{
  "checkedAt": "2026-10-27",
  "variants": {
    "founders": {
      "name": "DGX Spark Founders Edition",
      "ssd": 4000,
      "price": 7500,
      "shop": "Joybuy",
      "url": "https://www.joybuy.de/dp/nvidia-dgx-spark-founders-edition-nvidia/100011565021660"
    },
    "pgx": {
      "name": "Lenovo ThinkStation PGX",
      "ssd": 1000,
      "price": 5869,
      "shop": "easynotebooks",
      "url": "https://www.easynotebooks.de/PC-s/Workstations/Lenovo/1125960/Lenovo-ThinkStation-PGX-GB10-Grace-Blackwell-128GB-RAM-1TB-SSD-NVIDIA-DGX-OS"
    }
  },
  "priceProvenance": "Vom Nutzer vorgegebene Händlerpreise. Joybuy 403, easynotebooks nicht abrufbar; keine unabhängige Preisbestätigung.",
  "hardwareSources": [
    "https://docs.nvidia.com/dgx/dgx-spark/hardware.html",
    "https://psref.lenovo.com/Product/ThinkStation/ThinkStation_PGX?MT=30KL"
  ],
  "benchmarkSources": {
    "sparkggml": {
      "name": "llama.cpp · DGX-Spark-Testreihe",
      "url": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
      "date": "Systemprotokoll 05.02.2026 · Abruf 27.10.2026",
      "setup": "Ein DGX Spark · GB10 · 128 GB · Linux aarch64 · CUDA 13.0 · Treiber 580.95.05 · llama.cpp 11fb327bf (7941).",
      "protocol": "llama-bench: pp2048 / tg32, Kontexttiefe d0–d32768, Flash Attention an, 99 GPU-Layer, ubatch 2048, mmap aus, direct I/O an. Einzelstrom; keine aggregierten Batch-Werte. MXFP4-MoE, nicht mit dichten 120B gleichsetzen."
    },
    "sparkqwen": {
      "name": "DandinPower · eigene DGX-Messungen",
      "url": "https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md",
      "date": "Repository-Stand 05.12.2025 · Abruf 27.10.2026",
      "setup": "DGX Spark GB10 · 128 GB · llama.cpp / CUDA · Q4_K_M. llama.cpp-Build im Bericht nicht angegeben.",
      "protocol": "Bericht: Generation bei 512 Tokens Kontexttiefe und Batch 1; PP mit 512 Tokens Prompt. Testskript: 20 Threads, 99 GPU-Layer, FA an, fünf Wiederholungen. Bericht ordnet TG nicht eindeutig einer Ausgabelänge (64/256/512) zu. Qwen3-30B-A3B-Instruct-2507 ist eine eigene Modellversion."
    },
    "sparkollama": {
      "name": "Gopi Trinadh Maddikunta · Ollama-Rohlogs",
      "url": "https://huggingface.co/datasets/G3nadh/dgx-spark-benchmarks",
      "date": "Coding-Lauf 30.03.2026 · Abruf 27.10.2026",
      "setup": "DGX Spark · GB10 · 128 GB · 4 TB · Ollama 0.18.3 · CUDA 13.0 · Treiber 580.142.",
      "protocol": "Coding-Lauf: native Ollama-Modelltags, Quantisierung nicht dokumentiert. Prompt- und generierte Tokenzahl pro Zeile aus Rohlog. Kontextlimit/Cache-Zustand nicht vollständig dokumentiert. Keine Gleichsetzung mit Q4_K_M oder den Modellen ohne Coder-Zusatz. Keine direkte Messung der Lenovo PGX."
    }
  },
  "benchmarkRows": [
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 20B (MoE)",
      "quant": "MXFP4",
      "context": 0,
      "contextLabel": "d0 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 83.43,
      "pp": 4505.82,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 20B (MoE)",
      "quant": "MXFP4",
      "context": 4096,
      "contextLabel": "d4.096 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 79.22,
      "pp": 4158.34,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 20B (MoE)",
      "quant": "MXFP4",
      "context": 8192,
      "contextLabel": "d8.192 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 75.22,
      "pp": 3993.81,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 20B (MoE)",
      "quant": "MXFP4",
      "context": 16384,
      "contextLabel": "d16.384 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 70.36,
      "pp": 3449.98,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 20B (MoE)",
      "quant": "MXFP4",
      "context": 32768,
      "contextLabel": "d32.768 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 61.65,
      "pp": 2689.42,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 120B (MoE)",
      "quant": "MXFP4",
      "context": 0,
      "contextLabel": "d0 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 58.72,
      "pp": 2443.91,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 120B (MoE)",
      "quant": "MXFP4",
      "context": 4096,
      "contextLabel": "d4.096 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 55.67,
      "pp": 2309.84,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 120B (MoE)",
      "quant": "MXFP4",
      "context": 8192,
      "contextLabel": "d8.192 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 52.87,
      "pp": 2216.68,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 120B (MoE)",
      "quant": "MXFP4",
      "context": 16384,
      "contextLabel": "d16.384 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 49.45,
      "pp": 1956.31,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Linux aarch64 · CUDA 13.0",
      "model": "gpt-oss 120B (MoE)",
      "quant": "MXFP4",
      "context": 32768,
      "contextLabel": "d32.768 · pp2048 / tg32",
      "testNote": "Einzelstrom · vorgefüllte Kontexttiefe",
      "tg": 42.76,
      "pp": 1567.08,
      "source": "sparkggml",
      "backend": "llama.cpp · CUDA + FA",
      "build": "11fb327bf (7941)"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "20 Threads · GPU-Offload 99",
      "model": "Qwen3 8B",
      "quant": "Q4_K_M",
      "context": 512,
      "contextLabel": "TG @ d512 · PP512",
      "testNote": "Batch 1 · TG-Ausgabelänge nicht eindeutig",
      "tg": 43.7,
      "pp": 3167,
      "source": "sparkqwen",
      "backend": "llama.cpp · CUDA + FA",
      "build": "nicht angegeben"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "20 Threads · GPU-Offload 99",
      "model": "Qwen3 32B",
      "quant": "Q4_K_M",
      "context": 512,
      "contextLabel": "TG @ d512 · PP512",
      "testNote": "Batch 1 · TG-Ausgabelänge nicht eindeutig",
      "tg": 10.7,
      "pp": 762,
      "source": "sparkqwen",
      "backend": "llama.cpp · CUDA + FA",
      "build": "nicht angegeben"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "20 Threads · GPU-Offload 99",
      "model": "Qwen3-30B-A3B-Instruct-2507 (MoE)",
      "quant": "Q4_K_M",
      "context": 512,
      "contextLabel": "TG @ d512 · PP512",
      "testNote": "Batch 1 · TG-Ausgabelänge nicht eindeutig",
      "tg": 89.3,
      "pp": 2541,
      "source": "sparkqwen",
      "backend": "llama.cpp · CUDA + FA",
      "build": "nicht angegeben"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Coding-Prompt · Ollama-Rohlog",
      "model": "Llama 3.1 8B",
      "quant": "Ollama-Tag · Quantisierung nicht angegeben",
      "context": null,
      "contextLabel": "35 Prompt / 369 Output-Tokens",
      "testNote": "Kontextlimit unbekannt · kein standardisierter llama-bench",
      "tg": 42.51,
      "pp": 966.02,
      "source": "sparkollama",
      "backend": "Ollama · CUDA",
      "build": "Ollama 0.18.3",
      "reportedAt": "2026-03-30"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Coding-Prompt · Ollama-Rohlog",
      "model": "Qwen2.5-Coder 32B",
      "quant": "Ollama-Tag · Quantisierung nicht angegeben",
      "context": null,
      "contextLabel": "54 Prompt / 552 Output-Tokens",
      "testNote": "Kontextlimit unbekannt · kein standardisierter llama-bench",
      "tg": 10.33,
      "pp": 374.36,
      "source": "sparkollama",
      "backend": "Ollama · CUDA",
      "build": "Ollama 0.18.3",
      "reportedAt": "2026-03-30"
    },
    {
      "family": "spark",
      "hardware": "DGX Spark · GB10 · 128 GB",
      "host": "Coding-Prompt · Ollama-Rohlog",
      "model": "Llama 3.3 70B",
      "quant": "Ollama-Tag · Quantisierung nicht angegeben",
      "context": null,
      "contextLabel": "35 Prompt / 507 Output-Tokens",
      "testNote": "Kontextlimit unbekannt · kein standardisierter llama-bench",
      "tg": 4.67,
      "pp": 118.05,
      "source": "sparkollama",
      "backend": "Ollama · CUDA",
      "build": "Ollama 0.18.3",
      "reportedAt": "2026-03-30"
    }
  ]
}