{
  "$schema": "./gpu-config.schema.json",
  "id": "rtx-4090",
  "name": "NVIDIA GeForce RTX 4090",
  "match_names": ["RTX 4090", "GeForce RTX 4090"],
  "vram_gb": 24,
  "compute_capability": "8.9",
  "arch": "ada-lovelace",
  "memory_bandwidth_gbs": 1008,
  "fp8": true,
  "fp4": false,
  "flash_attn_3": false,
  "llama_server_flags": {
    "n_gpu_layers": 999,
    "ctx_size": 32768,
    "batch_size": 2048,
    "ubatch_size": 512,
    "n_parallel": 8,
    "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": 8,
    "ctx_checkpoint_interval": 8192,
    "draft_max": 24,
    "draft_min": 4,
    "draft_p_min": 0.5
  },
  "bundle_recommendations": {
    "voice": {
      "ctx_size": 8192,
      "max_parallel": 4,
      "batch_size": 1024,
      "ubatch_size": 256
    },
    "eliza-1-2b": {
      "ctx_size": 65536,
      "max_parallel": 16
    },
    "eliza-1-4b": {
      "ctx_size": 65536,
      "max_parallel": 8
    },
    "eliza-1-9b": {
      "ctx_size": 65536,
      "max_parallel": 8
    },
    "eliza-1-27b": {
      "ctx_size": 32768,
      "max_parallel": 2
    }
  },
  "expected_metrics": {
    "ttfa_p50_ms": 220,
    "ttfa_p95_ms": 320,
    "rtf": 0.4,
    "tokens_per_second_decode": 140,
    "_provenance": "extrapolated"
  },
  "known_limits": [
    "flash-attn-3 unsupported (Hopper-only); uses flash-attn-2",
    "FP4 not supported on sm_89",
    "27B + 32k context is single-slot only"
  ],
  "_provenance": "extrapolated; numbers derived from VRAM + mem-bw math in SPECS.md"
}