The RTX 3090

If you're new here: strata→signal is a small workshop (plus a friendly dog with a white patch) that builds on its own machines and writes up what it measures. Every outside date and figure below was read on 2026-11-04 (UTC) from the source credited beside it — NVIDIA's own announcements, specification pages and GA102 architecture paper, the Ethereum project's page on the Merge, Meta's LLaMA announcement and GitHub's record of one repository — and every measurement is quoted from the page of ours that published it. Where we found no source, the claim is not here.

What NVIDIA sold in 2020

On 1 September 2020 NVIDIA announced the GeForce RTX 30 series: the RTX 3080 at $699 from 17 September, the RTX 3090 at $1,499 from 24 September, and the RTX 3070 at $499 in October. The announcement called it the "BFGPU" ("Big Ferocious GPU"), promised 60 frames a second at 8K in many games, and put it "up to 50 percent faster than the current ultimate PC graphics card, TITAN RTX", NVIDIA's previous 24 GB card.

The chip is GA102: 28.3 billion transistors on a 628.4 mm² die, made on Samsung's 8 nm process. An RTX 3090 uses 82 of its streaming multiprocessors, 10,496 CUDA cores (the small arithmetic units a graphics card has thousands of) and 24 GB of GDDR6X memory at 19.5 Gbps on a 384-bit bus, which is 936 GB/s of memory bandwidth: how many bytes a second the chip can pull from its own memory. Its rated board power is 350 W. The RTX 3080 beside it carried 10 GB.

The 24 GB was never only for games, and NVIDIA's paper says so in one sentence: "Like the previous generation TITAN RTX, it ships with 24 GB of memory, allowing data scientists to process large data sets, while gamers can experience next-generation gaming with ray tracing and DLSS 8K." An RTX 3090 brought a Titan's memory into the GeForce line, and that memory is the reason this page exists.

2022: the Ti, and one week in September

On 29 March 2022 the GeForce RTX 3090 Ti went on sale at $1,999: 10,752 CUDA cores, "24GB of the fastest 21Gbps GDDR6X memory" in NVIDIA's words, a 1.86 GHz boost clock against a 3090's 1.70, and 450 W of board power against 350. It was the last and fastest card of the line. Our bench of one against the other is RTX 3090 vs RTX 3090 Ti.

Then came one week in September 2022. On the 15th, Ethereum's Merge ended mining — "Mining is no longer the means of producing valid blocks", in the project's own words — and the cards that had been hashing had nothing to do; the two 3090s this workshop bought used in 2021 had spent about a year at it. On the 16th, EVGA, which built many of the cards on this shelf, told two interviewers it was leaving the graphics-card business; it never published a last-built date, and a sealed 3090 Ti after EVGA left tells that story. On the 20th, NVIDIA announced the GeForce RTX 4090: 16,384 CUDA cores, the same 24 GB of GDDR6X, $1,599, on sale 12 October. A card whose first job had ended, whose maker on this shelf had quit and whose successor was weeks away should have been finished.

2023: models for home computers, and memory was the number

On 24 February 2023 Meta released LLaMA, language models of 7, 13, 33 and 65 billion parameters, to researchers under a non-commercial licence (the weights that got out). The llama.cpp repository, which runs such models on ordinary computers, was created on 10 March 2023; Ollama, which wraps it in one command, shipped its first release on 8 July 2023 and its Linux build with NVIDIA acceleration on 23 September 2023 (a short history of Ollama).

A graphics card writes a language model's answer one token at a time (a token is a little under a word), and for a dense model every token means reading every weight once. At four-bit quantisation, the usual compression for home use, a 24-billion-parameter model is 14.14 GiB of weights and takes about 17 GiB of a card once it serves a 32,768-token context window, the text the model holds in mind, reserved up front (a short history of Mistral). Speed follows bandwidth: at 350 W, on that page's cap ladder, a 3090 writes with that model at 50.02 tokens a second, about four fifths of its 936 GB/s spent fetching weights.

What 24 GB holds, measured on an EVGA GeForce RTX 3090 XC3 Ultra in RTX 3090 vs RTX 3090 Ti: the 26-billion-parameter mixture-of-experts model gemma4:26b (a model that wakes only about 4 billion of its parameters for each token) whole at a 131,072-token window, 20,129 MiB on the card; the dense mistral-small3.2:24b whole at 65,536 tokens with placement left to Ollama, 20,589 MiB, and at 98,304 once the layer count is named, 23,469 MiB; at 131,072 it refuses to load. A ZOTAC GeForce RTX 5080 Solid CORE OC, 16 GB, a 2025 card, did not hold either model whole at any window under Ollama's own planner, as Ollama ships them (RTX 5080 vs RTX 3080 Ti): an earlier build of gemma4:26b, 588 MiB smaller, took 17,652 to 17,926 MiB on the 24 GB cards at a 4,096-token window, more than that card's 16,303 MiB.

And the price fell under what it launched at, then climbed back to it. The rig page read prices on 13 September 2026: a used RTX 3090 sold for $1,094.44 on average in August 2026, by one marketplace's own record; the cheapest new one of any brand asked $1,829.99; a renewed EVGA GeForce RTX 3090 XC3 Ultra asked $1,849.99, and $1,879.99 four days later. On those readings a new or renewed card asked more than NVIDIA's $1,499 of 2020, and a used one cost less. A later read, on 8 October 2026 for a 10 GB RTX 3080 desktop, priced, put the same marketplace's average completed sale for September 2026 at $1,194.64, up from $680 in October 2025, the lowest month of its twelve. Prices read again on 1 November 2026 for the one-card RTX 3090 rig, repriced put the cheapest new one of any brand at $2,399.00 and a typical used one, the median (the middle sale: half sold above it, half below) of 161 completed sales on a second marketplace between 8 and 15 October as a buyer's guide counted them, at $1,499.00: NVIDIA's launch price to the dollar. A median on one marketplace and an average on another are different rulers; on the guide's own two readings, the nearest thing to one ruler, its used median rose 17.6 per cent, from $1,275 in August to $1,499 in October, the figure that page leads with.

What it does on this shelf, measured

Nothing here was measured for this page: every row was published earlier, and names the page. Language rows are four-bit (Q4_K_M) builds served by Ollama unless the page says otherwise; every watt is a cap on board power, read from the driver, never a wall socket. What that board power costs in electricity, per million tokens and per picture, is on electricity per million tokens, card by card and electricity per image, card by card.

Two things the table cannot show. The cap moves a 3090 more than its name does: 33.3 to 47.8 tokens a second on the dense model is fifty watts, "a bigger jump than any gap between two different cards at the same cap in the dense model's table" on that page. And whether two cheaper cards beat one 3090 has two answers here: two used 10 GB GeForce RTX 3080s, 20 GB between them, read a long document in two thirds of the time and beat one 3090 on the dense model at 250 W, at a window both could hold, but the single card was faster on three of the four things measured and used about half the energy at 300 W (two 3080s against one 3090).

The same cards did other work: one served the dense model behind the workshop's public products at 300 W until 25 September 2026, when a 3090 Ti took its slot at 350 W (one 3090 out, one 3090 Ti in); four of this workshop's five music adapters were trained on one RTX 3090 held to 250 W (the rig page); one re-drew the workshop's own game art, render for render against the 96 GB card's archive, 3.4 to 4.9 times slower at 768 pixels (a rig your friend already owns); one was the print lab's painter until 18 September 2026 (two 3090s out, one 3090 Ti in); one ran Mistral NeMo 12B at five quantization levels at 300 W, 60.5 tokens a second at eight bits to 88.2 at four (the quantization ladder), and painted FLUX 3 Image Klein on its release day, 15 October 2026, and again with its maker's prompt upsampler (ranked blind in a game's own painted style; run again with its maker's prompt upsampler); and one of them wrote its own story, through the model it serves (a 3090, in its own words).

Against the cards that came after

By memory, the one number that decides what fits:

• GeForce RTX 4090, announced 20 September 2022 at $1,599: 24 GB of GDDR6X, the same 24. Never measured here.
• GeForce RTX 5080, announced 6 January 2025 at $999: 16 GB. Measured here, and under Ollama's own planner it held neither of the two models above whole, at any window.
• GeForce RTX 5090, announced the same day at $1,999, on sale 30 January 2025: 32 GB of GDDR7, 21,760 CUDA cores, 575 W. Not measured here as a desktop card. The laptop part that carries its name, the NVIDIA GeForce RTX 5090 Laptop GPU (24 GB), was: held to 95 W it wrote 118.4 tokens a second on the mixture model to a 3090's 136.6 at 350 W, and with NVIDIA's Dynamic Boost running, as such a laptop arrives, 155.3, past a 3090 and a 3090 Ti at its own 450 W (a laptop, asked the desktop's questions).
• RTX PRO 6000 Blackwell: 96 GB of GDDR7, 600 W. Measured here against a 3090 in one box, at 420 W against 250: a 3090 was 39 per cent behind it on the mixture model and 63 per cent on the dense (a short history of Mistral).

That last page explains the gap in one line: "Unthrottled, both cards read the dense model at about four fifths of their own memory bandwidth" — 936 GB/s on a 3090 against the 1,792 GB/s NVIDIA publishes for the 96 GB card. On the dense model, then, a 3090 trails that card by about the ratio of their bandwidths, and it still holds what a 16 GB card of 2025 cannot. The 24 GB that was for data scientists in 2020 is, in 2026, the least memory of any single card on this shelf that holds a 24-billion-parameter model whole.

What this page does not say

• Nothing measured on a desktop RTX 4090 or RTX 5090. Neither has been in this workshop.
• No wall-socket figure. Every watt above is a cap on board power, from the driver; the meter readings for the box two of these cards sat in are on what it costs to run a high-end AI rig at home.
• Prices are dated reads, each taken by the page linked beside it on the day named, not a market; and no failure rate, for any card.

Who ran this, and thanks

Thanks 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.

Sources

Every outside link below was fetched and read on 2026-11-04 (UTC); our own pages were read the same day in the markdown copy each publishes beside itself.

• NVIDIA — the GeForce RTX 30 Series announcement, 1 September 2020 (prices, dates, "BFGPU", 8K, "up to 50 percent faster than the current ultimate PC graphics card, TITAN RTX")
• NVIDIA — the GA102 GPU architecture whitepaper, v2 (28.3 billion transistors, 628.4 mm², Samsung's 8 nm; 82 SMs, 10,496 CUDA cores, 19.5 Gbps, 936 GB/s and 350 W in Appendix A, Table 9; the "data scientists" sentence; the RTX 3080's 10,240 MB in Table 2; Micron)
• NVIDIA — GeForce RTX 3090 and 3090 Ti specifications (10,752 and 10,496 CUDA cores; boost clocks 1.86 and 1.70 GHz; 450 and 350 W)
• NVIDIA — GeForce RTX 3090 Ti Out Now, 29 March 2022 ("Starting at $1,999"; "24GB of the fastest 21Gbps GDDR6X memory")
• NVIDIA — the GeForce RTX 40 Series announcement, 20 September 2022 (the RTX 4090's $1,599, 12 October, 24 GB of GDDR6X, 16,384 CUDA cores)
• NVIDIA — the GeForce RTX 50 Series announcement, 6 January 2025 (the RTX 5090's $1,999 and 30 January; the RTX 5080's $999), and the GeForce RTX 5090 specifications (32 GB GDDR7, 21,760 CUDA cores, 575 W) and GeForce RTX 5080 specifications (16 GB GDDR7)
• NVIDIA — RTX PRO 6000 Blackwell Workstation Edition (96 GB GDDR7, 1,792 GB/s, 600 W)
• Ethereum — The Merge ("The Merge was executed on September 15, 2022"; "Mining is no longer the means of producing valid blocks")
• Meta — Introducing LLaMA, ai.meta.com/blog/large-language-model-llama-meta-ai/, 24 February 2023 (the four sizes; the non-commercial research licence)
• GitHub — ggml-org/llama.cpp (repository created 2023-03-10; MIT)
• This workshop's own pages, each linked where it is quoted above

Corrections and later measurements will be added below, each dated (UTC), with a window at both ends where one applies and saying in plain words what it counts.