{
  "evidence_date": "2026-10-29",
  "author": "Elizabeth Ann West",
  "old_hardware": {
    "gpu": "RTX 5070",
    "vram_gb": 12,
    "ram_gb": 32,
    "cpu": "AMD Ryzen 7 8700F"
  },
  "new_hardware": {
    "gpu": "RTX PRO 5000 72GB Blackwell",
    "vram_gb": 72,
    "ram_gb": 128,
    "configured_ram_mt_s": 3600,
    "cpu": "AMD Ryzen 7 8700F",
    "driver": "597.30"
  },
  "scope": {
    "new_cells": 96,
    "high_resolution_cells": 48,
    "matched_pairs": 18,
    "scenes": [
      "boarding",
      "departure",
      "pass"
    ],
    "profiles": [
      "h3-turbo4",
      "h3-base20",
      "ltx-native8",
      "ltx-two-stage8plus3"
    ],
    "clip_seconds": [
      5,
      10
    ],
    "delivery_fps": 24,
    "delivery_sizes": [
      [
        854,
        480
      ],
      [
        1280,
        720
      ],
      [
        1920,
        1080
      ],
      [
        2048,
        1152
      ]
    ],
    "generation_1080_size": [
      1920,
      1088
    ],
    "aspect": "landscape; not a native vertical production benchmark"
  },
  "inputs": "one opening image and written motion prompt; no old finished video",
  "matching": "All 18 pairs share reference image bytes, prompt text, seed, frame dimensions/counts, FPS, step counts and normalized workflow graphs. Normalization removes uploaded/output filenames and resolves the H3 five-second frame-count helper to the same recorded 124 frames. Cache, RAM, GPU, storage/runtime environment were not isolated.",
  "timing": {
    "server": "model load if required, inference, decode, initial save",
    "file_ready": "upload, server execution, polling, copy, crop/trim and encoding",
    "excluded": "setup, downloads, technical review, creative review, editing, Dropbox"
  },
  "resources": "machine-wide approximately five-second samples; not instantaneous peaks or isolated model allocations",
  "statistics": "median of three scenes, one measured run per scene/configuration; no randomized cache states or repeated-run confidence estimates",
  "limits": [
    "GPU, RAM, storage and cache effects not isolated",
    "technical success is not creative approval",
    "quantized local recipes are not verified as identical to hosted endpoints",
    "768p H3 Turbo adapter used experimentally at higher resolutions"
  ],
  "software": {
    "torch": "2.12.1+cu130",
    "torchvision": "0.27.1+cu130",
    "torchaudio": "2.11.0+cu130",
    "comfy-aimdo": "0.5.5",
    "comfy-kitchen": "0.2.37"
  }
}