{
 "source": "https://qdna.fr/outils/calculateur-memoire-llm",
 "version": "3.1",
 "mis_a_jour": "2026-10-12",
 "citation": "QDNA, calculateur de mémoire, version 3.1 du 2026-10-12, https://qdna.fr/outils/calculateur-memoire-llm",
 "unites": {
  "Go": "10^9 octets",
  "PFLOPS": "denses, sans parcimonie"
 },
 "champs": {
  "vramGB": "mémoire totale annoncée, Go (10^9 octets)",
  "bwTBs": "bande passante mémoire, To/s",
  "rates": "calcul dense en PFLOPS (bf16, fp8, fp4, int8) ; 0 = absent du matériel, null = non publié",
  "devices": "nombre d'accélérateurs",
  "memKind": "règle de mémoire utilisable",
  "powerW": "consommation maximale retenue pour le coût, W",
  "priceHT": "prix hors taxes retenu pour le coût, EUR (majoré de 30 % en octobre 2026)",
  "priceBasis": "base du prix",
  "flags": "machines écartées des recommandations"
 },
 "regles": {
  "memoire": {
   "utilGpu": 0.92,
   "reserveFixeGio": 2.5,
   "appleFraction": 0.75,
   "gb10Gio": 119.7,
   "strixGio": 96,
   "margePoids": 1
  },
  "etaBw": {
   "cuda": 0.82,
   "apple": 0.85,
   "amd": 0.8,
   "other": 0.8
  },
  "etaGemm": {
   "cuda": 0.4,
   "apple": 0.7,
   "amd": 0.4,
   "other": 0.4
  },
  "etaGemmMoeEp": 0.27,
  "etaAttn": 0.5,
  "tauUs": {
   "single": [
    20,
    45
   ],
   "nvlink": [
    40,
    140
   ],
   "ici": [
    40,
    75
   ],
   "pcie": [
    60,
    180
   ],
   "nop2p": [
    90,
    250
   ],
   "apple": [
    100,
    150
   ]
  },
  "hopMs": 1,
  "epsPar": {
   "single": 1,
   "nvlink": 1,
   "ici": 1,
   "pcie": 0.9,
   "nop2p": 0.8,
   "apple": 1,
   "rdma": 1
  },
  "cpMax": 4,
  "epsCp": 0.96,
  "rdma": {
   "hopUs": 50
  },
  "empilement": {
   "gb10": {
    "quantites": [
     1,
     2,
     3,
     4,
     8
    ],
    "tauUs": [
     300,
     265
    ],
    "lienGBs": {
     "2": 23.1,
     "3": 23.1,
     "4": 23.1,
     "8": 23.1
    },
    "reseau": {
     "2": {
      "cable": "cableSpark",
      "nb": 1
     },
     "3": {
      "commutateur": "crs804",
      "cable": "eclatement200",
      "nb": 2
     },
     "4": {
      "commutateur": "crs804",
      "cable": "eclatement200",
      "nb": 2
     },
     "8": {
      "commutateur": "crs804",
      "cable": "eclatement200",
      "nb": 4
     }
    }
   },
   "gb300": {
    "quantites": [
     1,
     2
    ],
    "tauUs": [
     150,
     135
    ],
    "lienGBs": {
     "2": 92.2
    },
    "reseau": {
     "2": {
      "cable": "dac400",
      "nb": 2
     }
    }
   }
  },
  "pieces": {
   "crs804": {
    "prixHT": 993,
    "powerW": 123
   },
   "cableSpark": {
    "prixHT": 89.43
   },
   "eclatement200": {
    "prixHT": 63.5
   },
   "dac400": {
    "prixHT": 54.86
   }
  },
  "kvBytes": {
   "4": 0.5625,
   "8": 1,
   "16": 2
  },
  "spec": {
   "drafts": 3,
   "acceptMtp": 2.4,
   "draftCost": 0.05
  },
  "familles": {
   "reference": 1,
   "natif": 1,
   "service": 0.77
  },
  "cout": {
   "kwhEurHt": 0.1624,
   "amortHeures": 26280,
   "tauxActivite": 0.3,
   "heuresMois": 730
  },
  "charge": {
   "partInvite": 0.8,
   "jetonsSortie": 512,
   "reflexionS": 5
  }
 },
 "machines": [
  {
   "id": "dell-gb10",
   "brand": "Dell",
   "name": "Dell Pro Precision with GB10 (anciennement Pro Max, DGX Spark)",
   "gpu": "NVIDIA GB10 Grace Blackwell",
   "category": "gb10",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "gb10",
   "tiers": null,
   "vramGB": 128,
   "bwTBs": 0.273,
   "rates": {
    "bf16": 0.125,
    "fp8": 0.25,
    "fp4": 0.5,
    "int8": 0.25
   },
   "fp4Type": "NVFP4",
   "powerW": 280,
   "powerBasis": "adaptateur Dell 280 W",
   "priceHT": 7400,
   "priceBasis": "relevé pi3g du 3 au 18/09/2026, configuration de stockage indiquée, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://docs.nvidia.com/dgx/dgx-spark/hardware.html"
  },
  {
   "id": "lenovo-gb10",
   "brand": "Lenovo",
   "name": "Lenovo ThinkStation PGX GB10",
   "gpu": "NVIDIA GB10 Grace Blackwell",
   "category": "gb10",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "gb10",
   "tiers": null,
   "vramGB": 128,
   "bwTBs": 0.273,
   "rates": {
    "bf16": 0.125,
    "fp8": 0.25,
    "fp4": 0.5,
    "int8": 0.25
   },
   "fp4Type": "NVFP4",
   "powerW": 240,
   "powerBasis": "bloc NVIDIA 240 W (hypothèse haute, comme le comparateur)",
   "priceHT": 6500,
   "priceBasis": "relevé pi3g du 3 au 18/09/2026, configuration de stockage indiquée, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://docs.nvidia.com/dgx/dgx-spark/hardware.html"
  },
  {
   "id": "asus-gb10",
   "brand": "ASUS",
   "name": "ASUS Ascent GX10",
   "gpu": "NVIDIA GB10 Grace Blackwell",
   "category": "gb10",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "gb10",
   "tiers": null,
   "vramGB": 128,
   "bwTBs": 0.273,
   "rates": {
    "bf16": 0.125,
    "fp8": 0.25,
    "fp4": 0.5,
    "int8": 0.25
   },
   "fp4Type": "NVFP4",
   "powerW": 240,
   "powerBasis": "bloc NVIDIA 240 W (hypothèse haute, comme le comparateur)",
   "priceHT": 8300,
   "priceBasis": "relevé pi3g du 3 au 18/09/2026, configuration de stockage indiquée, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://docs.nvidia.com/dgx/dgx-spark/hardware.html"
  },
  {
   "id": "hp-gb10",
   "brand": "HP",
   "name": "HP ZGX Nano G1n AI Station",
   "gpu": "NVIDIA GB10 Grace Blackwell",
   "category": "gb10",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "gb10",
   "tiers": null,
   "vramGB": 128,
   "bwTBs": 0.273,
   "rates": {
    "bf16": 0.125,
    "fp8": 0.25,
    "fp4": 0.5,
    "int8": 0.25
   },
   "fp4Type": "NVFP4",
   "powerW": 240,
   "powerBasis": "bloc NVIDIA 240 W (hypothèse haute, comme le comparateur)",
   "priceHT": 7100,
   "priceBasis": "relevé pi3g du 3 au 18/09/2026, configuration de stockage indiquée, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://docs.nvidia.com/dgx/dgx-spark/hardware.html"
  },
  {
   "id": "gigabyte-gb10",
   "brand": "GIGABYTE",
   "name": "Gigabyte AI TOP ATOM",
   "gpu": "NVIDIA GB10 Grace Blackwell",
   "category": "gb10",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "gb10",
   "tiers": null,
   "vramGB": 128,
   "bwTBs": 0.273,
   "rates": {
    "bf16": 0.125,
    "fp8": 0.25,
    "fp4": 0.5,
    "int8": 0.25
   },
   "fp4Type": "NVFP4",
   "powerW": 240,
   "powerBasis": "bloc NVIDIA 240 W (hypothèse haute, comme le comparateur)",
   "priceHT": 7100,
   "priceBasis": "relevé pi3g du 3 au 18/09/2026, configuration de stockage indiquée, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-spark/"
  },
  {
   "id": "supermicro-gb300",
   "brand": "Supermicro",
   "name": "Supermicro Super AI Station (ARS-511GD-NB-LCC)",
   "gpu": "NVIDIA GB300 Blackwell Ultra Desktop",
   "category": "gb300",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 252,
    "fastTBs": 7.1,
    "slowGB": 496,
    "slowTBs": 0.396
   },
   "vramGB": 748,
   "bwTBs": 7.1,
   "rates": {
    "bf16": 2.5,
    "fp8": 5,
    "fp4": 15,
    "int8": 0.1875
   },
   "fp4Type": "NVFP4",
   "powerW": 1600,
   "powerBasis": "fiche NVIDIA : 1 600 W",
   "priceHT": 114000,
   "priceBasis": "relevé pi3g du 14/09/2026, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-station/"
  },
  {
   "id": "msi-gb300",
   "brand": "MSI",
   "name": "MSI XpertStation WS300 (anciennement CT60-S8060)",
   "gpu": "NVIDIA GB300 Blackwell Ultra Desktop",
   "category": "gb300",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 252,
    "fastTBs": 7.1,
    "slowGB": 496,
    "slowTBs": 0.396
   },
   "vramGB": 748,
   "bwTBs": 7.1,
   "rates": {
    "bf16": 2.5,
    "fp8": 5,
    "fp4": 15,
    "int8": 0.1875
   },
   "fp4Type": "NVFP4",
   "powerW": 1600,
   "powerBasis": "fiche NVIDIA : 1 600 W",
   "priceHT": 117000,
   "priceBasis": "relevé pi3g du 14/09/2026, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-station/"
  },
  {
   "id": "exxact-gb300",
   "brand": "Exxact",
   "name": "Exxact Valence VWS-158270643 (intégrateur, base MSI)",
   "gpu": "NVIDIA GB300 Blackwell Ultra Desktop",
   "category": "gb300",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 252,
    "fastTBs": 7.1,
    "slowGB": 496,
    "slowTBs": 0.396
   },
   "vramGB": 748,
   "bwTBs": 7.1,
   "rates": {
    "bf16": 2.5,
    "fp8": 5,
    "fp4": 15,
    "int8": 0.1875
   },
   "fp4Type": "NVFP4",
   "powerW": 1600,
   "powerBasis": "fiche NVIDIA : 1 600 W",
   "priceHT": 124000,
   "priceBasis": "relevé pi3g du 14/09/2026, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-station/"
  },
  {
   "id": "gigabyte-gb300",
   "brand": "GIGABYTE",
   "name": "Gigabyte W775-V10-L01",
   "gpu": "NVIDIA GB300 Blackwell Ultra Desktop",
   "category": "gb300",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 252,
    "fastTBs": 7.1,
    "slowGB": 496,
    "slowTBs": 0.396
   },
   "vramGB": 748,
   "bwTBs": 7.1,
   "rates": {
    "bf16": 2.5,
    "fp8": 5,
    "fp4": 15,
    "int8": 0.1875
   },
   "fp4Type": "NVFP4",
   "powerW": 1600,
   "powerBasis": "fiche NVIDIA : 1 600 W",
   "priceHT": 129000,
   "priceBasis": "relevé pi3g du 14/09/2026, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-station/"
  },
  {
   "id": "hp-gb300",
   "brand": "HP",
   "name": "HP ZGX Fury G1n",
   "gpu": "NVIDIA GB300 Blackwell Ultra Desktop",
   "category": "gb300",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 252,
    "fastTBs": 7.1,
    "slowGB": 496,
    "slowTBs": 0.396
   },
   "vramGB": 748,
   "bwTBs": 7.1,
   "rates": {
    "bf16": 2.5,
    "fp8": 5,
    "fp4": 15,
    "int8": 0.1875
   },
   "fp4Type": "NVFP4",
   "powerW": 1600,
   "powerBasis": "fiche NVIDIA : 1 600 W",
   "priceHT": 143000,
   "priceBasis": "relevé pi3g du 14/09/2026, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-station/"
  },
  {
   "id": "asus-gb300",
   "brand": "ASUS",
   "name": "ASUS ExpertCenter Pro ET900N G3",
   "gpu": "NVIDIA GB300 Blackwell Ultra Desktop",
   "category": "gb300",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 252,
    "fastTBs": 7.1,
    "slowGB": 496,
    "slowTBs": 0.396
   },
   "vramGB": 748,
   "bwTBs": 7.1,
   "rates": {
    "bf16": 2.5,
    "fp8": 5,
    "fp4": 15,
    "int8": 0.1875
   },
   "fp4Type": "NVFP4",
   "powerW": 1600,
   "powerBasis": "fiche NVIDIA : 1 600 W",
   "priceHT": 144000,
   "priceBasis": "relevé pi3g du 14/09/2026, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-station/"
  },
  {
   "id": "dell-gb300",
   "brand": "Dell",
   "name": "Dell Pro Precision with GB300 (anciennement Pro Max, FCT6263)",
   "gpu": "NVIDIA GB300 Blackwell Ultra Desktop",
   "category": "gb300",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 252,
    "fastTBs": 7.1,
    "slowGB": 496,
    "slowTBs": 0.396
   },
   "vramGB": 748,
   "bwTBs": 7.1,
   "rates": {
    "bf16": 2.5,
    "fp8": 5,
    "fp4": 15,
    "int8": 0.1875
   },
   "fp4Type": "NVFP4",
   "powerW": 1600,
   "powerBasis": "fiche NVIDIA : 1 600 W",
   "priceHT": 225000,
   "priceBasis": "relevé pi3g du 14/09/2026, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-station/"
  },
  {
   "id": "bizon-gb300",
   "brand": "BIZON",
   "name": "BIZON GB300 (intégrateur)",
   "gpu": "NVIDIA GB300 Blackwell Ultra Desktop",
   "category": "gb300",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 252,
    "fastTBs": 7.1,
    "slowGB": 496,
    "slowTBs": 0.396
   },
   "vramGB": 748,
   "bwTBs": 7.1,
   "rates": {
    "bf16": 2.5,
    "fp8": 5,
    "fp4": 15,
    "int8": 0.1875
   },
   "fp4Type": "NVFP4",
   "powerW": 1600,
   "powerBasis": "fiche NVIDIA : 1 600 W",
   "priceHT": 128000,
   "priceBasis": "relevé en $ converti (BCE 25/09/2026), majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/dgx-station/"
  },
  {
   "id": "nvidia-dgx-b200",
   "brand": "NVIDIA",
   "name": "NVIDIA DGX B200 (8x B200)",
   "gpu": "8x NVIDIA Blackwell B200 SXM",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 1440,
   "bwTBs": 64,
   "rates": {
    "bf16": 18,
    "fp8": 36,
    "fp4": 72,
    "int8": 36
   },
   "fp4Type": "NVFP4",
   "powerW": 14300,
   "powerBasis": "NVIDIA : ~14,3 kW max",
   "priceHT": 527000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/dgx-b200/"
  },
  {
   "id": "nvidia-dgx-b300",
   "brand": "NVIDIA",
   "name": "NVIDIA DGX B300 (8x B300)",
   "gpu": "8x NVIDIA Blackwell Ultra B300 SXM",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 2304,
   "bwTBs": 64,
   "rates": {
    "bf16": 18,
    "fp8": 36,
    "fp4": 108,
    "int8": 1.5
   },
   "fp4Type": "NVFP4",
   "powerW": 15000,
   "powerBasis": "guide NVIDIA : 15 kW max",
   "priceHT": 736000,
   "priceBasis": "milieu de fourchette en $ converti (BCE 25/09/2026), majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/dgx-b300/"
  },
  {
   "id": "dell-xe9680l-b200",
   "brand": "Dell",
   "name": "Dell PowerEdge XE9680L (8x B200 DLC)",
   "gpu": "8x NVIDIA Blackwell B200 SXM",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 1440,
   "bwTBs": 64,
   "rates": {
    "bf16": 18,
    "fp8": 36,
    "fp4": 72,
    "int8": 36
   },
   "fp4Type": "NVFP4",
   "powerW": 14300,
   "powerBasis": "estimation (classe DGX B200)",
   "priceHT": 494000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/hgx/"
  },
  {
   "id": "nvidia-dgx-h200",
   "brand": "NVIDIA",
   "name": "NVIDIA DGX H200 (8x H200)",
   "gpu": "8x NVIDIA H200 SXM5",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 1128,
   "bwTBs": 38.4,
   "rates": {
    "bf16": 7.916,
    "fp8": 15.832,
    "fp4": 0,
    "int8": 15.832
   },
   "fp4Type": null,
   "powerW": 10200,
   "powerBasis": "guide DGX H100/H200 : 10,2 kW max",
   "priceHT": 429000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/h200/"
  },
  {
   "id": "dell-xe9680-h100",
   "brand": "Dell",
   "name": "Dell PowerEdge XE9680 (8x H100 SXM5)",
   "gpu": "8x NVIDIA H100 SXM5",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 640,
   "bwTBs": 26.8,
   "rates": {
    "bf16": 7.916,
    "fp8": 15.832,
    "fp4": 0,
    "int8": 15.832
   },
   "fp4Type": null,
   "powerW": 10200,
   "powerBasis": "calcul : référence DGX H100 (10,2 kW max)",
   "priceHT": 361000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/h100/"
  },
  {
   "id": "supermicro-821ge-h100",
   "brand": "Supermicro",
   "name": "Supermicro SYS-821GE-TNHR (8x H100)",
   "gpu": "8x NVIDIA H100 SXM5",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 640,
   "bwTBs": 26.8,
   "rates": {
    "bf16": 7.916,
    "fp8": 15.832,
    "fp4": 0,
    "int8": 15.832
   },
   "fp4Type": null,
   "powerW": 10200,
   "powerBasis": "calcul : référence DGX H100 (10,2 kW max)",
   "priceHT": 335000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/h100/"
  },
  {
   "id": "lenovo-sr675-v3",
   "brand": "Lenovo",
   "name": "Lenovo ThinkSystem SR675 V3 (4x H100 SXM5, configuration retirée)",
   "gpu": "4x NVIDIA H100 SXM5",
   "category": "hgx",
   "devices": 4,
   "tpMax": 4,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 320,
   "bwTBs": 13.4,
   "rates": {
    "bf16": 3.958,
    "fp8": 7.916,
    "fp4": 0,
    "int8": 7.916
   },
   "fp4Type": null,
   "powerW": 3600,
   "powerBasis": "calcul : 4 x 700 W + 800 W plate-forme",
   "priceHT": null,
   "priceBasis": null,
   "flags": [
    "retire"
   ],
   "compute_source": "https://lenovopress.lenovo.com/lp1611-thinksystem-sr675-v3-server"
  },
  {
   "id": "gigabyte-g593-sd0",
   "brand": "GIGABYTE",
   "name": "Gigabyte G593-SD0 (8x H100 HGX)",
   "gpu": "8x NVIDIA H100 SXM5",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 640,
   "bwTBs": 26.8,
   "rates": {
    "bf16": 7.916,
    "fp8": 15.832,
    "fp4": 0,
    "int8": 15.832
   },
   "fp4Type": null,
   "powerW": 10200,
   "powerBasis": "calcul : référence DGX H100 (10,2 kW max)",
   "priceHT": 325000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [
    "legacy"
   ],
   "compute_source": "https://www.nvidia.com/en-us/data-center/h100/"
  },
  {
   "id": "asus-esc-n8",
   "brand": "ASUS",
   "name": "ASUS ESC N8-E11 (8x H100 HGX)",
   "gpu": "8x NVIDIA H100 SXM5",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 640,
   "bwTBs": 26.8,
   "rates": {
    "bf16": 7.916,
    "fp8": 15.832,
    "fp4": 0,
    "int8": 15.832
   },
   "fp4Type": null,
   "powerW": 10200,
   "powerBasis": "calcul : référence DGX H100 (10,2 kW max)",
   "priceHT": 319000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [
    "eol"
   ],
   "compute_source": "https://www.nvidia.com/en-us/data-center/h100/"
  },
  {
   "id": "dell-xe9680-mi300x",
   "brand": "Dell",
   "name": "Dell PowerEdge XE9680 (8x AMD MI300X)",
   "gpu": "8x AMD Instinct MI300X OAM",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 1536,
   "bwTBs": 42.4,
   "rates": {
    "bf16": 10.46,
    "fp8": 20.92,
    "fp4": 0,
    "int8": 20.92
   },
   "fp4Type": null,
   "powerW": 10600,
   "powerBasis": "calcul : 8 x 750 W + 4 600 W plate-forme HGX (DGX H100 10,2 kW - 8 x 700 W)",
   "priceHT": 286000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/data-sheets/amd-instinct-mi300x-data-sheet.pdf"
  },
  {
   "id": "supermicro-mi300x",
   "brand": "Supermicro",
   "name": "Supermicro AS-8125GS-TNMR2 (8x MI300X)",
   "gpu": "8x AMD Instinct MI300X OAM",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 1536,
   "bwTBs": 42.4,
   "rates": {
    "bf16": 10.46,
    "fp8": 20.92,
    "fp4": 0,
    "int8": 20.92
   },
   "fp4Type": null,
   "powerW": 10600,
   "powerBasis": "calcul : 8 x 750 W + 4 600 W plate-forme HGX (DGX H100 10,2 kW - 8 x 700 W)",
   "priceHT": 277000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/data-sheets/amd-instinct-mi300x-data-sheet.pdf"
  },
  {
   "id": "dell-xe9680-mi325x",
   "brand": "Supermicro",
   "name": "Supermicro AS-8126GS-TNMR (8x AMD MI325X)",
   "gpu": "8x AMD Instinct MI325X OAM (256 Go HBM3e chacun)",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 2048,
   "bwTBs": 48,
   "rates": {
    "bf16": 10.46,
    "fp8": 20.92,
    "fp4": 0,
    "int8": 20.92
   },
   "fp4Type": null,
   "powerW": 12600,
   "powerBasis": "calcul : 8 x 1 000 W + 4 600 W plate-forme HGX",
   "priceHT": 335000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [
    "non-verifie"
   ],
   "compute_source": "https://www.amd.com/en/products/accelerators/instinct/mi300/mi325x.html"
  },
  {
   "id": "hpe-ex235a",
   "brand": "HPE",
   "name": "HPE Cray EX255a (El Capitan)",
   "gpu": "4x AMD Instinct MI300A par nœud (2 nœuds et 8 APU par lame)",
   "category": "hgx",
   "devices": 4,
   "tpMax": 4,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 512,
   "bwTBs": 21.2,
   "rates": {
    "bf16": 3.922,
    "fp8": 7.84,
    "fp4": 0,
    "int8": 7.84
   },
   "fp4Type": null,
   "powerW": 3000,
   "powerBasis": "calcul : 4 x 550 W + 800 W plate-forme",
   "priceHT": null,
   "priceBasis": null,
   "flags": [
    "prix-non-source"
   ],
   "compute_source": "https://www.amd.com/en/products/accelerators/instinct/mi300/mi300a.html"
  },
  {
   "id": "nvidia-gb200-nvl72",
   "brand": "NVIDIA",
   "name": "NVIDIA GB200 NVL72 (baie complète, 72 GPU)",
   "gpu": "72x GPU Blackwell B200",
   "category": "rack",
   "devices": 72,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 13392,
   "bwTBs": 576,
   "rates": {
    "bf16": 180,
    "fp8": 360,
    "fp4": 720,
    "int8": 360
   },
   "fp4Type": "NVFP4",
   "powerW": 132000,
   "powerBasis": "NVIDIA : ~132 kW",
   "priceHT": 4225000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/gb200-nvl72/"
  },
  {
   "id": "lenovo-sc777-v4",
   "brand": "Lenovo",
   "name": "Lenovo ThinkSystem SC777 V4 Neptune (GB200 NVL4)",
   "gpu": "4x NVIDIA B200 (GB200 NVL4, 186 Go chacun)",
   "category": "rack",
   "devices": 4,
   "tpMax": 4,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 744,
   "bwTBs": 32,
   "rates": {
    "bf16": 10,
    "fp8": 20,
    "fp4": 40,
    "int8": 20
   },
   "fp4Type": "NVFP4",
   "powerW": 5800,
   "powerBasis": "calcul : carte GB200 NVL4 5 480 W (Lenovo LP2048) + 320 W réseau et stockage",
   "priceHT": 393000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://lenovopress.lenovo.com/lp2048-thinksystem-sc777-v4"
  },
  {
   "id": "nvidia-dgx-a100-80",
   "brand": "NVIDIA",
   "name": "NVIDIA DGX A100 640 Go (8x A100 80 Go)",
   "gpu": "8x NVIDIA A100 SXM4 80 Go",
   "category": "legacy",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 640,
   "bwTBs": 16.3,
   "rates": {
    "bf16": 2.496,
    "fp8": 0,
    "fp4": 0,
    "int8": 4.992
   },
   "fp4Type": null,
   "powerW": 6500,
   "powerBasis": "fiche NVIDIA : 6,5 kW max",
   "priceHT": 150000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/nvidia-dgx-a100-datasheet.pdf"
  },
  {
   "id": "dell-r750xa-a100",
   "brand": "Dell",
   "name": "Dell PowerEdge R750xa (4x A100 PCIe)",
   "gpu": "4x NVIDIA A100 PCIe 80 Go",
   "category": "legacy",
   "devices": 4,
   "tpMax": 4,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 320,
   "bwTBs": 7.74,
   "rates": {
    "bf16": 1.248,
    "fp8": 0,
    "fp4": 0,
    "int8": 2.496
   },
   "fp4Type": null,
   "powerW": 2000,
   "powerBasis": "calcul : 4 x 300 W + 800 W plate-forme",
   "priceHT": 59000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a100/pdf/nvidia-a100-datasheet-us-nvidia-1758950-r4-web.pdf"
  },
  {
   "id": "workstation-4x-rtx6000",
   "brand": "Exxact",
   "name": "Station de travail 4x NVIDIA RTX 6000 Ada",
   "gpu": "4x NVIDIA RTX 6000 Ada Generation (PCIe)",
   "category": "ada",
   "devices": 4,
   "tpMax": 4,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 192,
   "bwTBs": 3.84,
   "rates": {
    "bf16": 1.457,
    "fp8": 2.914,
    "fp4": 0,
    "int8": 2.914
   },
   "fp4Type": null,
   "powerW": 1550,
   "powerBasis": "calcul : 4 x 300 W + 350 W plate-forme",
   "priceHT": 55000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/rtx-6000/proviz-print-rtx6000-datasheet-web-2504660.pdf"
  },
  {
   "id": "station-homelab-rtx5090",
   "brand": "Assemblage sur mesure",
   "name": "Station RTX 5090 (assemblage sur mesure)",
   "gpu": "1x NVIDIA GeForce RTX 5090 (Blackwell GB202)",
   "category": "homelab",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 32,
   "bwTBs": 1.792,
   "rates": {
    "bf16": 0.2095,
    "fp8": 0.419,
    "fp4": 1.676,
    "int8": 0.838,
    "bf16acc16": 0.419,
    "fp8acc16": 0.838
   },
   "fp4Type": "NVFP4",
   "powerW": 925,
   "powerBasis": "calcul : 575 W TGP + 350 W plate-forme",
   "priceHT": 6500,
   "priceBasis": "milieu de la fourchette communiquée par QDNA en octobre 2026 (hausse constatée du matériel)",
   "flags": [],
   "compute_source": "https://images.nvidia.com/aem-dam/Solutions/geforce/blackwell/nvidia-rtx-blackwell-gpu-architecture.pdf"
  },
  {
   "id": "apple-mac-studio-m5-ultra-512",
   "brand": "Apple",
   "name": "Apple Mac Studio M5 Ultra (512 Go unifiée, 1 200 Go/s)",
   "gpu": "Apple M5 Ultra, 80 cœurs GPU avec accélérateurs neuronaux",
   "category": "apple",
   "devices": 1,
   "tpMax": 1,
   "link": "apple",
   "memKind": "apple",
   "tiers": null,
   "vramGB": 512,
   "bwTBs": 1.2,
   "rates": {
    "bf16": 0.1,
    "fp8": 0,
    "fp4": 0,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 385,
   "powerBasis": "Apple (support 102027) : 385 W de consommation maximale ; 480 W = puissance continue maximale de l alimentation",
   "priceHT": null,
   "priceBasis": null,
   "flags": [],
   "compute_source": "https://www.apple.com/mac-studio/specs/"
  },
  {
   "id": "apple-mac-studio-m5-ultra-256",
   "brand": "Apple",
   "name": "Apple Mac Studio M5 Ultra (256 Go unifiée, 1 200 Go/s)",
   "gpu": "Apple M5 Ultra, 80 cœurs GPU avec accélérateurs neuronaux",
   "category": "apple",
   "devices": 1,
   "tpMax": 1,
   "link": "apple",
   "memKind": "apple",
   "tiers": null,
   "vramGB": 256,
   "bwTBs": 1.2,
   "rates": {
    "bf16": 0.1,
    "fp8": 0,
    "fp4": 0,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 385,
   "powerBasis": "Apple (support 102027) : 385 W de consommation maximale ; 480 W = puissance continue maximale de l alimentation",
   "priceHT": 10358,
   "priceBasis": "Apple Store France, relevé le 8 octobre 2026, TTC divisé par 1,2",
   "flags": [],
   "compute_source": "https://www.apple.com/mac-studio/specs/"
  },
  {
   "id": "apple-mac-studio-m5-ultra-96",
   "brand": "Apple",
   "name": "Apple Mac Studio M5 Ultra (96 Go unifiée, 1 200 Go/s)",
   "gpu": "Apple M5 Ultra, 64 cœurs GPU avec accélérateurs neuronaux (puce de base)",
   "category": "apple",
   "devices": 1,
   "tpMax": 1,
   "link": "apple",
   "memKind": "apple",
   "tiers": null,
   "vramGB": 96,
   "bwTBs": 1.2,
   "rates": {
    "bf16": 0.08,
    "fp8": 0,
    "fp4": 0,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 385,
   "powerBasis": "Apple (support 102027) : 385 W de consommation maximale ; 480 W = puissance continue maximale de l alimentation",
   "priceHT": 5499,
   "priceBasis": "Apple Store France, relevé le 8 octobre 2026, TTC divisé par 1,2",
   "flags": [],
   "compute_source": "https://www.apple.com/mac-studio/specs/"
  },
  {
   "id": "apple-mac-studio-m4-ultra",
   "brand": "Apple",
   "name": "Apple Mac Studio M3 Ultra (256 Go unifiée, 2025)",
   "gpu": "Apple M3 Ultra, 80 cœurs GPU",
   "category": "apple",
   "devices": 1,
   "tpMax": 1,
   "link": "apple",
   "memKind": "apple",
   "tiers": null,
   "vramGB": 256,
   "bwTBs": 0.819,
   "rates": {
    "bf16": 0.0296,
    "fp8": 0,
    "fp4": 0,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 270,
   "powerBasis": "Apple (support 102027) : 270 W de consommation maximale",
   "priceHT": null,
   "priceBasis": null,
   "flags": [
    "renommer-m3-ultra",
    "prix-a-resourcer"
   ],
   "compute_source": "https://www.apple.com/newsroom/2025/03/apple-unveils-new-mac-studio-the-most-powerful-mac-ever/"
  },
  {
   "id": "apple-mac-studio-m2-ultra",
   "brand": "Apple",
   "name": "Apple Mac Studio M2 Ultra (192 Go unifiée, occasion)",
   "gpu": "Apple M2 Ultra, 76 cœurs GPU",
   "category": "apple",
   "devices": 1,
   "tpMax": 1,
   "link": "apple",
   "memKind": "apple",
   "tiers": null,
   "vramGB": 192,
   "bwTBs": 0.8,
   "rates": {
    "bf16": 0.0272,
    "fp8": 0,
    "fp4": 0,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 295,
   "powerBasis": "Apple (support 102027) : 295 W de consommation maximale",
   "priceHT": 4200,
   "priceBasis": "TTC, régime de la marge (TVA non récupérable)",
   "flags": [
    "occasion-regime-marge"
   ],
   "compute_source": "https://support.apple.com/en-us/111835"
  },
  {
   "id": "cluster-dual-mac-studio",
   "brand": "Apple",
   "name": "Grappe de 2 Mac Studio M2 Ultra (384 Go, exo ou llama.cpp RPC)",
   "gpu": "2x Apple M2 Ultra 192 Go (76 cœurs GPU chacun)",
   "category": "apple",
   "devices": 2,
   "tpMax": 1,
   "link": "apple",
   "memKind": "apple",
   "tiers": null,
   "vramGB": 384,
   "bwTBs": 1.6,
   "rates": {
    "bf16": 0.0544,
    "fp8": 0,
    "fp4": 0,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 590,
   "powerBasis": "Apple (support 102027) : 2 x 295 W",
   "priceHT": 8400,
   "priceBasis": "2 x 4 200 € TTC, régime de la marge (TVA non récupérable)",
   "flags": [
    "occasion-regime-marge"
   ],
   "compute_source": "https://www.apple.com/mac-studio/specs/"
  },
  {
   "id": "workstation-2x-rtx4090",
   "brand": "Assemblage sur mesure",
   "name": "Station IA 2x NVIDIA RTX 4090 (48 Go)",
   "gpu": "2x NVIDIA GeForce RTX 4090 (Ada Lovelace)",
   "category": "homelab",
   "devices": 2,
   "tpMax": 2,
   "link": "nop2p",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 48,
   "bwTBs": 2.016,
   "rates": {
    "bf16": 0.3304,
    "fp8": 0.6606,
    "fp4": 0,
    "int8": 1.3212,
    "bf16acc16": 0.6606,
    "fp8acc16": 1.3212
   },
   "fp4Type": null,
   "powerW": 1250,
   "powerBasis": "calcul : 2 x 450 W + 350 W plate-forme",
   "priceHT": 4167,
   "priceBasis": "milieu de fourchette TTC / 1,2",
   "flags": [],
   "compute_source": "https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf"
  },
  {
   "id": "workstation-4x-rtx4090",
   "brand": "Assemblage sur mesure",
   "name": "Station IA 4x NVIDIA RTX 4090 (96 Go)",
   "gpu": "4x NVIDIA GeForce RTX 4090 (Ada Lovelace)",
   "category": "homelab",
   "devices": 4,
   "tpMax": 4,
   "link": "nop2p",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 96,
   "bwTBs": 4.032,
   "rates": {
    "bf16": 0.6608,
    "fp8": 1.3212,
    "fp4": 0,
    "int8": 2.6424,
    "bf16acc16": 1.3212,
    "fp8acc16": 2.6424
   },
   "fp4Type": null,
   "powerW": 2150,
   "powerBasis": "calcul : 4 x 450 W + 350 W plate-forme",
   "priceHT": 9167,
   "priceBasis": "milieu de fourchette TTC / 1,2",
   "flags": [],
   "compute_source": "https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf"
  },
  {
   "id": "amd-ryzen-ai-max-395",
   "brand": "AMD",
   "name": "Mini-PC IA AMD Ryzen AI Max+ 395 (Strix Halo 128 Go)",
   "gpu": "AMD Radeon 8060S (40 unités de calcul RDNA 3.5)",
   "category": "homelab",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "strix",
   "tiers": null,
   "vramGB": 128,
   "bwTBs": 0.256,
   "rates": {
    "bf16": 0.06,
    "fp8": 0,
    "fp4": 0,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 180,
   "powerBasis": "bloc externe 180 W",
   "priceHT": 3333,
   "priceBasis": "relevé du 8 octobre 2026 : médiane des mini-PC 128 Go vendus en euros, TTC divisé par 1,2",
   "flags": [],
   "compute_source": "https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html"
  },
  {
   "id": "tenstorrent-wormhole-n300",
   "brand": "Tenstorrent",
   "name": "Station Tenstorrent Wormhole n300 (carte n300 et hôte)",
   "gpu": "Carte n300 : 2 puces Wormhole à cœurs Tensix (RISC-V)",
   "category": "homelab",
   "devices": 2,
   "tpMax": 2,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 24,
   "bwTBs": 0.576,
   "rates": {
    "bf16": 0.131,
    "fp8": 0.466,
    "fp4": null,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 650,
   "powerBasis": "calcul : 300 W carte + 350 W hôte",
   "priceHT": 4300,
   "priceBasis": "station complète HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://docs.tenstorrent.com/aibs/wormhole/specifications.html"
  },
  {
   "id": "nvidia-gb300-nvl72",
   "brand": "NVIDIA",
   "name": "NVIDIA GB300 NVL72 (baie complète, 72 GPU Blackwell Ultra)",
   "gpu": "72x NVIDIA Blackwell Ultra B300 (288 Go HBM3e)",
   "category": "rack",
   "devices": 72,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 20736,
   "bwTBs": 576,
   "rates": {
    "bf16": 180,
    "fp8": 360,
    "fp4": 1080,
    "int8": 13.5
   },
   "fp4Type": "NVFP4",
   "powerW": 135000,
   "powerBasis": "NVIDIA : ~135 kW",
   "priceHT": 4745000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/gb300-nvl72/"
  },
  {
   "id": "amd-helios-mi455x",
   "brand": "AMD",
   "name": "AMD Helios (baie complète, 72 GPU Instinct MI455X)",
   "gpu": "72x AMD Instinct MI455X (432 Go HBM4)",
   "category": "rack",
   "devices": 72,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 31104,
   "bwTBs": 1677.6,
   "rates": {
    "bf16": 720,
    "fp8": 1440,
    "fp4": 2880,
    "int8": 1440
   },
   "fp4Type": "MXFP4",
   "powerW": null,
   "powerBasis": "non publiée par AMD",
   "priceHT": null,
   "priceBasis": null,
   "flags": [
    "prix-non-source"
   ],
   "compute_source": "https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html"
  },
  {
   "id": "amd-mi355x-8gpu",
   "brand": "AMD",
   "name": "Serveur 8x AMD Instinct MI355X OAM (CDNA 4)",
   "gpu": "8x AMD Instinct MI355X OAM (CDNA 4)",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 2304,
   "bwTBs": 64,
   "rates": {
    "bf16": 20,
    "fp8": 40,
    "fp4": 80.8,
    "int8": 40
   },
   "fp4Type": "MXFP4",
   "powerW": 15800,
   "powerBasis": "calcul : 8 x 1 400 W + 4 600 W plate-forme HGX",
   "priceHT": 380000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html"
  },
  {
   "id": "google-tpu-v7-pod",
   "brand": "Google",
   "name": "Pod Google TPU7x Ironwood (tranche de 64 puces)",
   "gpu": "64x puces Google TPU7x Ironwood",
   "category": "rack",
   "devices": 64,
   "tpMax": 4,
   "link": "ici",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 13194,
   "bwTBs": 472.3,
   "rates": {
    "bf16": 147.6,
    "fp8": 295.3,
    "fp4": 0,
    "int8": 295.3
   },
   "fp4Type": null,
   "powerW": null,
   "powerBasis": "non publié",
   "priceHT": null,
   "priceBasis": "estimation SemiAnalysis de 2,00 $ par puce-heure en location, convertie au taux BCE de 1,1403 ; Google ne publie pas de tarif TPU7x",
   "flags": [
    "location"
   ],
   "compute_source": "https://docs.cloud.google.com/tpu/docs/tpu7x"
  },
  {
   "id": "sambanova-sn50-rdu",
   "brand": "SambaNova",
   "name": "SambaNova SN50 (nœud de 8 RDU)",
   "gpu": "8x RDU SambaNova SN50 (processeurs à flux de données reconfigurables)",
   "category": "enterprise",
   "devices": 8,
   "tpMax": 8,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 512,
   "bwTBs": 14.7,
   "rates": {
    "bf16": 12.8,
    "fp8": 25.6,
    "fp4": 0,
    "int8": 0
   },
   "fp4Type": null,
   "powerW": 10000,
   "powerBasis": "calcul : SambaRack 20 kW / 16 puces x 8",
   "priceHT": 263000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [
    "bp-estimee"
   ],
   "compute_source": "https://sambanova.ai/products/rdu-ai-chips"
  },
  {
   "id": "dell-r760xa-h100-pcie",
   "brand": "Dell",
   "name": "Dell PowerEdge R760xa (4x H100 PCIe)",
   "gpu": "4x NVIDIA H100 PCIe 80 Go (350 W)",
   "category": "hgx",
   "devices": 4,
   "tpMax": 4,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 320,
   "bwTBs": 8,
   "rates": {
    "bf16": 3.026,
    "fp8": 6.052,
    "fp4": 0,
    "int8": 6.052
   },
   "fp4Type": null,
   "powerW": 2200,
   "powerBasis": "calcul : 4 x 350 W + 800 W plate-forme",
   "priceHT": 179000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.megware.com/fileadmin/user_upload/LandingPage%20NVIDIA/nvidia-h100-datasheet.pdf"
  },
  {
   "id": "supermicro-as-4125gs-tnrt",
   "brand": "Supermicro",
   "name": "Supermicro AS-4125GS-TNRT1 (8x H100 PCIe)",
   "gpu": "8x NVIDIA H100 PCIe 80 Go (350 W)",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 640,
   "bwTBs": 16,
   "rates": {
    "bf16": 6.052,
    "fp8": 12.104,
    "fp4": 0,
    "int8": 12.104
   },
   "fp4Type": null,
   "powerW": 4300,
   "powerBasis": "calcul : 8 x 350 W + 1 500 W plate-forme PCIe",
   "priceHT": 293000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.megware.com/fileadmin/user_upload/LandingPage%20NVIDIA/nvidia-h100-datasheet.pdf"
  },
  {
   "id": "nvidia-b200-nvl-pcie",
   "brand": "NVIDIA",
   "name": "2x NVIDIA H200 NVL (paire PCIe avec pont NVLink)",
   "gpu": "2x NVIDIA H200 NVL PCIe (141 Go HBM3e chacune)",
   "category": "hgx",
   "devices": 2,
   "tpMax": 2,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 282,
   "bwTBs": 9.6,
   "rates": {
    "bf16": 1.671,
    "fp8": 3.341,
    "fp4": 0,
    "int8": 3.341
   },
   "fp4Type": null,
   "powerW": 2000,
   "powerBasis": "calcul : 2 x 600 W + 800 W plate-forme",
   "priceHT": null,
   "priceBasis": null,
   "flags": [
    "remplace-h200-nvl",
    "prix-a-resourcer"
   ],
   "compute_source": "https://www.nvidia.com/en-us/data-center/hgx/"
  },
  {
   "id": "workstation-1x-rtx-pro-6000",
   "brand": "NVIDIA",
   "name": "Station 1x NVIDIA RTX PRO 6000 Blackwell (96 Go)",
   "gpu": "1x NVIDIA RTX PRO 6000 Blackwell (GB202)",
   "category": "workstation",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 96,
   "bwTBs": 1.792,
   "rates": {
    "bf16": 0.5038,
    "fp8": 1.0076,
    "fp4": 2.0152,
    "int8": 1.0076
   },
   "fp4Type": "NVFP4",
   "powerW": 950,
   "powerBasis": "calcul : 600 W + 350 W plate-forme",
   "priceHT": 23000,
   "priceBasis": "station configurée HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/quadro-product-literature/NVIDIA-RTX-Blackwell-PRO-GPU-Architecture-v1.0.pdf"
  },
  {
   "id": "rtx-pro-6000-maxq",
   "brand": "NVIDIA",
   "name": "NVIDIA RTX PRO 6000 Blackwell Max-Q Edition (96 Go)",
   "gpu": "1x NVIDIA RTX PRO 6000 Blackwell Max-Q (GB202)",
   "category": "workstation",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 96,
   "bwTBs": 1.792,
   "rates": {
    "bf16": 0.439,
    "fp8": 0.8779,
    "fp4": 1.7557,
    "int8": 0.878
   },
   "fp4Type": "NVFP4",
   "powerW": 650,
   "powerBasis": "calcul : 300 W + 350 W plate-forme",
   "priceHT": 20000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/quadro-product-literature/NVIDIA-RTX-Blackwell-PRO-GPU-Architecture-v1.0.pdf"
  },
  {
   "id": "server-4x-rtx-pro-6000",
   "brand": "Supermicro / Gigabyte",
   "name": "Serveur 4x NVIDIA RTX PRO 6000 Blackwell Server Edition (384 Go)",
   "gpu": "4x NVIDIA RTX PRO 6000 Blackwell Server Edition (96 Go chacune)",
   "category": "enterprise",
   "devices": 4,
   "tpMax": 4,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 384,
   "bwTBs": 6.388,
   "rates": {
    "bf16": 2,
    "fp8": 4,
    "fp4": 8,
    "int8": 4
   },
   "fp4Type": "NVFP4",
   "powerW": 3200,
   "powerBasis": "calcul : 4 x 600 W + 800 W plate-forme",
   "priceHT": 99000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/"
  },
  {
   "id": "server-8x-rtx-pro-6000",
   "brand": "Supermicro / Dell",
   "name": "Serveur 8x NVIDIA RTX PRO 6000 Blackwell Server Edition (768 Go)",
   "gpu": "8x NVIDIA RTX PRO 6000 Blackwell Server Edition (96 Go chacune)",
   "category": "enterprise",
   "devices": 8,
   "tpMax": 8,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 768,
   "bwTBs": 12.776,
   "rates": {
    "bf16": 4,
    "fp8": 8,
    "fp4": 16,
    "int8": 8
   },
   "fp4Type": "NVFP4",
   "powerW": 6300,
   "powerBasis": "calcul : 8 x 600 W + 1 500 W plate-forme PCIe",
   "priceHT": 192000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/"
  },
  {
   "id": "workstation-1x-rtx-pro-5000",
   "brand": "NVIDIA",
   "name": "Station 1x NVIDIA RTX PRO 5000 Blackwell (48 Go)",
   "gpu": "1x NVIDIA RTX PRO 5000 Blackwell (GB202)",
   "category": "workstation",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 48,
   "bwTBs": 1.344,
   "rates": {
    "bf16": 0.258,
    "fp8": 0.516,
    "fp4": 1.032,
    "int8": 0.516
   },
   "fp4Type": "NVFP4",
   "powerW": 650,
   "powerBasis": "calcul : 300 W + 350 W plate-forme",
   "priceHT": 9600,
   "priceBasis": "station configurée HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/products/workstations/professional-desktop-gpus/rtx-pro-5000-blackwell/workstation-datasheet-blackwell-rtx-pro-5000-5488550-nvidia.pdf"
  },
  {
   "id": "rtx-pro-5000-mobile",
   "brand": "Dell / Lenovo / HP",
   "name": "Station mobile RTX PRO 5000 Blackwell (24 Go)",
   "gpu": "1x GPU NVIDIA RTX PRO 5000 Blackwell pour portable",
   "category": "workstation",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 24,
   "bwTBs": 0.896,
   "rates": {
    "bf16": 0.228,
    "fp8": 0.456,
    "fp4": 0.912,
    "int8": 0.456
   },
   "fp4Type": "NVFP4",
   "powerW": 230,
   "powerBasis": "calcul : 175 W TGP + 55 W plate-forme",
   "priceHT": 5208,
   "priceBasis": "milieu de fourchette TTC / 1,2",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/products/workstations/professional-laptops/compare/"
  },
  {
   "id": "server-4x-rtx-pro-5000",
   "brand": "Supermicro / ASUS",
   "name": "Serveur 4x NVIDIA RTX PRO 5000 Blackwell (192 Go)",
   "gpu": "4x NVIDIA RTX PRO 5000 Blackwell 48 Go (GB202)",
   "category": "enterprise",
   "devices": 4,
   "tpMax": 4,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 192,
   "bwTBs": 5.376,
   "rates": {
    "bf16": 1.032,
    "fp8": 2.064,
    "fp4": 4.128,
    "int8": 2.064
   },
   "fp4Type": "NVFP4",
   "powerW": 2000,
   "powerBasis": "calcul : 4 x 300 W + 800 W plate-forme",
   "priceHT": 50000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/products/workstations/professional-desktop-gpus/rtx-pro-5000-blackwell/workstation-datasheet-blackwell-rtx-pro-5000-5488550-nvidia.pdf"
  },
  {
   "id": "server-8x-rtx-pro-5000",
   "brand": "Supermicro / Gigabyte",
   "name": "Serveur 8x NVIDIA RTX PRO 5000 Blackwell (384 Go)",
   "gpu": "8x NVIDIA RTX PRO 5000 Blackwell 48 Go (GB202)",
   "category": "enterprise",
   "devices": 8,
   "tpMax": 8,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 384,
   "bwTBs": 10.752,
   "rates": {
    "bf16": 2.064,
    "fp8": 4.128,
    "fp4": 8.256,
    "int8": 4.128
   },
   "fp4Type": "NVFP4",
   "powerW": 3900,
   "powerBasis": "calcul : 8 x 300 W + 1 500 W plate-forme PCIe",
   "priceHT": 96000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/products/workstations/professional-desktop-gpus/rtx-pro-5000-blackwell/workstation-datasheet-blackwell-rtx-pro-5000-5488550-nvidia.pdf"
  },
  {
   "id": "supermicro-2x-l40s",
   "brand": "Supermicro",
   "name": "Supermicro SYS-221GE-NR (2x L40S)",
   "gpu": "2x NVIDIA L40S PCIe 48 Go",
   "category": "hgx",
   "devices": 2,
   "tpMax": 2,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 96,
   "bwTBs": 1.728,
   "rates": {
    "bf16": 0.724,
    "fp8": 1.466,
    "fp4": 0,
    "int8": 1.466
   },
   "fp4Type": null,
   "powerW": 1500,
   "powerBasis": "calcul : 2 x 350 W + 800 W plate-forme",
   "priceHT": 33000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/l40s/"
  },
  {
   "id": "smc-gh200",
   "brand": "Supermicro",
   "name": "Supermicro ARS-111GL-NHR (1x GH200 Grace Hopper)",
   "gpu": "Superpuce NVIDIA GH200 Grace Hopper",
   "category": "hgx",
   "devices": 1,
   "tpMax": 1,
   "link": "single",
   "memKind": "grace",
   "tiers": {
    "fastGB": 96,
    "fastTBs": 4,
    "slowGB": 480,
    "slowTBs": 0.384
   },
   "vramGB": 576,
   "bwTBs": 4,
   "rates": {
    "bf16": 0.99,
    "fp8": 1.979,
    "fp4": 0,
    "int8": 1.979
   },
   "fp4Type": null,
   "powerW": 1300,
   "powerBasis": "calcul : module 1 000 W + 300 W plate-forme",
   "priceHT": 48000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [
    "eol"
   ],
   "compute_source": "https://dam-cdn.nvd.orangelogic.com/AssetLink/tkno03fa645gb7f33w60rf200d330420.pdf"
  },
  {
   "id": "dgx-station-a100",
   "brand": "NVIDIA",
   "name": "NVIDIA DGX Station A100 320 Go (4x A100 80 Go)",
   "gpu": "4x NVIDIA A100 SXM4 80 Go (HBM2e)",
   "category": "legacy",
   "devices": 4,
   "tpMax": 4,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 320,
   "bwTBs": 8.156,
   "rates": {
    "bf16": 1.248,
    "fp8": 0,
    "fp4": 0,
    "int8": 2.496
   },
   "fp4Type": null,
   "powerW": 1500,
   "powerBasis": "fiche NVIDIA : 1,5 kW",
   "priceHT": 76000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a100/pdf/nvidia-a100-datasheet-nvidia-us-2188504-web.pdf"
  },
  {
   "id": "dell-xe8640-l40s",
   "brand": "Dell",
   "name": "Dell PowerEdge XE8640 (4x H100 SXM5)",
   "gpu": "4x NVIDIA H100 SXM5 80 Go (700 W), reliés par NVLink",
   "category": "hgx",
   "devices": 4,
   "tpMax": 4,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 320,
   "bwTBs": 13.4,
   "rates": {
    "bf16": 3.958,
    "fp8": 7.916,
    "fp4": 0,
    "int8": 7.916
   },
   "fp4Type": null,
   "powerW": 3600,
   "powerBasis": "calcul : 4 x 700 W + 800 W plate-forme",
   "priceHT": null,
   "priceBasis": null,
   "flags": [
    "config-reelle-h100",
    "prix-a-resourcer"
   ],
   "compute_source": "https://www.nvidia.com/en-us/data-center/l40s/"
  },
  {
   "id": "hpe-dl380a-h100",
   "brand": "HPE",
   "name": "HPE ProLiant DL380a Gen11 (4x H100 PCIe 80 Go)",
   "gpu": "4x NVIDIA H100 PCIe 80 Go (HBM2e)",
   "category": "hgx",
   "devices": 4,
   "tpMax": 4,
   "link": "pcie",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 320,
   "bwTBs": 8,
   "rates": {
    "bf16": 3.026,
    "fp8": 6.052,
    "fp4": 0,
    "int8": 6.052
   },
   "fp4Type": null,
   "powerW": 2200,
   "powerBasis": "calcul : 4 x 350 W + 800 W plate-forme",
   "priceHT": 189000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://lenovopress.lenovo.com/lp1732-thinksystem-nvidia-h100-pcie-gen5-gpu"
  },
  {
   "id": "smc-421ge-h100nvl",
   "brand": "Supermicro",
   "name": "Supermicro SYS-421GE-TNRT3 (4x H100 NVL 94 Go)",
   "gpu": "4x NVIDIA H100 NVL 94 Go (2 paires reliées par NVLink)",
   "category": "hgx",
   "devices": 4,
   "tpMax": 4,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 376,
   "bwTBs": 15.6,
   "rates": {
    "bf16": 3.341,
    "fp8": 6.682,
    "fp4": 0,
    "int8": 6.682
   },
   "fp4Type": null,
   "powerW": 2400,
   "powerBasis": "calcul : 4 x 400 W + 800 W plate-forme",
   "priceHT": 220000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [],
   "compute_source": "https://www.nvidia.com/en-us/data-center/h100/"
  },
  {
   "id": "dell-xe9640-gh200",
   "brand": "Dell",
   "name": "Dell PowerEdge XE9640 (4x H100 SXM5)",
   "gpu": "4x NVIDIA H100 SXM5 80 Go (700 W), reliés par NVLink",
   "category": "hgx",
   "devices": 4,
   "tpMax": 4,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 320,
   "bwTBs": 13.4,
   "rates": {
    "bf16": 3.958,
    "fp8": 7.916,
    "fp4": 0,
    "int8": 7.916
   },
   "fp4Type": null,
   "powerW": 3600,
   "powerBasis": "calcul : 4 x 700 W + 800 W plate-forme",
   "priceHT": null,
   "priceBasis": null,
   "flags": [
    "config-reelle-h100",
    "prix-a-resourcer"
   ],
   "compute_source": "https://dam-cdn.nvd.orangelogic.com/AssetLink/tkno03fa645gb7f33w60rf200d330420.pdf"
  },
  {
   "id": "smc-b100",
   "brand": "Supermicro",
   "name": "Supermicro SYS-821GE-TNHR-B100 (8x B100, non commercialisé)",
   "gpu": "8x NVIDIA Blackwell B100 SXM (192 Go HBM3e, spécifications préliminaires)",
   "category": "hgx",
   "devices": 8,
   "tpMax": 8,
   "link": "nvlink",
   "memKind": "discrete",
   "tiers": null,
   "vramGB": 1536,
   "bwTBs": 64,
   "rates": {
    "bf16": 14,
    "fp8": 28,
    "fp4": 56,
    "int8": 28
   },
   "fp4Type": "NVFP4",
   "powerW": 10200,
   "powerBasis": "calcul : 8 x 700 W + 4 600 W plate-forme HGX",
   "priceHT": 410000,
   "priceBasis": "milieu de fourchette HT, majoré de 30 % (hausse constatée, octobre 2026)",
   "flags": [
    "non-commercialise"
   ],
   "compute_source": "https://catalogone.com/wp-content/uploads/2024/06/NVIDIA-Blackwell-Technical-Brief.pdf"
  }
 ],
 "modeles": {
  "kimi-k3": {
   "name": "Kimi K3 (MoE de 2,8 T, 104 Md actifs)",
   "arch": {
    "L": 93,
    "h": 7168,
    "V": 163840,
    "moe": {
     "E": 896,
     "k": 16
    },
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 1.174,
     "head": 1.174,
     "attn": 36.18,
     "mlp": 0.727,
     "shared": 12.155,
     "router": 0.591,
     "latent": 4.727,
     "routed": 2722.741,
     "mtp": 0
    },
    "groups": [
     {
      "kind": "mla",
      "layers": 24,
      "nq": 96,
      "dqk": 192,
      "dv": 128,
      "latent": 512,
      "rope": 64
     },
     {
      "kind": "linear",
      "layers": 69,
      "heads": 96,
      "dk": 128,
      "dv": 128,
      "stateBytes": 4,
      "stateNote": "FP32 supposé (KDA)"
     }
    ],
    "ctx": {
     "native": 1048576,
     "max": 1048576
    },
    "ckpt": {
     "int4": {
      "gb": 1560.9,
      "repo": "moonshotai/Kimi-K3",
      "note": "MXFP4 natif, poids seuls (W4A16)",
      "ne": 2,
      "r": 0.53125
     },
     "nvfp4": {
      "gb": 1609.9,
      "repo": "nvidia/Kimi-K3-NVFP4",
      "ne": 2,
      "r": 0.5625
     },
     "q4km": {
      "gb": 1508.67,
      "repo": "unsloth/Kimi-K3-GGUF",
      "note": "UD-Q4_K_XL, aucun Q4_K_M publié"
     }
    },
    "qlora": {
     "loraM": 1072.6,
     "interAct": 55296,
     "fused": false,
     "compatible": true,
     "prereq": "Experts publiés en MXFP4 : matérialisation BF16 (~5,6 To) puis NF4 ; code distant (trust_remote_code)",
     "prereq_en": "Experts published in MXFP4: BF16 materialisation (~5.6 TB) then NF4; remote code (trust_remote_code)"
    },
    "gguf": {
     "publie": true,
     "llamacpp": "partiel",
     "source": "https://github.com/ggml-org/llama.cpp/pull/26185",
     "note": "texte seulement, vision hors branche principale",
     "verifie_le": "2026-10-08"
    }
   }
  },
  "deepseek-v4-pro": {
   "name": "DeepSeek V4 Pro 0813 (MoE de 1,6 T, 49 Md actifs)",
   "arch": {
    "L": 61,
    "h": 7168,
    "V": 129280,
    "moe": {
     "E": 384,
     "k": 6
    },
    "hc": 4,
    "mtp": true,
    "parallel": "dp",
    "p": {
     "embed": 0.927,
     "head": 0.927,
     "attn": 19.463,
     "shared": 4.03,
     "router": 0.168,
     "routed": 1547.396,
     "mtp": 26.633
    },
    "groups": [
     {
      "kind": "dsv4",
      "csa": 30,
      "hca": 31,
      "nq": 128,
      "d": 512,
      "swa": 128,
      "csaRatio": 4,
      "hcaRatio": 128,
      "topk": 1024,
      "idxHeads": 64,
      "idxDim": 128,
      "entryFp8": 584,
      "entryBf16": 1152,
      "idxBytes": 132
     }
    ],
    "kvFixed": {
     "8": 5511,
     "16": 9848
    },
    "kvPerSeq": 23000000,
    "ctx": {
     "native": 1048576,
     "max": 1048576
    },
    "ckpt": {
     "int4": {
      "gb": 892.7,
      "repo": "deepseek-ai/DeepSeek-V4-Pro-0813",
      "note": "natif : experts FP4, reste FP8",
      "ne": 1,
      "r": 0.53125
     },
     "nvfp4": {
      "gb": 941.1,
      "repo": "nvidia/DeepSeek-V4-Pro-0813-NVFP4",
      "ne": 2,
      "r": 0.5625
     },
     "q4km": {
      "gb": 950.88,
      "repo": "DevQuasar/deepseek-ai.DeepSeek-V4-Pro-0813-GGUF"
     }
    },
    "qlora": {
     "loraM": 740.6,
     "interAct": 21504,
     "fused": true,
     "compatible": true,
     "prereq": "Poids natifs FP4/FP8 : déquantification BF16 avant bitsandbytes",
     "prereq_en": "Native FP4/FP8 weights: BF16 dequantisation before bitsandbytes"
    },
    "gguf": {
     "publie": true,
     "llamacpp": "partiel",
     "source": "https://github.com/ggml-org/llama.cpp/pull/24162",
     "note": "sans MTP",
     "verifie_le": "2026-10-08"
    }
   }
  },
  "deepseek-v41-flash": {
   "name": "DeepSeek V4.1 Flash 1017 (MoE de 552 Md, 16 Md actifs en génération)",
   "arch": {
    "L": 40,
    "h": 5120,
    "V": 129280,
    "moe": {
     "E": 384,
     "k": 6
    },
    "hc": 4,
    "mtp": true,
    "parallel": "dp",
    "actDecodeB": 16,
    "actPrefillB": 8,
    "p": {
     "embed": 0.662,
     "head": 0.662,
     "attn": 5.126,
     "shared": 1.416,
     "router": 0.079,
     "routed": 543.582,
     "mtp": 14.651,
     "engram": 196.928
    },
    "engramBytes": 1.0315,
    "groups": [
     {
      "kind": "dsv4flash",
      "layers": 40,
      "nq": 64,
      "d": 512,
      "swa": 128,
      "topk": 512,
      "idxHeads": 32,
      "idxDim": 128
     }
    ],
    "kvFixed": {
     "4": 890,
     "8": 1582,
     "16": 3164
    },
    "kvFixedNote": {
     "4": "publié (README DeepSeek)",
     "8": "estimation (≥ 890 × 8/4,5)",
     "16": "estimation"
    },
    "kvPerSeq": 3000000,
    "ctx": {
     "native": 1048576,
     "max": 1048576
    },
    "ckpt": {
     "int4": {
      "gb": 510.3,
      "repo": "deepseek-ai/DeepSeek-V4.1-Flash-1017",
      "note": "natif, tables Engram comprises (203,1 Go)",
      "ne": 1,
      "r": 0.53125
     },
     "nvfp4": {
      "gb": 527.3,
      "repo": "nvidia/DeepSeek-V4.1-Flash-NVFP4",
      "note": "tables Engram FP8 comprises",
      "ne": 2,
      "r": 0.5625
     },
     "q4km": {
      "gb": 444.74,
      "repo": "vcruz305/DeepSeek-V4.1-Flash-GGUF",
      "note": "tables Engram en Q4_K incluses"
     }
    },
    "qlora": {
     "loraM": 248.6,
     "interAct": 16128,
     "fused": true,
     "compatible": false,
     "reason": "Architecture deepseek_v41 absente de transformers et de PEFT au 07/10/2026",
     "reason_en": "The deepseek_v41 architecture is missing from transformers and PEFT as of 7 October 2026"
    },
    "gguf": {
     "publie": true,
     "llamacpp": false,
     "source": "https://github.com/ggml-org/llama.cpp/pull/28696",
     "note": "architecture deepseek41 non fusionnée : fichiers utilisables seulement avec une branche dérivée",
     "verifie_le": "2026-10-08"
    }
   }
  },
  "glm-5-3": {
   "name": "GLM 5.3 (MoE de 744 Md, 40 Md actifs)",
   "arch": {
    "L": 78,
    "h": 6144,
    "V": 154880,
    "moe": {
     "E": 256,
     "k": 8
    },
    "hc": 1,
    "mtp": true,
    "parallel": "tp",
    "p": {
     "embed": 0.952,
     "head": 0.952,
     "attn": 13.068,
     "mlp": 0.679,
     "shared": 2.831,
     "router": 0.118,
     "routed": 724.776,
     "mtp": 9.953
    },
    "groups": [
     {
      "kind": "dsa",
      "layers": 78,
      "nq": 64,
      "dqk": 256,
      "dv": 256,
      "latent": 512,
      "rope": 64,
      "topk": 2048,
      "idxLayers": 21,
      "idxHeads": 32,
      "idxDim": 128,
      "idxBytes": 132,
      "denseFirst": 3
     }
    ],
    "ctx": {
     "native": 1048576,
     "max": 1048576
    },
    "ckpt": {
     "8": {
      "gb": 755.7,
      "repo": "zai-org/GLM-5.3",
      "note": "FP8 natif",
      "ne": 1,
      "r": 1
     },
     "nvfp4": {
      "gb": 464.1,
      "repo": "nvidia/GLM-5.3-NVFP4",
      "note": "experts FP4 ; attention, experts partagés et tête en BF16",
      "ne": 1.96,
      "r": 0.5625
     },
     "q4km": {
      "gb": 467.29,
      "repo": "unsloth/GLM-5.3-GGUF",
      "note": "UD-Q4_K_XL, aucun Q4_K_M publié"
     }
    },
    "qlora": {
     "loraM": 578.2,
     "interAct": 18432,
     "fused": true,
     "compatible": true,
     "prereq": "Poids FP8 natifs : déquantification BF16 avant bitsandbytes",
     "prereq_en": "Native FP8 weights: BF16 dequantisation before bitsandbytes"
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/25407",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "glm-5-3-flash": {
   "name": "GLM 5.3 Flash (MoE de 320 Md, 18 Md actifs)",
   "arch": {
    "L": 45,
    "h": 4096,
    "V": 154880,
    "moe": {
     "E": 288,
     "k": 8
    },
    "hc": 4,
    "mtp": true,
    "parallel": "tp",
    "p": {
     "embed": 0.634,
     "head": 0.634,
     "attn": 6.053,
     "mlp": 0.453,
     "shared": 1.057,
     "router": 0.05,
     "routed": 304.406,
     "mtp": 7.433
    },
    "groups": [
     {
      "kind": "dsa",
      "layers": 11,
      "nq": 64,
      "dqk": 256,
      "dv": 256,
      "latent": 512,
      "rope": 0,
      "topk": 2048,
      "idxLayers": 11,
      "idxHeads": 32,
      "idxDim": 128,
      "idxBytes": 33,
      "idxPool": 4
     },
     {
      "kind": "linear",
      "layers": 34,
      "heads": 64,
      "dk": 128,
      "dv": 128,
      "stateBytes": 4,
      "stateNote": "FP32 supposé (KDA)"
     }
    ],
    "ctx": {
     "native": 1048576,
     "max": 1048576
    },
    "ckpt": {
     "8": {
      "gb": 328.3,
      "repo": "zai-org/GLM-5.3-Flash",
      "note": "FP8 natif",
      "ne": 1,
      "r": 1
     },
     "nvfp4": {
      "gb": 204.4,
      "repo": "nvidia/GLM-5.3-Flash-NVFP4",
      "ne": 2,
      "r": 0.5625
     }
    },
    "qlora": {
     "loraM": 300.2,
     "interAct": 18432,
     "fused": true,
     "compatible": true,
     "prereq": "Poids FP8 natifs : déquantification BF16 avant bitsandbytes",
     "prereq_en": "Native FP8 weights: BF16 dequantisation before bitsandbytes"
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/27773",
     "note": "Q4_K_M publiés de 188,7 Go (batiai) à 202,0 Go (AesSedai) ; plusieurs séquences exigent --kv-unified",
     "verifie_le": "2026-10-08"
    }
   }
  },
  "mistral-large-4": {
   "name": "Mistral Large 4 « Le Chonk » (MoE de 1,05 T, 49 Md actifs)",
   "arch": {
    "unknown": true,
    "L": null,
    "h": null,
    "V": null,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "total": 1050,
     "active": 49
    },
    "groups": [],
    "ctx": {
     "native": 524288,
     "max": 1048576
    },
    "ckpt": {},
    "qlora": {
     "compatible": false,
     "reason": "Poids non publiés au 07/10/2026",
     "reason_en": "Weights not published as of 7 October 2026"
    },
    "gguf": {
     "publie": false,
     "llamacpp": null,
     "source": null,
     "note": "poids non publiés au 8 octobre 2026",
     "verifie_le": "2026-10-08"
    }
   }
  },
  "minimax-m3": {
   "name": "MiniMax M3 (MoE de 428 Md, 23 Md actifs)",
   "arch": {
    "L": 60,
    "h": 6144,
    "V": 200064,
    "moe": {
     "E": 128,
     "k": 4
    },
    "hc": 1,
    "mtp": true,
    "parallel": "tp",
    "p": {
     "embed": 1.229,
     "head": 1.229,
     "attn": 6.641,
     "mlp": 0.679,
     "shared": 3.228,
     "router": 0.045,
     "routed": 413.122
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 3,
      "nq": 64,
      "kvh": 4,
      "hd": 128
     },
     {
      "kind": "sgqa",
      "layers": 57,
      "nq": 64,
      "kvh": 4,
      "hd": 128,
      "cap": 2176,
      "idxPerTok": 128,
      "idxHeads": 4,
      "idxDim": 128,
      "note": "MSA : 16 blocs de 128 + bloc local ; clés d’index approximées à 128 octets par couche"
     }
    ],
    "ctx": {
     "native": 1048576,
     "max": 1048576
    },
    "ckpt": {
     "nvfp4": {
      "gb": 250.1,
      "repo": "nvidia/MiniMax-M3-NVFP4",
      "ne": 2,
      "r": 0.5625
     },
     "q4km": {
      "gb": 261.28,
      "repo": "bartowski/MiniMax-M3-GGUF"
     }
    },
    "qlora": {
     "loraM": 319.9,
     "interAct": 15360,
     "fused": true,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "partiel",
     "source": "https://github.com/ggml-org/llama.cpp/pull/24908",
     "note": "attention MSA en repli dense",
     "verifie_le": "2026-10-08"
    }
   }
  },
  "qwen-3-5": {
   "name": "Qwen 3.5 (MoE de 397 Md, 17 Md actifs)",
   "arch": {
    "L": 60,
    "h": 4096,
    "V": 248320,
    "moe": {
     "E": 512,
     "k": 10
    },
    "hc": 1,
    "mtp": true,
    "parallel": "tp",
    "p": {
     "embed": 1.017,
     "head": 1.017,
     "attn": 6.881,
     "shared": 0.755,
     "router": 0.126,
     "routed": 386.547,
     "mtp": 6.596
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 15,
      "nq": 32,
      "kvh": 2,
      "hd": 256
     },
     {
      "kind": "linear",
      "layers": 45,
      "heads": 64,
      "dk": 128,
      "dv": 128,
      "stateBytes": 4,
      "stateNote": "FP32 déclaré (mamba_ssm_dtype)"
     }
    ],
    "ctx": {
     "native": 262144,
     "max": 1010000
    },
    "ckpt": {
     "nvfp4": {
      "gb": 251.2,
      "repo": "nvidia/Qwen3.5-397B-A17B-NVFP4",
      "ne": 2,
      "r": 0.5625
     },
     "q4km": {
      "gb": 244.09,
      "repo": "unsloth/Qwen3.5-397B-A17B-GGUF"
     }
    },
    "qlora": {
     "loraM": 241.2,
     "interAct": 11264,
     "fused": true,
     "compatible": true,
     "warn": "QLoRA déconseillé par Unsloth sur Qwen3.5",
     "warn_en": "QLoRA discouraged by Unsloth on Qwen3.5"
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/19435",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "qwen-3-8-flash": {
   "name": "Qwen 3.8 Flash Next (MoE de 176 Md, 6 Md actifs)",
   "arch": {
    "L": 48,
    "h": 2560,
    "V": 248320,
    "moe": {
     "E": 512,
     "k": 10
    },
    "hc": 4,
    "mtp": true,
    "parallel": "tp",
    "actDecodeB": 6,
    "p": {
     "embed": 0.636,
     "head": 0.636,
     "attn": 2.6,
     "shared": 0.24,
     "router": 0.06,
     "routed": 120.8,
     "mtp": 4,
     "ngram": 51.2
    },
    "estimate": "Répartition attention / experts partagés estimée (aucun en-tête safetensors relu) ; total 180 Md dont 51,2 Md de tables n-grammes",
    "groups": [
     {
      "kind": "sgqa",
      "layers": 12,
      "nq": 24,
      "kvh": 2,
      "hd": 256,
      "cap": 2048,
      "idxPerTok": 32,
      "idxHeads": 1,
      "idxDim": 128,
      "note": "QSA : budget 2 048, indexeur 1x128 compressé x4"
     },
     {
      "kind": "linear",
      "layers": 36,
      "heads": 48,
      "dk": 128,
      "dv": 128,
      "stateBytes": 4,
      "stateNote": "FP32 déclaré (mamba_ssm_dtype)"
     }
    ],
    "ctx": {
     "native": 262144,
     "max": 1000000
    },
    "ckpt": {
     "nvfp4": {
      "gb": 132.7,
      "repo": "nvidia/Qwen3.8-Flash-Next-NVFP4",
      "note": "experts FP4, attention BF16, n-grammes FP8",
      "ne": 2,
      "r": 0.5625
     },
     "q4km": {
      "gb": 119.6,
      "repo": "bartowski/Qwen3.8-Flash-Next-GGUF",
      "note": "table n-gram quantifiée au-dessus de 4 bits"
     }
    },
    "qlora": {
     "loraM": 160,
     "interAct": 7040,
     "fused": true,
     "compatible": true,
     "estimate": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/27742",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "deepseek-v3": {
   "name": "DeepSeek-V3 / R1 (MoE de 671 Md, 37 Md actifs)",
   "arch": {
    "L": 61,
    "h": 7168,
    "V": 129280,
    "moe": {
     "E": 256,
     "k": 8
    },
    "hc": 1,
    "mtp": true,
    "parallel": "dp",
    "p": {
     "embed": 0.927,
     "head": 0.927,
     "attn": 11.413,
     "mlp": 1.189,
     "shared": 2.554,
     "router": 0.106,
     "routed": 653.909,
     "mtp": 13.464
    },
    "groups": [
     {
      "kind": "mla",
      "layers": 61,
      "nq": 128,
      "dqk": 192,
      "dv": 128,
      "latent": 512,
      "rope": 64
     }
    ],
    "ctx": {
     "native": 131072,
     "max": 163840
    },
    "ckpt": {
     "nvfp4": {
      "gb": 423.6,
      "repo": "nvidia/DeepSeek-V3-0324-NVFP4",
      "ne": 2,
      "r": 0.5625
     },
     "q4km": {
      "gb": 404.43,
      "repo": "unsloth/DeepSeek-V3-0324-GGUF"
     }
    },
    "qlora": {
     "loraM": 505.4,
     "interAct": 18432,
     "fused": true,
     "compatible": true,
     "prereq": "Poids FP8 natifs : déquantification BF16 avant bitsandbytes",
     "prereq_en": "Native FP8 weights: BF16 dequantisation before bitsandbytes"
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/11049",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "deepseek-coder-v2": {
   "name": "DeepSeek Coder V2 (MoE de 236 Md, 21 Md actifs)",
   "arch": {
    "L": 60,
    "h": 5120,
    "V": 102400,
    "moe": {
     "E": 160,
     "k": 6
    },
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 0.524,
     "head": 0.524,
     "attn": 8.954,
     "mlp": 0.189,
     "shared": 2.784,
     "router": 0.048,
     "routed": 222.718
    },
    "groups": [
     {
      "kind": "mla",
      "layers": 60,
      "nq": 128,
      "dqk": 192,
      "dv": 128,
      "latent": 512,
      "rope": 64
     }
    ],
    "ctx": {
     "native": 131072,
     "max": 163840
    },
    "ckpt": {
     "q4km": {
      "gb": 142.45,
      "repo": "bartowski/DeepSeek-Coder-V2-Instruct-GGUF"
     }
    },
    "qlora": {
     "loraM": 454.2,
     "interAct": 12288,
     "fused": true,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/7519",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "llama-405b": {
   "name": "Llama 3.1 405B (dense)",
   "arch": {
    "L": 126,
    "h": 16384,
    "V": 128256,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 2.1015,
     "head": 2.1015,
     "lin": 401.65
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 126,
      "nq": 128,
      "kvh": 8,
      "hd": 128
     }
    ],
    "ctx": {
     "native": 131072,
     "max": 131072
    },
    "ckpt": {
     "nvfp4": {
      "gb": 234.3,
      "repo": "nvidia/Llama-3.1-405B-Instruct-NVFP4",
      "lin": 0.5625
     },
     "q4km": {
      "gb": 245.72,
      "repo": "bullerwins/Meta-Llama-3.1-405B-Instruct-GGUF"
     }
    },
    "qlora": {
     "loraM": 2493.8,
     "interAct": 53248,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/8676",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "nemotron-340b": {
   "name": "Nemotron-4 340B (dense)",
   "arch": {
    "L": 96,
    "h": 18432,
    "V": 256000,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 4.7185,
     "head": 4.7185,
     "lin": 331.59
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 96,
      "nq": 96,
      "kvh": 8,
      "hd": 192
     }
    ],
    "ctx": {
     "native": 4096,
     "max": 4096
    },
    "ckpt": {
     "q4km": {
      "gb": 210.24,
      "repo": "mradermacher/Nemotron-4-340B-Instruct-hf-GGUF"
     }
    },
    "qlora": {
     "loraM": 1830.8,
     "interAct": 73728,
     "compatible": true,
     "gated": false
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/8922",
     "note": "via la conversion Hugging Face communautaire",
     "verifie_le": "2026-10-08"
    }
   }
  },
  "mistral-123b": {
   "name": "Mistral Large 2 (123 Md, dense)",
   "arch": {
    "L": 88,
    "h": 12288,
    "V": 32768,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 0.4025,
     "head": 0.4025,
     "lin": 121.805
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 88,
      "nq": 96,
      "kvh": 8,
      "hd": 128
     }
    ],
    "ctx": {
     "native": 131072,
     "max": 131072
    },
    "ckpt": {
     "q4km": {
      "gb": 73.3,
      "repo": "bartowski/Mistral-Large-Instruct-2407-GGUF",
      "note": "calcul : 4,78 bits mesurés sur le fichier"
     }
    },
    "qlora": {
     "loraM": 1118.8,
     "interAct": 28672,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://huggingface.co/bartowski/Mistral-Large-Instruct-2407-GGUF",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "llama-70b": {
   "name": "Llama 3.3 / 3.1 70B (dense)",
   "arch": {
    "L": 80,
    "h": 8192,
    "V": 128256,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 1.0505,
     "head": 1.0505,
     "lin": 68.453
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 80,
      "nq": 64,
      "kvh": 8,
      "hd": 128
     }
    ],
    "ctx": {
     "native": 131072,
     "max": 131072
    },
    "ckpt": {
     "nvfp4": {
      "gb": 42.7,
      "repo": "nvidia/Llama-3.3-70B-Instruct-NVFP4",
      "lin": 0.5625
     },
     "q4km": {
      "gb": 42.52,
      "repo": "bartowski/Llama-3.3-70B-Instruct-GGUF"
     }
    },
    "qlora": {
     "loraM": 828.4,
     "interAct": 28672,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/8676",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "qwen-72b": {
   "name": "Qwen 2.5 72B (dense)",
   "arch": {
    "L": 80,
    "h": 8192,
    "V": 152064,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 1.2455,
     "head": 1.2455,
     "lin": 70.215
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 80,
      "nq": 64,
      "kvh": 8,
      "hd": 128
     }
    ],
    "ctx": {
     "native": 32768,
     "max": 131072
    },
    "ckpt": {
     "q4km": {
      "gb": 47.42,
      "repo": "bartowski/Qwen2.5-72B-Instruct-GGUF"
     }
    },
    "qlora": {
     "loraM": 842.1,
     "interAct": 29568,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/5037",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "qwen-32b-coder": {
   "name": "Qwen 2.5 Coder 32B (dense)",
   "arch": {
    "L": 64,
    "h": 5120,
    "V": 152064,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 0.7785,
     "head": 0.7785,
     "lin": 31.207
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 64,
      "nq": 40,
      "kvh": 8,
      "hd": 128
     }
    ],
    "ctx": {
     "native": 32768,
     "max": 131072
    },
    "ckpt": {
     "q4km": {
      "gb": 19.85,
      "repo": "bartowski/Qwen2.5-Coder-32B-Instruct-GGUF"
     }
    },
    "qlora": {
     "loraM": 536.9,
     "interAct": 27648,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/5037",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "qwen38-27b": {
   "name": "Qwen 3.8 27B (dense)",
   "arch": {
    "L": 64,
    "h": 5120,
    "V": 248320,
    "moe": null,
    "hc": 1,
    "mtp": true,
    "parallel": "tp",
    "p": {
     "embed": 1.271,
     "head": 1.271,
     "lin": 24.351,
     "mtp": 0.425
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 16,
      "nq": 24,
      "kvh": 4,
      "hd": 256
     },
     {
      "kind": "linear",
      "layers": 48,
      "heads": 48,
      "dk": 128,
      "dv": 128,
      "stateBytes": 4,
      "stateNote": "FP32 déclaré (mamba_ssm_dtype)"
     }
    ],
    "ctx": {
     "native": 262144,
     "max": 1000000
    },
    "ckpt": {
     "8": {
      "gb": 30.87,
      "repo": "Qwen/Qwen3.8-27B-FP8",
      "note": "tour de vision comprise"
     },
     "16": {
      "gb": 55.56,
      "repo": "Qwen/Qwen3.8-27B",
      "note": "BF16, tour de vision comprise"
     },
     "nvfp4": {
      "gb": 21.9,
      "repo": "nvidia/Qwen3.8-27B-NVFP4",
      "note": "NVFP4 et FP8 mêlés",
      "lin": 0.656
     },
     "q4km": {
      "gb": 16.81,
      "repo": "lmstudio-community/Qwen3.8-27B-GGUF"
     }
    },
    "qlora": {
     "loraM": 466.9,
     "interAct": 17408,
     "compatible": true,
     "warn": "QLoRA déconseillé par Unsloth sur la famille Qwen3.5 (applicabilité au 27B non vérifiée)",
     "warn_en": "QLoRA discouraged by Unsloth on the Qwen3.5 family (applicability to the 27B not verified)"
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/19435",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "devstral-24b": {
   "name": "Mistral Small / Devstral 24B (dense)",
   "arch": {
    "L": 40,
    "h": 5120,
    "V": 131072,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 0.671,
     "head": 0.671,
     "lin": 22.23
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 40,
      "nq": 32,
      "kvh": 8,
      "hd": 128
     }
    ],
    "ctx": {
     "native": 131072,
     "max": 131072
    },
    "ckpt": {
     "q4km": {
      "gb": 14.33,
      "repo": "bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF"
     }
    },
    "qlora": {
     "loraM": 369.6,
     "interAct": 32768,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/17644",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  },
  "llama-8b": {
   "name": "Llama 3.1 8B (dense)",
   "arch": {
    "L": 32,
    "h": 4096,
    "V": 128256,
    "moe": null,
    "hc": 1,
    "mtp": false,
    "parallel": "tp",
    "p": {
     "embed": 0.525,
     "head": 0.525,
     "lin": 6.98
    },
    "groups": [
     {
      "kind": "gqa",
      "layers": 32,
      "nq": 32,
      "kvh": 8,
      "hd": 128
     }
    ],
    "ctx": {
     "native": 131072,
     "max": 131072
    },
    "ckpt": {
     "nvfp4": {
      "gb": 6,
      "repo": "nvidia/Llama-3.1-8B-Instruct-NVFP4",
      "lin": 0.5625
     },
     "q4km": {
      "gb": 4.92,
      "repo": "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF"
     }
    },
    "qlora": {
     "loraM": 167.8,
     "interAct": 14336,
     "compatible": true
    },
    "gguf": {
     "publie": true,
     "llamacpp": "oui",
     "source": "https://github.com/ggml-org/llama.cpp/pull/8676",
     "note": null,
     "verifie_le": "2026-10-08"
    }
   }
  }
 }
}