{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "a-rig-your-friend-already-owns",
 "title": "A Rig Your Friend Already Owns",
 "dek": "We re-rendered our game's actual art — same prompts, same seeds, same checkpoint files — on a several-year-old consumer RTX 3090 and a gaming laptop, paired render-for-render against our 96 GB card's own archived output. Two of the three runs happened in the same physical computer. This page is the stopwatch, told honestly — including the one number we don't yet trust.",
 "published": "2026-08-26",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/a-rig-your-friend-already-owns/",
  "md_url": "https://research.strata2signal.com/a-rig-your-friend-already-owns/index.md",
  "md_sha": "36a9969e7a6147c89a494287df58545c5f4a05f73925bbe36baccf06fcf20f24",
  "html_sha": "a726fb641634bddcb78027913e637ec1b1bb4bfd0062de5c4f465f1496e82d53",
  "anchors_sha": "34c03bd155199c8590e69d63b1cd72ed2d8f691104cd5e9fff8090858f5b2243"
 },
 "short": {
  "paragraph": "We re-rendered our game's actual art - same prompts, same seeds, same checkpoint files - on a several-year-old consumer RTX 3090 and a gaming laptop, paired render-for-render against our 96 GB card's own archived output. Two of the three runs happened in the same physical computer. This page is the stopwatch, told honestly - including the one number we don't yet trust.",
  "from_dek": true,
  "counts": {
   "words": 2311,
   "minutes": 11,
   "tables": 3,
   "kit": true
  },
  "bullets": [
   {
    "text": "a 3090's klein warm-up paid 9.25 s",
    "figure": "9.25 s",
    "cite": "what-we-ran-and-the-rule-behind-every-number",
    "quote": "Timing is the engine's own execution window - denoising, image decode, and file save; not model load (that's what the warm-ups absorb: the 3090's klein warm-up paid 9.25 s against a 1.65 s timed median) and not network transport, which a player pays on top.",
    "span": [
     7278,
     7285
    ]
   },
   {
    "text": "the big card is about 3.4×",
    "figure": "3.4×",
    "cite": "the-sketch-ladder-the-number-the-game-lives-on",
    "quote": "The big card is about 3.4× the 3090 and 2.7-2.9× the laptop down this ladder - and in our own playtests, below about a second a render simply reads as \"the picture appeared.\" We state that as our reading, not a law of nature, and the engine clock is the render only: a player also pays their app's own overhead on top.",
    "span": [
     10288,
     10293
    ]
   },
   {
    "text": "the 96 GB card's hidream times are nearly flat at 2.50 s",
    "figure": "2.50 s",
    "cite": "three-findings-we-didnt-order",
    "quote": "But split it by resolution and the story cracks: the 96 GB card's own hidream times are nearly flat - 2.49 s at 512, 2.50 s at 768, 2.60 s at 1024 - for a 4× increase in pixels.",
    "span": [
     10952,
     10959
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "four-words-this-page-leans-on",
   "heading": "Four words this page leans on",
   "level": 2,
   "span": [
    1086,
    1597
   ],
   "chunks": [
    [
     1086,
     1597
    ]
   ],
   "chars": 511,
   "digest": "the article defines diffusion model, steps, distilled / turbo, and VRAM to establish a common vocabulary for the technical discussion.",
   "digest_skipped": null
  },
  {
   "id": "the-moment-this-page-is-about",
   "heading": "The moment this page is about",
   "level": 2,
   "span": [
    1597,
    2733
   ],
   "chunks": [
    [
     1597,
     2733
    ]
   ],
   "chars": 1136,
   "digest": "RealKeep is a living-world RPG that runs local art and logic on home computers. the authors investigate how well the game performs on hardware a normal person owns, such as an RTX 3090 or a laptop, rather than flagship cards.",
   "digest_skipped": null
  },
  {
   "id": "what-we-ran-and-the-rule-behind-every-number",
   "heading": "What we ran, and the rule behind every number",
   "level": 2,
   "span": [
    2733,
    8033
   ],
   "chunks": [
    [
     2733,
     8033
    ]
   ],
   "chars": 5300,
   "digest": "the authors tested three rigs, including a 96 GB card and an RTX 5090 Laptop GPU, using the same engine and checkpoint files. they measured 171 renders per rig across various models and resolutions, following a reporting rule to use per-model medians.",
   "digest_skipped": null
  },
  {
   "id": "the-table-that-answers-the-question",
   "heading": "The table that answers the question",
   "level": 2,
   "span": [
    8033,
    9340
   ],
   "chunks": [
    [
     8033,
     9340
    ]
   ],
   "chars": 1307,
   "digest": "a table provides median seconds per render for six models across three different hardware setups. it includes multipliers comparing the 96 GB card and an RTX 5090 Laptop GPU to an RTX 3090 24G.",
   "digest_skipped": null
  },
  {
   "id": "the-sketch-ladder-the-number-the-game-lives-on",
   "heading": "The sketch ladder — the number the game lives on",
   "level": 2,
   "span": [
    9340,
    10587
   ],
   "chunks": [
    [
     9340,
     10587
    ]
   ],
   "chars": 1247,
   "digest": "the FLUX.2 klein-4B model is used for in-game sketches. the table shows median render times for three resolutions, noting that a 512×512 sketch takes 0.26 s on a 96G card.",
   "digest_skipped": null
  },
  {
   "id": "three-findings-we-didnt-order",
   "heading": "Three findings we didn't order",
   "level": 2,
   "span": [
    10587,
    13291
   ],
   "chunks": [
    [
     10587,
     13291
    ]
   ],
   "chars": 2704,
   "digest": "the authors report three findings: an anomalous baseline for hidream-o1-dev, the fact that speed class survives the hardware drop across the tested cards, and that 24 GB of VRAM provides the option of native precision.",
   "digest_skipped": null
  },
  {
   "id": "what-this-means-for-the-game",
   "heading": "What this means for the game",
   "level": 2,
   "span": [
    13291,
    14066
   ],
   "chunks": [
    [
     13291,
     14066
    ]
   ],
   "chars": 775,
   "digest": "the production art stack runs at felt-instant to comfortable speeds on a single several-year-old consumer card or a gaming laptop. the authors note that the game runs on a rig an enthusiast friend already owns.",
   "digest_skipped": null
  },
  {
   "id": "the-kit",
   "heading": "The kit",
   "level": 2,
   "span": [
    14066,
    15065
   ],
   "chunks": [
    [
     14066,
     15065
    ]
   ],
   "chars": 999,
   "digest": "the kit contains the manifest, images, and run logs for the 342 renders. it includes every timing field, per-model licence citations, and the engine's own render-seconds.",
   "digest_skipped": null
  }
 ],
 "chips": [
  {
   "id": "c-ce6c70a1",
   "q": "what terminology defines the model's speed and memory requirements?",
   "a": "The article defines diffusion models as AI that refines noise into images, steps as the number of refinement passes, distilled or turbo models as those requiring only 4–8 steps, and VRAM as the graphics card memory required to run a model at full speed.",
   "cites": [
    "four-words-this-page-leans-on"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-d826bd00",
   "q": "how does the game handle local image generation?",
   "a": "RealKeep is a living-world RPG where the mind, art, and dice run locally on home computers without cloud AI services or API meters. The speed of the experience depends on the strength of the local machines used to carry the game.",
   "cites": [
    "the-moment-this-page-is-about"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-6f73ad2e",
   "q": "how were the hardware comparisons standardized?",
   "a": "The comparison used a consumer-grade tower that remained unchanged except for the graphics card, alongside an RTX 5090 Laptop GPU. All rigs ran the same checkpoint files through the same engine, ensuring the software environment remained constant across tests.",
   "cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-63c56900",
   "q": "what are the median render times for various models?",
   "a": "Median render times vary by model, such as 1.65 s for FLUX.2 klein-4B, 4.44 s for z-image-turbo, and 28.95 s for kandinsky5-lite when using an RTX 3090. The table provides these medians to show how different models perform across different hardware.",
   "cites": [
    "the-table-that-answers-the-question"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-6d24a427",
   "q": "how fast does the primary sketch artist perform at different sizes?",
   "a": "For the FLUX.2 klein-4B model, the median time for a 512×512 sketch is 0.88 s on an RTX 3090, 1.65 s at 768×768, and 2.86 s at 1024×1024. On a 96G card, these times are 0.26 s, 0.48 s, and 0.82 s respectively.",
   "cites": [
    "the-sketch-ladder-the-number-the-game-lives-on"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-080d4f08",
   "q": "what anomaly was observed in the hidream-o1-dev results?",
   "a": "The hidream-o1-dev model showed an anomalous baseline where the 96 GB card's times were nearly flat across different resolutions. This caused the laptop to appear to beat the 96 GB card at 512 resolution, a result the authors distrust due to the inconsistent per-resolution data.",
   "cites": [
    "three-findings-we-didnt-order"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-7cae75c7",
   "q": "what is the practical implication for the game's performance?",
   "a": "The production art stack runs at felt-instant to comfortable speeds on a single several-year-old consumer card or a gaming laptop. This means the game can run on hardware an enthusiast already owns rather than requiring a dedicated server.",
   "cites": [
    "what-this-means-for-the-game"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § The table that answers the question · § At a glance"
  },
  {
   "slug": "fifty-seven-milliseconds",
   "why": "shares ground with § The sketch ladder — the number the game lives on · § Finding four: the argument with our own settings — and what it changed"
  },
  {
   "slug": "two-used-3080s-priced",
   "why": "shares ground with § The table that answers the question · § Renders: one 3080 against one 3090, seconds per image"
  }
 ],
 "thanks": "",
 "kit": {
  "url": "https://research.strata2signal.com/a-rig-your-friend-already-owns/data/",
  "licence": "CC BY 4.0",
  "files": [
   "rematch-3090-0826-manifest.json",
   "rematch-rigB-0826-manifest.json",
   "rematch-3090-0826.log",
   "rematch-rigB-0826.log",
   "graph-flux2-klein-4b.json",
   "graph-ctrl-z-image-turbo.json",
   "graph-krea2-turbo.json",
   "graph-hidream-o1-dev.json",
   "graph-kandinsky5-lite.json",
   "graph-z-image-base.json",
   "provenance.json",
   "README.md"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 3,
  "digests_written": 8,
  "digests_kept": 8,
  "chips_written": 8,
  "chips_kept": 7,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 0,
   "length": 1,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0
  },
  "new_noun_tokens_checked": [
   "0.26",
   "0.48",
   "0.82",
   "0.88",
   "1.65",
   "1024×1024",
   "171",
   "2.50",
   "2.86",
   "24",
   "24G",
   "28.95",
   "3.4×",
   "30",
   "3090",
   "3090's",
   "342",
   "4-8",
   "4.44",
   "5090",
   "512",
   "512×512",
   "768×768",
   "9.25",
   "96",
   "96G",
   "AI",
   "API",
   "B-class",
   "FLUX.2",
   "FLUX.2-dev",
   "GB",
   "GPU",
   "Laptop",
   "RPG",
   "RTX",
   "VRAM",
   "fp8",
   "hidream-o1-dev",
   "kandinsky5-lite",
   "klein-4B",
   "krea2-turbo"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "drops": [
   {
    "kind": "chip",
    "reason": "length",
    "item": "what impact does 24 GB of VRAM have on model selection?",
    "detail": "q not lowercase"
   }
  ]
 },
 "generated_utc": "2026-09-18T21:54:44Z"
}