\uac00 \ubaa9\uc801\uc774\ub2e4. \ubd84\ud560\uc758 \ub3d9\uae30\uac00 \uc0c1\uc774\ud558\ubbc0\ub85c \ubd84\ud560\uc758 \ucd95 \ub610\ud55c \uc0c1\uc774\ud558\ub2e4.<\/p>\n<h2>4. \uad6c\uc870\uc640 \ubb34\uad00\ud558\uac8c \uc794\uc874\ud558\ub294 NUMA<\/h2>\n<p>Compute\ub97c \ub2e8\uc77c Die\ub85c \uc720\uc9c0\ud558\ub354\ub77c\ub3c4 \uba54\ubaa8\ub9ac \uc811\uadfc\uc774 \uade0\uc77c\ud574\uc9c0\ub294 \uac83\uc740 \uc544\ub2c8\ub2e4. Microsoft\ub294 Maia 300\uc758 \uba54\ubaa8\ub9ac \ubaa8\ub378\uc744 \uba85\uc2dc\uc801 NUMA\ub85c \uaddc\uc815\ud558\uace0, \ube48\ubc88\ud788 \uc811\uadfc\ud558\ub294 \ub370\uc774\ud130\ub97c \ud574\ub2f9 Compute \uadfc\ucc98\uc5d0 \ubc30\uce58\ud558\ub294 \ucc45\uc784\uc744 Software-Defined Dataflow\uc5d0 \ubd80\uc5ec\ud55c\ub2e4. Memory Die 4\uac1c\uac00 \uac01\uac01 HBM4 2 Stack\uc744 \ub2f4\ub2f9\ud558\ub294 \uad6c\uc870\ub77c\uba74 Compute Die\uc758 Cluster\uc640 Memory Die \uc0ac\uc774\uc758 \uac70\ub9ac\uc5d0 \ub530\ub77c \uc9c0\uc5f0\uc774 \uc0c1\uc774\ud574\uc9c0\ub294 \uac83\uc740 \uad6c\uc870\uc0c1 \ud544\uc5f0\uc774\ub2e4.<\/p>\n<p>Rebel100\uc5d0\uc11c\ub294 NUMA\uac00 \uad6c\uc870\uc801\uc73c\ub85c \ub354\uc6b1 \ub69c\ub837\ud558\ub2e4. HBM\uc774 Chiplet\ubcc4\ub85c 1 Stack\uc529 \uadc0\uc18d\ub418\ubbc0\ub85c \ud0c0 Chiplet\uc5d0 \uacb0\ud569\ub41c HBM \uc811\uadfc\uc740 \ubc18\ub4dc\uc2dc UCIe \ub9c1\ud06c\ub97c \uacbd\uc720\ud558\uba70, \uc9d1\uacc4 4TB\/s\uc758 UCIe \ub300\uc5ed\ud3ed\uc744 4\uac1c Chiplet\uc774 \ubd84\ud560\ud558\uc5ec \uc0ac\uc6a9\ud55c\ub2e4. “\ub2e8\uc77c \ud504\ub85c\uc138\uc11c\ub85c \uc778\uc2dd\ub41c\ub2e4”\ub294 \uc124\uba85\uc740 \uc8fc\uc18c \uacf5\uac04\uacfc NoC \ud1a0\ud3f4\ub85c\uc9c0 \uad00\uc810\uc758 \uc11c\uc220\uc774\uace0, \uc9c0\uc5f0\uacfc \ub300\uc5ed\ud3ed\uc758 \ube44\uade0\uc77c\uc131\uc740 \uc794\uc874\ud55c\ub2e4. \ub530\ub77c\uc11c KV cache\uc640 Weight\ub97c \uac01 Chiplet\uc758 Local HBM\uc5d0 \ubd84\ud560 \ubc30\uce58\ud558\uace0 Chiplet \uacbd\uacc4\ub97c \ub118\ub294 \uc811\uadfc\uc744 \ucd5c\uc18c\ud654\ud558\ub294 \uac83\uc774 \uc131\ub2a5\uc758 \uc804\uc81c\uac00 \ub41c\ub2e4.<\/p>\n<p>AMD\uc5d0\uc11c\ub294 \uc774 \ubb38\uc81c\uac00 \ud558\ub4dc\uc6e8\uc5b4 \uc77c\uad00\uc131 \uc218\uc900\uc5d0\uc11c \ub4dc\ub7ec\ub098\uba70, \uc138\ub300\uc5d0 \ub530\ub77c \uadf8 \ub2e8\uc704\uac00 \ubcc0\uacbd\ub418\uc5c8\ub2e4. MI300X\uc640 MI355X\uc5d0\uc11c\ub294 XCD\ub9c8\ub2e4 L2\uac00 \ub3c5\ub9bd\uc801\uc73c\ub85c \uc874\uc7ac\ud558\uace0 XCD \uac04 L2\uac00 \ube44\uc77c\uad00\uc131(Non-Coherent)\uc774\uc5c8\uc73c\ubbc0\ub85c, \ud558\ub098\uc758 \uc791\uc5c5\uc744 \ud558\ub098\uc758 XCD \ub0b4\ubd80\uc5d0 \ud55c\uc815\ud558\uba74(XCD Confinement) Cache Maintenance \uc5c6\uc774 L2\ub97c \uc7ac\uc0ac\uc6a9\ud560 \uc218 \uc788\uc5c8\uc73c\ub098 XCD \uacbd\uacc4\ub97c \ub118\ub294 \uc21c\uac04 IOD \uacbd\uc720 \ube44\uc6a9\uacfc \uc77c\uad00\uc131 \ube44\uc6a9\uc774 \ub3d9\uc2dc\uc5d0 \ubc1c\uc0dd\ud558\uc600\ub2e4. MI455X\ub294 L2\ub97c XCD\uc5d0\uc11c FCD\ub85c \uc774\uc804\ud558\uc5ec XCD 4\uac1c\uac00 96MB L2 \ud558\ub098\ub97c \uacf5\uc720\ud558\ub3c4\ub85d \ud558\uc600\uace0, \uc774\ub85c\uc368 XCD \uac04 Atomic \uc5f0\uc0b0\uc774 \uac1c\uc120\ub418\uace0 MI300 \uacc4\uc5f4\uc5d0\uc11c \uc804\uc5ed \uac00\uc2dc\uc131 \ud655\ubcf4\uc5d0 \ud544\uc694\ud558\uc600\ub358 Kernel \uacbd\uacc4\uc758 Cache Flush\uac00 \uc81c\uac70\ub418\uc5c8\ub2e4. \ub2e4\ub9cc \ud55c FCD\uc758 WGP\ub294 \ub2e4\ub978 FCD\uc758 L2\uc5d0 \uc811\uadfc\ud560 \uc218 \uc5c6\uc73c\ubbc0\ub85c, \uc77c\uad00\uc131 \ub2e8\uc704\uac00 XCD 8\uac1c\uc5d0\uc11c FCD 2\uac1c\ub85c \ud655\ub300\ub418\uc5c8\uc744 \ubfd0 FCD \uacbd\uacc4\uc758 NUMA\ub294 \uc794\uc874\ud55c\ub2e4. \uac01 L2\uac00 \uc804\uc5ed \uc8fc\uc18c \uc804\uccb4\ub97c Cache\ud560 \uc218 \uc788\uc5b4 Die-to-Die \ud2b8\ub798\ud53d\uc774 \uc644\ud654\ub418\uc9c0\ub9cc, \uc791\uc5c5\uc744 FCD \ub2e8\uc704\ub85c \ubc30\uce58\ud558\ub294 \uac83\uc774 \uc5ec\uc804\ud788 \uc720\ub9ac\ud558\ub2e4. Maia 300\uc774 NUMA\ub97c \uc18c\ud504\ud2b8\uc6e8\uc5b4\uc5d0 \ub178\ucd9c\uc2dc\ud0a8 \uacb0\uc815\uc740, \ud558\ub4dc\uc6e8\uc5b4\uac00 Locality \ub2e8\uc704\ub97c \uc740\ud3d0\ud558\uc9c0 \uc54a\uace0 \ub4dc\ub7ec\ub0b4\ub294 \ud3b8\uc774 Inference \uac00\uc18d\uae30\uc5d0\uc11c\ub3c4 \uc720\ub9ac\ud558\ub2e4\ub294 \ud310\ub2e8\uc73c\ub85c \ud574\uc11d\ub41c\ub2e4. “\uc791\uc5c5\uc744 \ud558\ub4dc\uc6e8\uc5b4 Locality \ub2e8\uc704\uc5d0 \ud55c\uc815\ud558\uace0 \ub2e8\uc704 \uac04 \ud1b5\uc2e0\uc744 \uba85\uc2dc\uc801\uc73c\ub85c \uad00\ub9ac\ud55c\ub2e4”\ub294 \uc6d0\uce59\uc774 \uc774\uc885 Compute \ubd84\ud560\ud615(AMD), Memory\u00b7I\/O \ubd84\ud560\ud615(Microsoft), \ub3d9\uc885 Compute \ubd84\ud560\ud615(Rebellions) \uc138 \uad6c\uc870 \ubaa8\ub450\uc5d0 \uc801\uc6a9\ub418\ub294 \uac83\uc774\ub2e4.<\/p>\n<h2>5. \uc804\ub825\uacfc Rack \uc218\uc900\uc758 \ud568\uc758<\/h2>\n<p>Maia 300\uc758 TDP\ub294 950W\ub85c Maia 200(750W) \ub300\ube44 27% \uc0c1\uc2b9\ud558\uc600\uc73c\uba70, Rack\ub2f9 96\uac1c\ub97c 167kW\uc5d0 \uc218\uc6a9\ud55c\ub2e4. MI355X\ub294 1,400W\uc774\uace0 MI455X\ub294 \uacf5\uc2dd TDP\uac00 \ubbf8\uacf5\uac1c\uc774\ub098 Helios Rack \uc804\ub825\uc73c\ub85c\ubd80\ud130 2kW \uc774\uc0c1\uc73c\ub85c \ucd94\uc815\ub41c\ub2e4. \ub300\uc5ed\ud3ed \uc911\uc2ec \uc124\uacc4\uac00 \uc5f0\uc0b0 \uc911\uc2ec \uc124\uacc4\ubcf4\ub2e4 \uc804\ub825 \uc99d\uac00 \ud3ed\uc774 \uc791\ub2e4\ub294 \uc810\uc774 \uc218\uce58\ub85c \ud655\uc778\ub41c\ub2e4. GPU Cluster \uc6b4\uc601 \uad00\uc810\uc5d0\uc11c \ub450 \uac00\uc9c0 \ud568\uc758\uac00 \ub3c4\ucd9c\ub41c\ub2e4. \uccab\uc9f8, \uc790\uc6d0 \uad00\ub9ac \ub2e8\uc704\uac00 “GPU \ud55c \uc7a5”\uc5d0\uc11c Die \ub610\ub294 NUMA domain\uc73c\ub85c \uc138\ubd84\ud654\ub420 \uac00\ub2a5\uc131\uc758 \uc99d\ub300\uc774\ub2e4. MI300 \uacc4\uc5f4\uc758 DPX\u00b7QPX Partition \ubaa8\ub4dc\ub294 \uadf8 \ucd08\uae30 \ud615\ud0dc\uc774\uba70, MI455X\uc758 FCD 2\uac1c \uad6c\uc870, Maia 300\uc758 Cluster\u00b7Memory Die \uad6c\uc870, Rebel100\uc758 Chiplet \ub2e8\uc704 \uad6c\uc870 \ub610\ud55c \ub3d9\uc77c \ubc29\ud5a5\uc758 \ubd84\ud560\uc744 \ud5c8\uc6a9\ud560 \uc5ec\uc9c0\uac00 \uc788\ub2e4. \ub458\uc9f8, Inference \uc804\uc6a9 \uac00\uc18d\uae30\uc640 \ubc94\uc6a9 GPU\uc758 \uc0ac\uc591 \uaca9\ucc28\uac00 \ub300\uc5ed\ud3ed\/\uc5f0\uc0b0 \ube44\ub85c \uc815\ub7c9\ud654\ub418\ubbc0\ub85c, \uc6cc\ud06c\ub85c\ub4dc\ubcc4 \ubc30\uce58 \uae30\uc900\uc744 \uc774 \ube44\uc728\ub85c \uc218\ub9bd\ud558\ub294 \uac83\uc774 \uac00\ub2a5\ud574\uc9c4\ub2e4.<\/p>\n<h2>6. \uad6d\ub0b4 \uc0ac\ub840: \ub450 \uac08\ub798 \ubd84\ud560 \ubc29\uc2dd\uc758 \uad6d\ub0b4 \uc7ac\ud604<\/h2>\n<p>\uad6d\ub0b4 \ub450 \ud68c\uc0ac\ub294 Chiplet \ubd84\ud560\uc758 \uc0c1\uc774\ud55c \uac08\ub798\ub97c \uac01\uac01 \ucc44\ud0dd\ud558\uc600\ub2e4. Rebellions\ub294 \uc804\uc220\ud55c \ubc14\uc640 \uac19\uc774 Compute\ub97c \ub3d9\uc885 Chiplet 4\uac1c\ub85c \ubd84\ud560\ud558\ub294 \ubc29\uc2dd\uc774\ub2e4. FuriosaAI\uc758 \ud604\ud589 \uc81c\ud488 RNGD\ub294 TSMC 5nm \ub2e8\uc77c Die\uc5d0 HBM3 48GB(1.5TB\/s), SRAM 256MB, FP8 512 TFLOPS, TDP 180W\uc758 Monolithic \uc124\uacc4\uc774\uba70, 2026\ub144 5\uc6d4 Broadcom\uacfc \ubc1c\ud45c\ud55c 3\uc138\ub300 \uac00\uc18d\uae30\uc5d0\uc11c 2nm Compute Die\uc640 Scale-up Networking\uc6a9 \ubcc4\ub3c4 I\/O Die\ub97c Broadcom\uc758 Advanced Packaging\uc73c\ub85c \ud1b5\ud569\ud558\uace0 HBM4\/4E\ub97c \ucc44\ud0dd\ud560 \uc608\uc815\uc774\ub2e4(2028\ub144 \uc0c1\ubc18\uae30 \uc0d8\ud50c\ub9c1). \uc989 FuriosaAI\uc758 \ucc28\uc138\ub300\ub294 Maia 300\uacfc \ub3d9\uc77c\ud55c Compute \ud1b5\ud569\u00b7I\/O \ubd84\ub9ac\ud615\uc5d0 \ud574\ub2f9\ud55c\ub2e4. \uad6d\ub0b4\uc5d0\uc11c\ub3c4 Compute \ubd84\ud560\ud615\uacfc I\/O \ubd84\ub9ac\ud615\uc774 \ubcd1\uc874\ud558\uac8c \ub418\uba70, \uc5b4\ub290 \ubc29\uc2dd\uc774\ub4e0 Locality \uad00\ub9ac \ucc45\uc784\uc774 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc2a4\ud0dd\uc73c\ub85c \uc774\uc804\ub41c\ub2e4\ub294 \uc810\uc740 \ub3d9\uc77c\ud558\ub2e4.<\/p>\n<h2>7. \uacb0\ub860<\/h2>\n<p>Maia 300, AMD Instinct, Rebellions Rebel100\uc740 \ubaa8\ub450 Chiplet\uc774\ub098 \ubd84\ud560\uc758 \ucd95\uc774 \uc0c1\uc774\ud558\ub2e4. AMD\ub294 Compute\ub97c 8\uac1c XCD\ub85c \ubd84\ud560\ud558\uc5ec \uc218\uc728\uacfc \uba74\uc801 \ubb38\uc81c\ub97c \ud574\uc18c\ud558\uc600\uace0, MI455X\uc5d0\uc11c L2\ub97c FCD\ub85c \uc774\uc804\ud558\uc5ec \uc77c\uad00\uc131 \ub2e8\uc704\ub97c XCD 8\uac1c\uc5d0\uc11c FCD 2\uac1c\ub85c \ud655\ub300\ud558\uc600\uc73c\ub098 FCD \uacbd\uacc4\uc758 NUMA\ub294 \uc18c\ud504\ud2b8\uc6e8\uc5b4\uac00 \uad00\ub9ac\ud558\uc5ec\uc57c \ud55c\ub2e4. Microsoft\ub294 Compute Die\ub97c \ub2e8\uc77c\ub85c \uc720\uc9c0\ud558\uace0 Memory\u00b7I\/O\ub97c \ubd84\ub9ac\ud558\uc5ec HBM4 \ub300\uc5ed\ud3ed\uc744 2.5\ubc30 \ud655\ub300\ud558\uc600\uc73c\uba70, NUMA\ub97c \uc124\uacc4 \ucd08\uae30\ubd80\ud130 \uba85\uc2dc\uc801\uc73c\ub85c \ub178\ucd9c\ud558\uc600\ub2e4. Rebellions\ub294 IOD \uc5c6\uc774 \ub3d9\uc885 Compute Chiplet\uc744 UCIe\ub85c \uc9c1\uc811 \uc5f0\uacb0\ud558\uc5ec \ud328\ud0a4\uc9c0 \uc804\uccb4\ub97c \ub2e8\uc77c Mesh\ub85c \uacb0\ud569\ud558\uc600\uace0, \uadf8 \ub300\uc2e0 Chiplet\ubcc4\ub85c \uadc0\uc18d\ub41c HBM \uac04\uc758 \ube44\uade0\uc77c \uc811\uadfc\uc744 \uc18c\ud504\ud2b8\uc6e8\uc5b4\uac00 \uad00\ub9ac\ud558\uc5ec\uc57c \ud55c\ub2e4. \ud5a5\ud6c4 \uc5b4\ub290 \ubc29\uc2dd\uc774 \uc8fc\ub958\uac00 \ub420\uc9c0\ub294 Inference \ube44\uc911\uc774 \uacb0\uc815\ud560 \uac83\uc774\ub098, \uc138 \uad6c\uc870 \ubaa8\ub450\uc5d0\uc11c Locality\ub97c \uc18c\ud504\ud2b8\uc6e8\uc5b4\uac00 \uad00\ub9ac\ud558\uc5ec\uc57c \ud55c\ub2e4\ub294 \uc810\uc740 \uc774\ubbf8 \ud655\uc815\ub41c \uc0ac\uc2e4\uc774\ub2e4.<\/p>\n<h2>\uc2e4\ubb3c \uc0ac\uc9c4 \ucd9c\ucc98<\/h2>\n<p>\ubcf8\ubb38 \ub3c4\ud574\ub294 \ubaa8\ub450 \uacf5\uac1c \uc0ac\uc591\uc744 \ubc14\ud0d5\uc73c\ub85c \uc9c1\uc811 \uc791\uc131\ud55c \uac83\uc774\uba70, \uc2e4\ubb3c \uc0ac\uc9c4\uc740 \uc544\ub798 \ucd9c\ucc98\ub97c \ucc38\uc870\ud568.<\/p>\n<p>Microsoft Maia 300 \uacf5\uc2dd \uc774\ubbf8\uc9c0(Microsoft \uc81c\uacf5, news.microsoft.com\/maia-2026):<\/p>\n<ul>\n<li>\ud504\ub85c\ud1a0\ud0c0\uc785 \uce69 hero \uc774\ubbf8\uc9c0: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-prototype-chip-00_hero.png\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-prototype-chip-00_hero.png<\/a><\/li>\n<li>\uce69 render: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Azure-Maia-300-chip-render.png\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Azure-Maia-300-chip-render.png<\/a><\/li>\n<li>\uce69\u00b7\uc11c\ubc84\u00b7rack render: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Azure-Maia-300-Chip-to-Server-to-Rack-Render.png\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Azure-Maia-300-Chip-to-Server-to-Rack-Render.png<\/a><\/li>\n<li>\ubcf4\ub4dc \uc0ac\uc9c4: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-board-1.jpg\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-board-1.jpg<\/a><\/li>\n<li>\uc11c\ubc84 \uc0ac\uc9c4: <a href=\"https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-server-1.jpg\">https:\/\/msftstories.thesourcemediaassets.com\/sites\/760\/2026\/10\/Maia-300-server-1.jpg<\/a><\/li>\n<li>\uacf5\uc2dd \ubc1c\ud45c \ube14\ub85c\uadf8: <a href=\"https:\/\/blogs.microsoft.com\/blog\/2026\/10\/12\/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments\/\">https:\/\/blogs.microsoft.com\/blog\/2026\/10\/12\/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments\/<\/a><\/li>\n<\/ul>\n<p>Rebellions Rebel100 \/ REBEL-Quad:<\/p>\n<ul>\n<li>ServeTheHome Hot Chips 2025 \uae30\uc0ac(\ud328\ud0a4\uc9c0\u00b7\ubcf4\ub4dc \uc0ac\uc9c4 \uc218\ub85d): <a href=\"https:\/\/www.servethehome.com\/rebellions-rebel-quad-ucie-and-144gb-hbm3e-accelerator-at-hot-chips-2025\/\">https:\/\/www.servethehome.com\/rebellions-rebel-quad-ucie-and-144gb-hbm3e-accelerator-at-hot-chips-2025\/<\/a><\/li>\n<li>Rebellions \uacf5\uc2dd \ub274\uc2a4\ub8f8: <a href=\"https:\/\/rebellions.ai\/newsroom\/rebellions-debuts-rebel-quad-at-hot-chips-2025-breaking-ais-energy-tax-with-high-performance-chiplet-innovation\/\">https:\/\/rebellions.ai\/newsroom\/rebellions-debuts-rebel-quad-at-hot-chips-2025-breaking-ais-energy-tax-with-high-performance-chiplet-innovation\/<\/a><\/li>\n<li>Rebellions Rebel100 \uc81c\ud488 \ud398\uc774\uc9c0: <a href=\"https:\/\/rebellions.ai\/rebellions-product\/rebel100\/\">https:\/\/rebellions.ai\/rebellions-product\/rebel100\/<\/a><\/li>\n<\/ul>\n<p>AMD MI455X \ud328\ud0a4\uc9c0\u00b7Die \uc790\ub8cc:<\/p>\n<ul>\n<li>AMD Instinct MI400 Series \uc81c\ud488 \ud398\uc774\uc9c0: <a href=\"https:\/\/www.amd.com\/en\/products\/accelerators\/instinct\/mi400.html\">https:\/\/www.amd.com\/en\/products\/accelerators\/instinct\/mi400.html<\/a><\/li>\n<li>AMD CDNA 5 Architecture Whitepaper: <a href=\"https:\/\/www.amd.com\/content\/dam\/amd\/en\/documents\/products\/technologies\/cdna\/amd-cdna5-whitepaper.pdf\">https:\/\/www.amd.com\/content\/dam\/amd\/en\/documents\/products\/technologies\/cdna\/amd-cdna5-whitepaper.pdf<\/a><\/li>\n<li>ServeTheHome MI455X deep dive(\ud328\ud0a4\uc9c0 \uc0ac\uc9c4 \uc218\ub85d): <a href=\"https:\/\/www.servethehome.com\/amd-instinct-mi455x-deep-dive-cdna-5-marks-the-next-era-of-instinct\/\">https:\/\/www.servethehome.com\/amd-instinct-mi455x-deep-dive-cdna-5-marks-the-next-era-of-instinct\/<\/a><\/li>\n<li>StorageReview, MI455X and Helios: <a href=\"https:\/\/www.storagereview.com\/news\/amd-mi455x-and-helios-432gb-hbm4-72-gpu-racks-and-a-real-answer-to-vera-rubin\">https:\/\/www.storagereview.com\/news\/amd-mi455x-and-helios-432gb-hbm4-72-gpu-racks-and-a-real-answer-to-vera-rubin<\/a><\/li>\n<\/ul>\n<h2>\ucc38\uace0 \uc790\ub8cc<\/h2>\n<ul>\n<li>StorageReview, “Microsoft Maia 300 Puts 288GB of HBM4 at 18TB\/s on an N2P Chiplet”: <a href=\"https:\/\/www.storagereview.com\/news\/microsoft-maia-300-hbm4-n2p-chiplet-96-accelerator-rack\">https:\/\/www.storagereview.com\/news\/microsoft-maia-300-hbm4-n2p-chiplet-96-accelerator-rack<\/a><\/li>\n<li>IN Electronics, “Microsoft details Maia 300 chiplet accelerator architecture”: <a href=\"https:\/\/www.inelectronics.co.uk\/microsoft-details-maia-300-chiplet-accelerator-architecture\/\">https:\/\/www.inelectronics.co.uk\/microsoft-details-maia-300-chiplet-accelerator-architecture\/<\/a><\/li>\n<li>Glenn Lockwood, Maia 200 \uc815\ub9ac: <a href=\"https:\/\/glennklockwood.com\/garden\/processors\/Maia-200\">https:\/\/glennklockwood.com\/garden\/processors\/Maia-200<\/a><\/li>\n<li>HPCwire, Maia 300 \ubc1c\ud45c \ubcf4\ub3c4\uc790\ub8cc: <a href=\"https:\/\/hpcwire.com\/off-the-wire\/microsoft-announces-maia-300-its-next-gen-custom-ai-chip\">https:\/\/hpcwire.com\/off-the-wire\/microsoft-announces-maia-300-its-next-gen-custom-ai-chip<\/a><\/li>\n<li>ServeTheHome, AMD Instinct MI455X Deep Dive: <a href=\"https:\/\/www.servethehome.com\/amd-instinct-mi455x-deep-dive-cdna-5-marks-the-next-era-of-instinct\/\">https:\/\/www.servethehome.com\/amd-instinct-mi455x-deep-dive-cdna-5-marks-the-next-era-of-instinct\/<\/a><\/li>\n<li>AMD Instinct MI300X, MI355X, MI455X \uacf5\uac1c \uc0ac\uc591(AMD \uc81c\ud488 \ud398\uc774\uc9c0)<\/li>\n<li>Chips and Cheese, “AMD’s Instinct MI455X: Aiming for the Sun”: <a href=\"https:\/\/chipsandcheese.com\/p\/amds-instinct-mi455x-aiming-for-the\">https:\/\/chipsandcheese.com\/p\/amds-instinct-mi455x-aiming-for-the<\/a><\/li>\n<li>ServeTheHome, “Rebellions REBEL-Quad UCIe and 144GB HBM3E Accelerator at Hot Chips 2025”: <a href=\"https:\/\/www.servethehome.com\/rebellions-rebel-quad-ucie-and-144gb-hbm3e-accelerator-at-hot-chips-2025\/\">https:\/\/www.servethehome.com\/rebellions-rebel-quad-ucie-and-144gb-hbm3e-accelerator-at-hot-chips-2025\/<\/a><\/li>\n<li>abit.ee, “Rebellions’ Rebel100 AI accelerator: quad-chiplet UCIe design (ISSCC 2026)”: <a href=\"https:\/\/abit.ee\/en\/artificial-intelligence\/rebellions-rebel100-ucie-ai-accelerator-isscc-2026-nvidia-h200-hbm3e-multi-chiplet-en\">https:\/\/abit.ee\/en\/artificial-intelligence\/rebellions-rebel100-ucie-ai-accelerator-isscc-2026-nvidia-h200-hbm3e-multi-chiplet-en<\/a><\/li>\n<li>FuriosaAI, RNGD Specifications: <a href=\"https:\/\/furiosa.ai\/renegade-spec\">https:\/\/furiosa.ai\/renegade-spec<\/a><\/li>\n<li>FuriosaAI\u2013Broadcom 3\uc138\ub300 \uac00\uc18d\uae30 \ubc1c\ud45c(2026.05): <a href=\"https:\/\/finance.yahoo.com\/news\/furiosaai-partners-broadcom-build-next-232300613.html\">https:\/\/finance.yahoo.com\/news\/furiosaai-partners-broadcom-build-next-232300613.html<\/a><\/li>\n<\/ul>\n<\/main><\/body><\/html>\n","protected":false},"excerpt":{"rendered":"<p>\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8 \uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8 \ub300\ud45c \uc774\ubbf8\uc9c0: Chiplet \ud328\ud0a4\uc9c0\uc758 Exploded View \uc77c\ub7ec\uc2a4\ud2b8(Higgsfield gpt_image_2_5 \uc0dd\uc131). \uc544\ub798\ubd80\ud130 Package Substrate,<span class=\"more-button\"><a href=\"https:\/\/kyunam.com\/?p=1043\" class=\"more-link\">\uc77d\uc5b4 \ubcfc\uae4c?<span class=\"screen-reader-text\">\uce69\ub81b(Chiplet) \uad6c\uc870\ub780 \ubb34\uc5c7\uc778\uac00? AMD, Azure, \ub9ac\ubca8\ub9ac\uc628\uc73c\ub85c \uc0b4\ud3b4\ubcf4\ub294 \uce69\ub81b\uc758 \uc758\ubbf8<\/span><\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":1059,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[72,63],"tags":[179,398,399,396,168,405,400,402,403,401,171,404,406,397],"class_list":["post-1043","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-gpgpu","category-parallel-computing","tag-ai","tag-amd","tag-azure","tag-chiplet","tag-gpu","tag-gpugpu","tag-maia300","tag-mi300x","tag-mi355x","tag-mi455x","tag-npu","tag-404","tag-406","tag-397"],"_links":{"self":[{"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1043","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1043"}],"version-history":[{"count":3,"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1043\/revisions"}],"predecessor-version":[{"id":1061,"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1043\/revisions\/1061"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/media\/1059"}],"wp:attachment":[{"href":"https:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1043"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1043"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1043"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}