{
  "$schema": "./gpu-config.schema.json",
  "id": "h200",
  "name": "NVIDIA H200 (SXM 141 GiB)",
  "match_names": ["H200", "NVIDIA H200"],
  "vram_gb": 141,
  "compute_capability": "9.0",
  "arch": "hopper",
  "memory_bandwidth_gbs": 4800,
  "fp8": true,
  "fp4": false,
  "flash_attn_3": true,
  "llama_server_flags": {
    "n_gpu_layers": 999,
    "ctx_size": 262144,
    "batch_size": 4096,
    "ubatch_size": 2048,
    "n_parallel": 16,
    "cache_type_k": "qjl1_256",
    "cache_type_v": "q4_polar",
    "flash_attn": true,
    "split_mode": "none",
    "main_gpu": 0,
    "mlock": true,
    "no_mmap": false,
    "no_kv_offload": false,
    "ctx_checkpoints": 16,
    "ctx_checkpoint_interval": 8192,
    "draft_max": 32,
    "draft_min": 8,
    "draft_p_min": 0.5
  },
  "bundle_recommendations": {
    "voice": {
      "ctx_size": 16384,
      "max_parallel": 32,
      "batch_size": 2048,
      "ubatch_size": 512
    },
    "eliza-1-2b": {
      "ctx_size": 1048576,
      "max_parallel": 64
    },
    "eliza-1-4b": {
      "ctx_size": 1048576,
      "max_parallel": 32
    },
    "eliza-1-9b": {
      "ctx_size": 1048576,
      "max_parallel": 32
    },
    "eliza-1-27b": {
      "ctx_size": 131072,
      "max_parallel": 16
    }
  },
  "expected_metrics": {
    "ttfa_p50_ms": 110,
    "ttfa_p95_ms": 180,
    "rtf": 0.2,
    "tokens_per_second_decode": 320,
    "_provenance": "extrapolated"
  },
  "known_limits": [
    "FP4 not supported on sm_90 (Blackwell-only)",
    "PCIe spill path defeats the bandwidth advantage; keep KV in HBM"
  ],
  "_provenance": "extrapolated; numbers derived from VRAM + mem-bw math in SPECS.md"
}