{
  "kind": "illustrative arithmetic, not a benchmark or measured hardware memory limit",
  "units": {
    "GB": 1000000000,
    "GiB": 1073741824
  },
  "model_assumptions": {
    "parameters": 70000000000,
    "weight_bits": 4,
    "metadata_fraction_of_raw_weights": 0.1,
    "layers": 80,
    "kv_heads": 8,
    "head_dimension": 128,
    "kv_element_bytes": 2,
    "kv_bytes_per_token_per_sequence": 327680,
    "runtime_and_os_reserve_GiB": 16
  },
  "memory_cases": [
    {
      "tokens_per_sequence": 8192,
      "concurrent_sequences": 1,
      "bytes": {
        "weights": 35000000000,
        "quantization_metadata": 3500000000,
        "kv_cache": 2684354560,
        "runtime_and_os_reserve": 17179869184
      },
      "GiB": {
        "weights": 32.59629011154175,
        "quantization_metadata": 3.259629011154175,
        "kv_cache": 2.5,
        "runtime_and_os_reserve": 16.0
      },
      "total_bytes": 58364223744,
      "total_GiB": 54.35591912269592
    },
    {
      "tokens_per_sequence": 32768,
      "concurrent_sequences": 1,
      "bytes": {
        "weights": 35000000000,
        "quantization_metadata": 3500000000,
        "kv_cache": 10737418240,
        "runtime_and_os_reserve": 17179869184
      },
      "GiB": {
        "weights": 32.59629011154175,
        "quantization_metadata": 3.259629011154175,
        "kv_cache": 10.0,
        "runtime_and_os_reserve": 16.0
      },
      "total_bytes": 66417287424,
      "total_GiB": 61.85591912269592
    },
    {
      "tokens_per_sequence": 8192,
      "concurrent_sequences": 4,
      "bytes": {
        "weights": 35000000000,
        "quantization_metadata": 3500000000,
        "kv_cache": 10737418240,
        "runtime_and_os_reserve": 17179869184
      },
      "GiB": {
        "weights": 32.59629011154175,
        "quantization_metadata": 3.259629011154175,
        "kv_cache": 10.0,
        "runtime_and_os_reserve": 16.0
      },
      "total_bytes": 66417287424,
      "total_GiB": 61.85591912269592
    },
    {
      "tokens_per_sequence": 32768,
      "concurrent_sequences": 4,
      "bytes": {
        "weights": 35000000000,
        "quantization_metadata": 3500000000,
        "kv_cache": 42949672960,
        "runtime_and_os_reserve": 17179869184
      },
      "GiB": {
        "weights": 32.59629011154175,
        "quantization_metadata": 3.259629011154175,
        "kv_cache": 40.0,
        "runtime_and_os_reserve": 16.0
      },
      "total_bytes": 98629542144,
      "total_GiB": 91.85591912269592
    }
  ],
  "decode_bandwidth_rooflines": [
    {
      "memory_bandwidth_GB_s": 273,
      "weights_read_GB_per_token": 35,
      "ideal_tokens_s": 7.8
    },
    {
      "memory_bandwidth_GB_s": 1200,
      "weights_read_GB_per_token": 35,
      "ideal_tokens_s": 34.285714285714285
    }
  ],
  "decode_assumptions": [
    "Dense model, one sequence, one token per decoding step",
    "Every step streams 35 GB of raw weights at the stated peak bandwidth",
    "Ignores metadata, KV traffic, compute, synchronization and operating-system traffic",
    "Not a measured or expected token rate; not applicable to MoE, batching or speculative decoding"
  ]
}