{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "one-3090-against-one-3090-ti",
 "title": "RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench",
 "dek": "One EVGA GeForce RTX 3090 XC3 Ultra 24 GB and one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB, in the same x16 slot of the same desktop, on the same supply and the same script, 17–18 September 2026. Numbers only: what each writes, what each holds, what a picture costs, how hot each ran, and what a power cap takes away.",
 "published": "2026-09-18",
 "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/one-3090-against-one-3090-ti/",
  "md_url": "https://research.strata2signal.com/one-3090-against-one-3090-ti/index.md",
  "md_sha": "46e255f604eeca343573c8c779ad4adf4c69943c4133f5fed39808bc05af99b2",
  "html_sha": "4178b2fc871bad3ed9d5c3940e89d58039db9265d8a75fc930982092f6ac0f9d",
  "anchors_sha": "a0a01a41f1266116106bbb49399ea360e007a088410f0921d0828f62741a9dba"
 },
 "short": {
  "paragraph": "This 3090 Ti holds its memory at 10,251 MHz and a 3090 at 9,501 - 7.9 per cent more bandwidth, present at every setting, but whether the work can reach it is decided by a number nobody sets: held at 300 W a Ti pulls 66.3 per cent of the ceiling that clock implies, at its own 450 W limit 88.5. Run each board at the limit it ships with and this Ti is 8.0 per cent faster on a mixture model, 22.4 on a dense one and 9 to 12 per picture, for 17.7 and 53.2 per cent more board power. It cost $2,119.99 new; a renewed listing of the exact 3090 read $1,879.99.",
  "from_dek": false,
  "counts": {
   "words": 6236,
   "minutes": 28,
   "tables": 10,
   "kit": false
  },
  "bullets": [
   {
    "text": "the 3090 Ti was ordered on 15 September 2026",
    "figure": "15 September 2026",
    "cite": "the-two-cards-named-once",
    "quote": "Ordered 15 September 2026, unboxed on the 17th, first powered on for this bench - and given no burn-in: the first hours of load in its life are the four cap blocks on this page, its language arms 19:07 to 23:46 UTC and its renders 19:41 to 00:01.",
    "span": [
     2152,
     2170
    ]
   },
   {
    "text": "the XC3 Ultra weighs 1,203 g",
    "figure": "1,203 g",
    "cite": "the-card-itself-as-an-object",
    "quote": "Measured on 17 September 2026 by an operator with a kitchen scale and a tape measure, not a caliper or a bench balance: the XC3 Ultra 1,203 g and about 2 inches thick, the FTW3 Ultra Ti 2,157 g and about 2.75 inches - 954 g heavier, 1.79 times the weight, and about three quarters of an inch thicker.",
    "span": [
     27468,
     27478
    ]
   },
   {
    "text": "the GeForce RTX 3090 Ti went on sale on 29 March 2022",
    "figure": "29 March 2022",
    "cite": "a-short-history",
    "quote": "A short history The GeForce RTX 3090 Ti went on sale on 29 March 2022 at $1,999, in NVIDIA's own words with \"24GB of the fastest 21Gbps GDDR6X memory\" - the last and fastest of the GeForce RTX 30 line's flagships, six months before the generation above it replaced it, and its 2 GB chips are what let the whole 24 GB sit on one face of the board.",
    "span": [
     28375,
     28391
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "the-two-cards-named-once",
   "heading": "The two cards, named once",
   "level": 2,
   "span": [
    1382,
    4395
   ],
   "chunks": [
    [
     1382,
     4395
    ]
   ],
   "chars": 3013,
   "digest": "the article compares an EVGA GeForce RTX 3090 XC3 Ultra Gaming 24 GB and an EVGA GeForce RTX 3090 Ti FTW3 Ultra Gaming 24 GB. the 3090 is a used board from 2021, while the Ti is a new unit ordered 15 September 2026. both are stock with factory thermal pads and coolers.",
   "digest_skipped": null
  },
  {
   "id": "at-a-glance",
   "heading": "At a glance",
   "level": 2,
   "span": [
    4395,
    7451
   ],
   "chunks": [
    [
     4395,
     7451
    ]
   ],
   "chars": 3056,
   "digest": "a summary table compares the two cards across various workloads including mixture models, dense models, and renders. it notes that the 3090 Ti wins most throughput rows, while the 3090 is more energy efficient in certain configurations. the table also compares pricing and core temperatures.",
   "digest_skipped": null
  },
  {
   "id": "the-question",
   "heading": "The question",
   "level": 2,
   "span": [
    7451,
    8638
   ],
   "chunks": [
    [
     7451,
     8638
    ]
   ],
   "chars": 1187,
   "digest": "the text explores whether the 3090 Ti's memory bandwidth advantage of 21Gbps per pin against 19.5 is reflected in performance. both boards held their memory one bin under their own reported maximum. the comparison focuses on how writing a token on a quantised model waits on weight streaming.",
   "digest_skipped": null
  },
  {
   "id": "the-instrument-and-the-control",
   "heading": "The instrument, and the control",
   "level": 2,
   "span": [
    8638,
    9614
   ],
   "chunks": [
    [
     8638,
     9614
    ]
   ],
   "chars": 976,
   "digest": "the testing used a single harness with a frozen 2,101-byte prompt and median reported results. the control is the slot, with one card swapped for the other in the same case. models used include gemma4:26b and mistral-small3.2:24b on ollama 0.32.13.",
   "digest_skipped": null
  },
  {
   "id": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "heading": "What a power cap takes, and what each card does at its own limit",
   "level": 2,
   "span": [
    9614,
    16705
   ],
   "chunks": [
    [
     9614,
     16705
    ]
   ],
   "chars": 7091,
   "digest": "this section details performance at various power caps. the 3090 Ti wins both models at its own 450 W limit. the 3090 is more energy efficient on the dense model at shared caps. the text notes that a power cap slows the core, not the memory.",
   "digest_skipped": null
  },
  {
   "id": "how-a-cap-is-set",
   "heading": "How a cap is set",
   "level": 2,
   "span": [
    16705,
    17697
   ],
   "chunks": [
    [
     16705,
     17697
    ]
   ],
   "chars": 992,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-two-models-want-different-caps",
   "heading": "The two models want different caps",
   "level": 2,
   "span": [
    17697,
    19341
   ],
   "chunks": [
    [
     17697,
     19341
    ]
   ],
   "chars": 1644,
   "digest": "the mixture model reaches a knee at 350 W, while the dense model shows no knee. the Ti's mixture model performance sits within the run-to-run spread above 350 W. the text notes that board power includes the fans.",
   "digest_skipped": null
  },
  {
   "id": "renders",
   "heading": "Renders",
   "level": 2,
   "span": [
    19341,
    22510
   ],
   "chunks": [
    [
     19341,
     22510
    ]
   ],
   "chars": 3169,
   "digest": "the Ti takes every graph at both shared 300 W and 350 W caps. the Ti's ladder saturates at 400 W, as the cap flag collapses. the 3090 remains power-limited at its own 350 W limit.",
   "digest_skipped": null
  },
  {
   "id": "what-fits-and-what-the-cache-costs",
   "heading": "What fits, and what the cache costs",
   "level": 2,
   "span": [
    22510,
    24264
   ],
   "chunks": [
    [
     22510,
     24264
    ]
   ],
   "chars": 1754,
   "digest": "both cards hold the same largest windows for mixture and dense models. the 24 GB capacity is identical for both. using an unquantised cache is fastest on both boards at 350 W.",
   "digest_skipped": null
  },
  {
   "id": "three-more-arms",
   "heading": "Three more arms",
   "level": 2,
   "span": [
    24264,
    26676
   ],
   "chunks": [
    [
     24264,
     26676
    ]
   ],
   "chars": 2412,
   "digest": "the Ti leads in throughput for multiple simultaneous conversations. a short gate-shaped call is the one arm where the boards are indistinguishable. the Ti is also cheaper to leave on based on median idle power.",
   "digest_skipped": null
  },
  {
   "id": "the-card-itself-as-an-object",
   "heading": "The card itself, as an object",
   "level": 2,
   "span": [
    26676,
    28297
   ],
   "chunks": [
    [
     26676,
     28297
    ]
   ],
   "chars": 1621,
   "digest": "the 3090 Ti is heavier and thicker than the 3090. the XC3 Ultra weighs 1,203 g and is about 2 inches thick, while the FTW3 Ultra Ti weighs 2,157 g and is about 2.75 inches thick.",
   "digest_skipped": null
  },
  {
   "id": "a-short-history",
   "heading": "A short history",
   "level": 2,
   "span": [
    28297,
    29397
   ],
   "chunks": [
    [
     28297,
     29397
    ]
   ],
   "chars": 1100,
   "digest": "the GeForce RTX 3090 Ti launched on 29 March 2022. EVGA announced it was ending its partnership with NVIDIA on 16 September 2022 and would build no further graphics cards.",
   "digest_skipped": null
  },
  {
   "id": "what-it-cost",
   "heading": "What it cost",
   "level": 2,
   "span": [
    29397,
    30500
   ],
   "chunks": [
    [
     29397,
     30500
    ]
   ],
   "chars": 1103,
   "digest": "the Ti was bought new for $2,119.99 on 15 September 2026. a renewed listing for the 3090 read $1,879.99 on 17 September 2026.",
   "digest_skipped": null
  },
  {
   "id": "what-the-240-buys-on-those-two-readings",
   "heading": "What the $240 buys, on those two readings",
   "level": 2,
   "span": [
    30500,
    32230
   ],
   "chunks": [
    [
     30500,
     32230
    ]
   ],
   "chars": 1730,
   "digest": "the $240 difference between the two prices is analyzed. per dollar, the 3090 wins most rows except for the dense model. the Ti provides 22.4 per cent more performance past the 3090's limit.",
   "digest_skipped": null
  },
  {
   "id": "the-setting-this-page-would-choose-and-the-dissent",
   "heading": "The setting this page would choose, and the dissent",
   "level": 2,
   "span": [
    32230,
    33221
   ],
   "chunks": [
    [
     32230,
     33221
    ]
   ],
   "chars": 991,
   "digest": "a pre-written rule sets the choice at 350 W. the dissent notes that the dense model is not done at 350 W, gaining 6.71 per cent at 400 W. 450 W is a measurement, not a recommendation.",
   "digest_skipped": null
  },
  {
   "id": "what-this-page-does-not-measure",
   "heading": "What this page does not measure",
   "level": 2,
   "span": [
    33221,
    34847
   ],
   "chunks": [
    [
     33221,
     34847
    ]
   ],
   "chars": 1626,
   "digest": "the article does not measure answer quality, memory-die temperature, wall power, or loudness. it also did not run a 3090 Ti at 250 W or a 3090 at 400 W or 450 W.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    34847,
    35691
   ],
   "chunks": [
    [
     34847,
     35691
    ]
   ],
   "chars": 844,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    35691,
    37026
   ],
   "chunks": [
    [
     35691,
     37026
    ]
   ],
   "chars": 1335,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-43ef8fbd",
   "q": "how do the two boards differ in their history and condition?",
   "a": "One board is a used unit from 2021 that previously mined Ethereum and flew Flight Simulator, while the other is a new, unboxed unit that received no burn-in before testing. Both remain on their factory thermal pads and BIOS.",
   "cites": [
    "the-two-cards-named-once"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-191365ed",
   "q": "how does memory bandwidth affect the performance of these models?",
   "a": "The Ti's advantage lies in its 21Gbps per pin memory bandwidth compared to 19.5 for the 3090. Because writing tokens on quantized models depends on how fast a card streams weights, this difference is expected to manifest in the results.",
   "cites": [
    "the-question"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-3c8c1184",
   "q": "what is the relationship between a power cap and core clock speed?",
   "a": "A power cap does not slow the memory down; it slows the core down. A core with no power cannot drain a fast bus. The driver reports whether the cap held the clock down via a cap-active flag.",
   "cites": [
    "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-4bea70c3",
   "q": "how does the mixture model respond to increasing power limits?",
   "a": "The mixture model reaches its knee at 350 W. Increasing the power beyond 350 W results in performance that sits within the natural run-to-run spread, showing no significant gains.",
   "cites": [
    "the-two-models-want-different-caps"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-04da5f34",
   "q": "how does the ti perform in image generation tasks?",
   "a": "The Ti takes every graph at both shared 300 W and 350 W caps. When running as they come, the Ti is 9 to 12 per cent faster than the 3090 across the tested graphs.",
   "cites": [
    "renders"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-cb5d9fc6",
   "q": "what are the physical differences between the two boards?",
   "a": "The Ti is significantly larger and heavier, weighing 2,157 g and measuring about 2.75 inches thick, whereas the 3090 weighs 1,203 g and is about 2 inches thick. The Ti is 954 g heavier and 1.79 times the weight.",
   "cites": [
    "the-card-itself-as-an-object"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "two-used-3080s-priced",
   "why": "shares ground with § The instrument, and the control · § The instrument, and how the control was run"
  },
  {
   "slug": "home-inference-and-diffusion-rig-cost",
   "why": "shares ground with § The two cards, named once · § Why these two are EVGA, and why the cheapest card is a different row"
  },
  {
   "slug": "what-150-watts-buys",
   "why": "shares ground with § What a power cap takes, and what each card does at its own limit · § The answer, in one paragraph"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThe cards are an **EVGA GeForce RTX 3090 XC3 Ultra 24 GB** and an **EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB**, and **EVGA** gets the credit twice: for boards still doing serious work four and five years after they were sold for games, and for publishing the figures this page checked itself against. **NVIDIA** published the launch specifications quoted in the history, and its driver is the instrument every reading came out of.\n\nThe language work stands on **ollama** and, beneath it, **llama.cpp** and **NVIDIA's CUDA**; the pictures on **ComfyUI**. The models are Google's **Gemma 4**, **Mistral Small 3.2** from Mistral AI, Black Forest Labs' **FLUX.2 Klein 4B** and its decoder, which drew every picture timed here, and **Qwen** from Alibaba Cloud as the text encoder in those graphs. None of them owed us anything, and several ask for no credit, which is exactly why it is given.\n\nA small human team owns the cards, moved one out of the slot and the other in, set the rules and the refusals before the runs and signed the numbers; a fleet of AI agents ran the harness and did the arithmetic under that team's rulings. Thanks to the readers of the page before this one, who asked the question this one answers: *what happens if you put the faster card in the same slot?*",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 3,
  "digests_written": 16,
  "digests_kept": 15,
  "chips_written": 8,
  "chips_kept": 6,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 2,
   "figure": 1,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 0
  },
  "new_noun_tokens_checked": [
   "0.32.13",
   "072",
   "1",
   "1.79",
   "101-byte",
   "119.99",
   "12",
   "131",
   "15",
   "157",
   "16",
   "17",
   "19.5",
   "2",
   "2.75",
   "2021",
   "2022",
   "2026",
   "203",
   "21Gbps",
   "22.4",
   "24",
   "240",
   "24b",
   "250",
   "26b",
   "29",
   "300",
   "3090",
   "3090's",
   "350",
   "366",
   "400",
   "450",
   "480",
   "6.71",
   "879.99",
   "9",
   "954",
   "BIOS",
   "EVGA",
   "Ethereum",
   "FTW3",
   "Flight",
   "GB",
   "Gaming",
   "GeForce",
   "March",
   "NVIDIA",
   "RTX",
   "September",
   "Simulator",
   "Ti",
   "Ti's",
   "Ultra",
   "W",
   "XC3",
   "gemma4",
   "mistral-small3.2"
  ],
  "new_noun_tokens_withheld": 2,
  "figure_definition": "v3",
  "drops": [
   {
    "kind": "digest",
    "reason": "figure",
    "item": "caps were set using the driver's own tool, nvidia-smi, with persistence mode enabled. the allowed range for the 3090 is 350 W to 366 W, and for the Ti is 450 W to 480 W. this is a power limit, not an ",
    "detail": "figures absent: ['3090']"
   },
   {
    "kind": "chip",
    "reason": "new_noun",
    "item": null,
    "item_chars": 4,
    "withheld": "a string naming something the article does not",
    "detail": ""
   },
   {
    "kind": "chip",
    "reason": "new_noun",
    "item": null,
    "item_chars": 4,
    "withheld": "a string naming something the article does not",
    "detail": ""
   }
  ]
 },
 "generated_utc": "2026-09-18T03:43:08Z"
}