{ "schema_version": "1.0", "page_kind": "article", "pack_kind": "generate", "slug": "a-short-history-of-the-rtx-3090", "title": "The RTX 3090: the gaming card that would not retire", "dek": "Announced in 2020 as a card for games at 8K, the GeForce RTX 3090 is the card this workshop has measured most. Its history: what it was built as, what replaced it, why its 24 GB still decides which open models run at home, what it holds against the cards that came after, and what it has cost, read by date through November 2026.", "published": "2026-11-06", "series": [ "notes" ], "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-short-history-of-the-rtx-3090/", "md_url": "https://research.strata2signal.com/a-short-history-of-the-rtx-3090/index.md", "md_sha": "0baad2594443f15cc861cf0506d9ff4bf712fb2edbe0a43fbbd72a4381851768", "html_sha": "d5dd2fd5be27648cc0bcaa0f2874b2dce1730fa1177c9f63eb933c69c6c6b561", "anchors_sha": "33395af0f7539f6ff6a32cf896bdc13f4efbbc582d5a8fd46dcc4bbcb1efff02" }, "short": { "paragraph": "The GeForce RTX 3090 went on sale on 24 September 2020 with 24 GB of GDDR6X memory, which NVIDIA's own architecture paper said was there so that \"data scientists\" could \"process large data sets\". Its faster twin, the RTX 3090 Ti, arrived on 29 March 2022 at $1,999; about six months later, in one September week, Ethereum mining ended, EVGA left the graphics-card business and NVIDIA announced the RTX 4090 with the same 24 GB. Then in 2023 came Meta's LLaMA models and llama.cpp, a program that runs such models on ordinary computers, and the number that decides whether a model runs at home turned out to be memory, not age: a 24-billion-parameter model at four bits takes about 17 GiB of a card once it serves a 32,768-token window, which fits a 3090 and does not fit a 16 GB card from 2025. On this shelf a 3090 writes with that model at 47.8 tokens a second held to 300 W, holds a 26-billion-parameter mixture model whole at a 131,072-token window, and sold used for $1,094.44 on average in August 2026 and $1,194.64 in September, both under its launch price; by another source's count, a typical used one sold for $1,499.00 in mid-October, the launch price to the dollar.", "from_dek": false, "counts": { "words": 3006, "minutes": 14, "tables": 1, "kit": false }, "bullets": [] }, "sections": [ { "id": "what-nvidia-sold-in-2020", "heading": "What NVIDIA sold in 2020", "level": 2, "span": [ 2437, 3768 ], "chunks": [ [ 2437, 3768 ] ], "chars": 1331, "digest": null, "digest_skipped": null }, { "id": "2022-the-ti-and-one-week-in-september", "heading": "2022: the Ti, and one week in September", "level": 2, "span": [ 3768, 5048 ], "chunks": [ [ 3768, 5048 ] ], "chars": 1280, "digest": "The GeForce RTX 3090 Ti launched on 29 March 2022 at $1,999. In September 2022, Ethereum's Merge ended mining on 15 September, EVGA announced it was leaving the graphics-card business on 16 September, and NVIDIA announced the GeForce RTX 4090 on 20 September.", "digest_skipped": null }, { "id": "2023-models-for-home-computers-and-memory-was-the-number", "heading": "2023: models for home computers, and memory was the number", "level": 2, "span": [ 5048, 10343 ], "chunks": [ [ 5048, 10343 ] ], "chars": 5295, "digest": null, "digest_skipped": null }, { "id": "what-it-does-on-this-shelf-measured", "heading": "What it does on this shelf, measured", "level": 2, "span": [ 10343, 14500 ], "chunks": [ [ 10343, 14500 ] ], "chars": 4157, "digest": null, "digest_skipped": null }, { "id": "against-the-cards-that-came-after", "heading": "Against the cards that came after", "level": 2, "span": [ 14500, 16250 ], "chunks": [ [ 14500, 16250 ] ], "chars": 1750, "digest": null, "digest_skipped": null }, { "id": "what-this-page-does-not-say", "heading": "What this page does not say", "level": 2, "span": [ 16250, 16770 ], "chunks": [ [ 16250, 16770 ] ], "chars": 520, "digest": "This section clarifies that no desktop RTX 4090 or RTX 5090 was measured, no wall-socket power figures are provided, and the prices listed are dated reads rather than a market.", "digest_skipped": null }, { "id": "who-ran-this-and-thanks", "heading": "Who ran this, and thanks", "level": 2, "span": [ 16770, 18060 ], "chunks": [ [ 16770, 18060 ] ], "chars": 1290, "digest": null, "digest_skipped": "credits" }, { "id": "sources", "heading": "Sources", "level": 2, "span": [ 18060, 20977 ], "chunks": [ [ 18060, 20977 ] ], "chars": 2917, "digest": "This section lists the sources used for the article, including NVIDIA announcements, architecture papers, Ethereum, Meta, GitHub, and various workshop pages, noting that all outside links were fetched on 2026-11-04 (UTC).", "digest_skipped": null } ], "chips": [ { "id": "c-2b030950", "q": "what were the launch prices for the rtx 30 series in 2020?", "a": "The RTX 3080 launched at $699 on 17 September, the RTX 3090 at $1,499 on 24 September, and the RTX 3070 at $499 in October.", "cites": [ "what-nvidia-sold-in-2020" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-aaa7745e", "q": "how did ethereum's merge affect graphics card usage in september 2022?", "a": "Ethereum's Merge ended mining on 15 September 2022, meaning cards previously used for hashing had no further role in producing valid blocks.", "cites": [ "2022-the-ti-and-one-week-in-september" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-abaeec5b", "q": "how does power consumption affect token generation speed on a 3090?", "a": "Increasing power from 250 W to 300 W on the dense model increases speed from 33.3 to 47.8 tokens a second.", "cites": [ "what-it-does-on-this-shelf-measured" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-a1ed28ce", "q": "what information is excluded from the provided measurements?", "a": "The article does not include measurements for desktop RTX 4090 or RTX 5090 cards, wall-socket electricity figures, or card failure rates.", "cites": [ "what-this-page-does-not-say" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" } ], "related": [ { "slug": "one-3090-against-one-3090-ti", "why": "shares ground with § 2022: the Ti, and one week in September · § A short history" }, { "slug": "a-sealed-3090-ti-after-evga-left", "why": "shares ground with § 2022: the Ti, and one week in September · § What EVGA said, in September 2022" }, { "slug": "a-3090-in-its-own-words", "why": "shares ground with § 2022: the Ti, and one week in September · § Part one: proof of work" } ], "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThanks to **NVIDIA**, whose 2020 and 2022 announcements, specification pages and GA102 architecture paper are dated and still up, which is why the chip's own numbers here are its maker's; to **EVGA**, which built many of the boards on this shelf, and **Gigabyte** and **ZOTAC**, which built one each of the cards quoted here; to **Micron**, named in NVIDIA's paper as the company it worked with to develop GDDR6X memory; to the **Ethereum** project for a Merge page that dates itself; to **Meta** for the LLaMA announcement; and to **GitHub** for the repository record that dates llama.cpp. Every measurement quoted stood on **Ollama** (MIT) over **llama.cpp** (MIT), **vLLM** (Apache-2.0) and **ComfyUI** (GPL-3.0), on **Gemma 4** (Google DeepMind, Apache 2.0), **Mistral Small 3.2** (Mistral AI, Apache-2.0) and **FLUX.2 klein 4B** (Black Forest Labs, Apache 2.0), with **ACE-Step 1.5** (MIT) under the music adapters, over **CUDA** (NVIDIA, proprietary, the one closed piece in the stack). None of them owed us anything. A small human team asked for this page, chose what it would and would not claim, and signed off on it; a fleet of AI agents fetched and read every source listed below and drafted it under that team's rulings.", "kit": null, "seat_class": { "writer": "gemma-class", "judge": null, "writer_runtime": "vllm" }, "bench": { "bullets_written": 3, "bullets_kept": 0, "digests_written": 7, "digests_kept": 3, "chips_written": 7, "chips_kept": 4, "chips_grounded": 0, "chips_published": 0, "dropped_by": { "new_noun": 0, "figure": 0, "length": 1, "cite": 0, "judge": 0, "quote": 0, "directive": 0, "redaction": 2, "profanity": 0, "identifier": 0, "bare_figure": 0, "house_voice": 0, "editorial": 0, "hand-edit": 7 }, "new_noun_tokens_checked": [ "1", "1.86", "10", "136.618", "14.14", "15", "16", "17", "20", "2020", "2022", "2026-11-04", "24", "24-billion-parameter", "24b", "250", "26b", "28.3", "29", "30", "300", "3070", "3080", "3090", "32", "33.3", "350", "4-bit", "4090", "47.8", "496", "499", "5080", "5090", "6000", "699", "768-token", "792", "936", "96", "999", "CORE", "CUDA", "EVGA", "Ethereum", "Ethereum's", "GA102", "GB", "GB/s", "GDDR6X", "GDDR7", "GHz", "GeForce", "GiB", "GitHub", "LLaMA", "March", "Merge", "Meta", "NVIDIA", "OC", "October", "Ollama", "Ollama's", "PRO", "RTX", "September", "Solid", "Ti", "UTC", "W", "ZOTAC", "gemma4", "mistral-small3.2" ], "new_noun_tokens_withheld": 2, "figure_definition": "v3", "figure_boundary": "guarded", "drops": [ { "kind": "digest", "reason": "redaction", "item": null, "item_chars": 311, "withheld": "a string carrying a fenced literal", "detail": "" }, { "kind": "digest", "reason": "redaction", "item": null, "item_chars": 204, "withheld": "a string carrying a fenced literal", "detail": "" }, { "kind": "chip", "reason": "length", "item": "how much memory is required for a 24-billion-parameter model at 4-bit quantization?", "detail": "q 83 chars" }, { "kind": "bullet", "reason": "hand-edit", "item": null, "item_chars": 39, "text_sha": "be06bae12755", "figure": "1.86", "withheld": "one bullet removed from this pack by hand", "removed_utc": "2026-11-06", "why": "the lead's ruling R2, 2026-11-06 (UTC): a maker spec, not a finding: it restates the boost clock NVIDIA's own specification page gives a 3090 Ti, 1.86 GHz, rather than anything the page measured or worked out (the rule the lead's E6 drew on exhibit 103)", "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: the lead's ruling R2, 2026-11-06 (UTC): a maker spec, not a finding: it restates the boost clock NVIDIA's own specification page gives a 3090 Ti, 1.86 GHz, rather than anything the page measured or worked out (the rule the lead's E6 drew on exhibit 103). Not a gate drop, and recorded here so …" }, { "kind": "bullet", "reason": "hand-edit", "item": null, "item_chars": 44, "text_sha": "8a43a4fa8350", "figure": "350", "withheld": "one bullet removed from this pack by hand", "removed_utc": "2026-11-06", "why": "the lead's ruling R2, 2026-11-06 (UTC): a maker spec, not a finding: it restates the rated board power NVIDIA's own architecture paper gives an RTX 3090, 350 W, rather than anything the page measured or worked out (the rule the lead's E6 drew on exhibit 103)", "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: the lead's ruling R2, 2026-11-06 (UTC): a maker spec, not a finding: it restates the rated board power NVIDIA's own architecture paper gives an RTX 3090, 350 W, rather than anything the page measured or worked out (the rule the lead's E6 drew on exhibit 103). Not a gate drop, and recorded her…" }, { "kind": "digest", "reason": "hand-edit", "item": null, "item_chars": 314, "text_sha": "9f71942c3462", "section": "2023-models-for-home-computers-and-memory-was-the-number", "withheld": "one section digest removed from this pack by hand", "removed_utc": "2026-11-06", "why": "hand-read 2026-11-06 (UTC): the digest says the article 'discusses how 24 GB of memory affects language model performance', where the section's point is two separate rules: memory decides whether a model fits whole (14.14 GiB of weights, about 17 GiB at a 32,768-token window, 20,129 to 23,469 MiB on a 24 GB card, refused at 131,072), and 'Speed follows bandwidth' sets how fast it writes; the digest folds the fit into speed, the confusion the section exists to undo", "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-06 (UTC): the digest says the article 'discusses how 24 GB of memory affects language model performance', where the section's point is two separate rules: memory decides whether a model fits whole (14.14 GiB of weights, about 17 GiB at a 32,768-token window, 20,129 to 23,469…" }, { "kind": "digest", "reason": "hand-edit", "item": null, "item_chars": 285, "text_sha": "1bbba5536429", "section": "what-nvidia-sold-in-2020", "withheld": "one section digest removed from this pack by hand", "removed_utc": "2026-11-06", "why": "hand-read 2026-11-06 (UTC): the digest hangs the architecture paper's one quoted sentence on the chip ('on a GA102 chip, which NVIDIA says allows data scientists to process large data sets'), where the section quotes the paper saying it is the 24 GB of memory that allows it, and makes that memory the reason the page exists", "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-06 (UTC): the digest hangs the architecture paper's one quoted sentence on the chip ('on a GA102 chip, which NVIDIA says allows data scientists to process large data sets'), where the section quotes the paper saying it is the 24 GB of memory that allows it, and makes that me…" }, { "kind": "bullet", "reason": "hand-edit", "item": null, "item_chars": 45, "text_sha": "0fe5b08d2363", "figure": "136.618", "withheld": "one bullet removed from this pack by hand", "removed_utc": "2026-11-06", "why": "hand-read 2026-11-06 (UTC): a bare figure: '136.618' loses its unit and its cap, both in the table cell it comes from ('136.139 tokens a second at 300 W; 136.618 at 350 W'), so a reader cannot tell what was counted or under what limit; the loss gate 8 exists for, which the house removed by hand on exhibits forty-seven and forty-eight", "detail": "removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-06 (UTC): a bare figure: '136.618' loses its unit and its cap, both in the table cell it comes from ('136.139 tokens a second at 300 W; 136.618 at 350 W'), so a reader cannot tell what was counted or under what limit; the loss gate 8 exists for, which the house removed by hand on exhibits forty-seven and forty-eight. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "chip", "reason": "hand-edit", "item": null, "item_chars": 226, "text_sha": "453a246f3fdb", "chip_id": "c-c6a518f3", "withheld": "one chip removed from this pack by hand", "removed_utc": "2026-11-06", "why": "hand-read 2026-11-06 (UTC): 'certain 24-billion-parameter models' misstates the section's two models, the dense 24-billion-parameter model and the 26-billion mixture model the page measures throughout, and the answer drops 'at any window', the condition the section gives for the 16 GB card holding neither whole", "detail": "removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-06 (UTC): 'certain 24-billion-parameter models' misstates the section's two models, the dense 24-billion-parameter model and the 26-billion mixture model the page measures throughout, and the answer drops 'at any window', the condition the section gives for the 16 GB card holding neither whole. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "chip", "reason": "hand-edit", "item": null, "item_chars": 189, "text_sha": "e53792978d58", "chip_id": "c-e2ec7aad", "withheld": "one chip removed from this pack by hand", "removed_utc": "2026-11-06", "why": "hand-read 2026-11-06 (UTC): the question asks when the architecture papers and announcements were published, and the answer gives no paper's date (the section gives the whitepaper none) and answers with the series' announcement day and a card's on-sale day, so the button promises what its answer does not give", "detail": "removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-06 (UTC): the question asks when the architecture papers and announcements were published, and the answer gives no paper's date (the section gives the whitepaper none) and answers with the series' announcement day and a card's on-sale day, so the button promises what its answer does not give. Not a gate drop, and recorded here so the ledger adds up." } ] }, "generated_utc": "2026-11-06T03:39:21Z" }