NVIDIA GeForce RTX 4090

The NVIDIA GeForce RTX 4090 is a consumer graphics card built on the AD102 processor of NVIDIA's Ada Lovelace architecture. NVIDIA announced it on September 20, 2022, and it went on sale on October 12, 2022 at a starting price of $1,599 ^[1]. The card pairs 16,384 CUDA cores and 512 fourth-generation Tensor Cores with 24 GB of GDDR6X memory at 1,008 GB/s, and it was the first GeForce card with FP8 Tensor Core support ^[2]. Its 24 GB of memory made it a common platform for running and fine-tuning open-weight language models, generating images and doing academic research, and GPU marketplaces rent it out by the hour ^[3][4][5][6]. From November 2023 it could not be shipped to China without a US export license, which led NVIDIA to sell a cut-down RTX 4090 D there, and press reports said production ended in late 2024, before its successor, the RTX 5090, arrived in January 2025 ^{[7][8][9][10]}.

Development and launch

NVIDIA introduced the GeForce RTX 40 series with the RTX 4090 as its flagship, alongside two configurations of the RTX 4080 ^[1]. The company called the RTX 4090 "the world's fastest gaming GPU" and said it was up to twice as fast as the previous-generation RTX 3090 Ti in existing games while drawing the same 450 W ^[1]. NVIDIA built a limited run of Founders Edition cards itself, and add-in-board partners including ASUS, Gigabyte, MSI, PNY and Zotac sold their own designs ^[1].

The RTX 4090 was the first GeForce card based on Ada Lovelace ^[2]. Its AD102 die is made on TSMC's 4N process and holds 76.3 billion transistors in 608.5 mm² ^[2]. A full AD102 has 144 streaming multiprocessors (SMs) with 18,432 CUDA cores and 576 Tensor Cores; the RTX 4090 enables 128 SMs, giving 16,384 CUDA cores and 512 Tensor Cores ^[2]. The same silicon is used in NVIDIA's professional RTX 6000 Ada Generation card and in its data-center L40 and L40S accelerators ^[7].

The card dropped the NVLink bridge connector that NVIDIA had offered for linking earlier high-end GeForce cards ^[11]. Jensen Huang told reporters after the announcement that NVIDIA had needed the chip's input/output area for something else and had used it to add more AI processing, and he said multi-GPU systems could exchange data peer to peer over PCI Express instead ^[11]. The RTX 4090 shipped with a PCIe 4.0 interface ^[2]. In February 2023 an NVIDIA representative confirmed in a user's forum thread that "Peer to Peer is not supported on 4090," after Puget Systems and owners had found that peer-to-peer tests such as NVIDIA's simpleP2P sample failed on dual-card systems ^[12]. Data moving between two RTX 4090s therefore travels through system memory instead of directly from card to card ^[12].

Specifications

The figures come from Appendix A of NVIDIA's Ada whitepaper ^[2], NVIDIA's product page ^[13] and the launch announcement ^[1].

How fast is the RTX 4090 at AI math?

NVIDIA's Ada whitepaper lists peak Tensor Core throughput at the 2.52 GHz boost clock in two forms: a dense rate and an "effective" rate that assumes 2:4 structured sparsity and is exactly twice the dense figure ^[2]. The rates also depend on whether the Tensor Cores accumulate their results in FP32 or in FP16 ^[2].

The "1,321 AI TOPS" on NVIDIA's product page matches the sparse INT8 and sparse FP8 (FP16-accumulate) entries ^[13][2]. NVIDIA corrected the FP8 FP32-accumulate figure in version 2.02 of the whitepaper ^[2]. The FP4 entry comes from NVIDIA's later RTX Blackwell whitepaper, which lists FP4 as unavailable on the RTX 4090 ^[14].

Why the FP16-accumulate figures are not comparable with data-center specifications

On GeForce cards, NVIDIA lists Tensor Core throughput with FP32 accumulation at half the rate with FP16 accumulation ^[2]. Its data-center GPUs carry no such penalty: the H100 SXM5 is rated at 989.4 dense TFLOPS for FP16 with either FP16 or FP32 accumulation, and at 1,978.9 dense TFLOPS for FP8 with either ^[15]. A data-center headline number therefore corresponds to the RTX 4090's FP32-accumulate rows, and setting it beside the GeForce card's doubled FP16-accumulate figures overstates the consumer card for any work that accumulates in FP32. Compared on the same basis, the RTX 4090 has about one-sixth of an H100 SXM5's dense FP16, BF16 and FP8 throughput, under a third of its memory bandwidth and less than a third of its memory ^[2][15].

The H100 uses high-bandwidth memory (HBM3) rather than GDDR6X; its figures are from version 1.04 of NVIDIA's Hopper whitepaper ^[15].

How is the RTX 4090 used for AI?

Local inference

For inference, memory capacity decides which models fit on one card. Answer.AI's March 2024 arithmetic for a 70-billion-parameter model is that even after 4-bit quantization its weights take 35 GB, "larger than the 24GB gaming GPUs" it wanted to use, so such a model needs a second card or partial offloading to the CPU ^[4][5]. Smaller models fit more easily. When OpenAI released its gpt-oss open-weight models in August 2025, NVIDIA said that on RTX PCs they could be tried in Ollama on GPUs with at least 24 GB of memory, and in other applications on GPUs with at least 16 GB ^[16].

The card is supported by the common local inference tools. Ollama's documentation lists the RTX 4090 among supported GPUs with CUDA compute capability 8.9 ^[17], and NVIDIA says it contributes performance work to llama.cpp and the GGML tensor library for its RTX cards ^[16]. The serving engine vLLM performs native FP8 computation only on GPUs with compute capability 8.9 or higher, a group that starts with Ada Lovelace and Hopper; on older cards such as the Ampere-generation RTX 3090 it falls back to weight-only FP8 ^[18].

Research systems have stretched the card further. PowerInfer, an inference engine described in a December 2023 paper, keeps frequently activated "hot" neurons on the GPU and computes the rest on the CPU. Its authors reported that on a single RTX 4090 it ran up to 11.69 times faster than llama.cpp, and that on the OPT-30B model it reached 82% of the token generation rate of a server-grade A100 ^[5].

Fine-tuning and training

Lambda's October 2022 training benchmarks, run in NVIDIA's PyTorch container, found that the RTX 4090 trained vision, language, speech and recommendation models 1.3 to 1.9 times as fast as the RTX 3090 in TF32 and 1.3 to 1.8 times as fast in FP16 mixed precision. At list prices of $1,599 and $1,400 that worked out to 1.2 to 1.6 times the training throughput per dollar, while throughput per watt was similar to the older card; most models trained close to twice as fast on two GPUs as on one ^[3].

Parameter-efficient methods brought larger models within reach. The QLoRA paper of May 2023 reduced memory use enough to fine-tune a 65-billion-parameter model on a single 48 GB GPU ^[19]. In March 2024 Answer.AI released an open-source system that combines QLoRA with PyTorch's Fully Sharded Data Parallel (FSDP) to spread a quantized model across several cards; the company said it could train a 70-billion-parameter model on a desktop with two or more 24 GB gaming GPUs such as the RTX 3090 or RTX 4090 ^[4]. Small multi-GPU machines were built around the card as well: Tiny Corp's TinyBox, which went on sale in August 2024, cost $25,000 in its "green" configuration with six RTX 4090s and 144 GB of combined GPU memory, against $15,000 for a version with six AMD Radeon RX 7900 XTX cards ^[20].

Image generation

In Tom's Hardware's Stable Diffusion tests published in December 2023, the RTX 4090 generated 75 Stable Diffusion 1.5 images per minute at 512x512 with NVIDIA's TensorRT optimizations, about three times the 26 images per minute of AMD's fastest card at the time, the Radeon RX 7900 XTX. At 1024x1024 it produced 13.4 images per minute ^[21]. NVIDIA has said that Black Forest Labs' FLUX.dev model needs more than 23 GB of memory at FP16, so that among GeForce cards only the RTX 4090 and RTX 5090 can run it at that precision, and that the RTX 4090 generates a 30-step FLUX.dev image in 15 seconds ^[14].

Benchmark results

Phoronix's January 2025 tests with Meta's Llama 3.1 8B also found the RTX 5090 about 17% faster than the RTX 4090 in prompt processing, and judged the two cards comparable in performance per watt for text generation ^[22].

Export controls

October 2023 rules

On October 17, 2023, the US government announced new export controls on advanced chips. NVIDIA's filing that day listed the RTX 4090, along with the A100, A800, H100, H800, L40 and L40S, among products that would need a license for export to China and to Country Groups D1, D4 and D5, with the rule taking effect 30 days after publication ^[23]. A later NVIDIA filing on the rules did not mention the RTX 4090, which left its status unclear until Tom's Hardware confirmed that exports of the card to China would stop on November 17, 2023 ^[7].

The card was caught because its total processing performance (TPP), the rules' measure of compute, exceeded the 4,800 threshold: Tom's Hardware put the RTX 4090's TPP at about 5,280, based on its 660 TFLOPS of FP8 throughput ^[24]. Because it is not designed for data centers, it fell under "Notified Advanced Computing," a new license exception for non-data-center chips, rather than the regular license regime ^[25]. According to Georgetown University's Center for Security and Emerging Technology, exporting an RTX 4090 to China required an "expedited license" whose review standard was unknown, while shipping one to a subsidiary of a Chinese company such as Baidu in Vietnam needed no license at all ^[25]. Tom's Hardware reported that license applications were reviewed with a presumption of denial ^[24]. By November 19, 2023, NVIDIA and its partners had stopped selling the card in China, and NVIDIA had removed it from its Chinese website ^[24]. Tom's Hardware noted that the rules would also block assembly of RTX 4090 cards in China, which could force board partners to move that work to places such as Taiwan ^[7].

RTX 4090 D

NVIDIA introduced a China-only variant, the GeForce RTX 4090 D, on December 28, 2023. It has 114 SMs (14,592 CUDA cores, 12.8% fewer) and a 425 W power rating instead of 450 W, but keeps 24 GB of GDDR6X on a 384-bit bus and the 2.52 GHz boost clock; it was priced at 12,999 yuan (about $1,828), the same as the RTX 4090 ^[8]. The Financial Times reported in January 2024 that the cuts could leave the RTX 4090 D too slow for large language model training ^[26].

Gray market and conversions

The ban created a gray market. In November 2023 Tom's Hardware reported on Chinese factories that stripped the AD102 chip and memory from RTX 4090 gaming cards and mounted them on boards with blower-style coolers, which fit more densely into AI servers ^[27]. The Financial Times reported in January 2024 that one factory manager had disassembled more than 4,000 cards in December 2023, up from about 1,000 in November, and that the main buyers were public enterprises and small AI labs ^[26]. In February 2024 Nikkei reported that buyers at Taipei's Guanghua market and Seoul's Yongsan market were paying more than $2,500 per card, with prices up as much as 60% in some Asian markets, to resell the cards in China and Hong Kong ^[28]. Memory upgrades followed: in April 2025 DigiTimes Asia reported that some Chinese AI data centers were refurbishing and reselling 48 GB versions of the China-market RTX 4090 D, at 20,000 to 40,000 yuan (about $2,735 to $5,470) each ^[29].

Market and pricing

Retail prices

Digital Trends wrote in November 2024 that the RTX 4090 had sold above its list price for most of its time on the market ^[9]. US prices climbed in late 2023 as the China ban took effect, and again in late 2024 as production wound down ^[30][9].

End of production

NVIDIA has not announced an end-of-life date for the RTX 4090, and its product page was still online in October 2026 ^[13]. In November 2024 Digital Trends, citing the supply-chain outlet Board Channels, reported that the RTX 4090 was among the first RTX 40 series cards to go out of production, while noting that NVIDIA had not officially announced the end of the series ^[9]. NVIDIA unveiled the RTX 5090 on January 6, 2025 and put it on sale on January 30, 2025 at $1,999 ^[10].

Data-center use and the GeForce license

GPU rental marketplaces and smaller clouds sell RTX 4090 time as AI compute, listing the card next to data-center GPUs. Vast.ai had a rental page for the card by August 2024 ^[32], RunPod offered it in both its Secure Cloud and lower-priced Community Cloud tiers ^[6], and SaladCloud said in 2025 that it rented "unused, latent compute resources from gamers and e-sports arenas worldwide," with more than 60,000 consumer GPUs on its network ^[33].

NVIDIA's driver license limits this kind of use. The version on NVIDIA's US website in October 2026 says GeForce or Titan software "is licensed for use only on GeForce or Titan hardware products you own" and "is not licensed for datacenter deployment," and separately bars users from providing "commercial hosting services with the SOFTWARE" ^[34]. The restriction is older than the card: The Register reported on January 3, 2018 that NVIDIA had recently added a clause reading "No Datacenter Deployment. The SOFTWARE is not licensed for datacenter deployment, except that blockchain processing in a datacenter is permitted." ^[35] An NVIDIA spokesperson told The Register in January 2018 that GeForce and Titan GPUs "were never designed for data center deployments," that the company did not intend to prohibit research uses "that do not operate at data center scale," and that "there are many different types of data centers" ^[35]. The license covers NVIDIA's driver software rather than the hardware itself ^[35]. For a data center that wants AD102 silicon under NVIDIA's data-center terms, the same chip is sold as the L40 and L40S ^[7].

Rental prices

The following list prices were captured from providers' pricing pages on the dates shown. Prices across providers are compared in cloud GPU pricing comparison.

Reception

Tom's Hardware gave the RTX 4090 its Editor's Choice award in its October 2022 review, which said the card "delivers on the technological and performance fronts, easily besting previous generation offerings," while listing its price and power draw among the drawbacks ^[38].

Within weeks of launch, owners reported melted 16-pin 12VHPWR power connectors, both on the adapters that shipped with cards and on the cards' sockets ^[39]. In November 2022 NVIDIA said it was "aware of about 50 cases globally" and that "a common issue is that connectors are not fully plugged into the graphics card"; it recommended attaching the adapter to the card before installing the card, and said it and its partners would expedite returns "regardless of the cable or card used" ^[39]. In July 2023 Tom's Hardware reported that PCI-SIG was revising the standard. The replacement 12V-2x6 connector moves the tips of the sense pins farther from the edge of the socket, from 0.45 mm to 1.7 mm, supplies no power when the sense contacts are open, and remains backward compatible with 12VHPWR ^[40]. Tom's Hardware commented that the revision implied not every failure had been user error, as NVIDIA had suggested ^[40].

AI users judged the card largely on cost. Lambda's benchmarks emphasized its training throughput per dollar ^[3], and Answer.AI designed its fine-tuning system around 24 GB gaming cards because a 48 GB card was "very expensive" ^[4]. Tom's Hardware pointed out the trade-off in February 2023: the professional RTX 6000 Ada Generation, built on the same AD102 chip and priced at about $6,800, supports the peer-to-peer transfers that the $1,599 RTX 4090 lacks ^[12].

References

1. ^1 ^2 ^3 ^4 ^5 ^6"NVIDIA Delivers Quantum Leap in Performance, Introduces New Era of Neural Rendering With GeForce RTX 40 Series." NVIDIA Newsroom, 2022-09-20. nvidianews.nvidia.com/...ith-geforce-rtx-40-series
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12"NVIDIA Ada GPU Architecture" (whitepaper, version 2.02). NVIDIA, 2023-11-03. images.nvidia.com/...nvidia-ada-gpu-architecture.pdf
3. ^1 ^2 ^3 ^4"NVIDIA GeForce RTX 4090 vs RTX 3090 Deep Learning Benchmark." Lambda, 2022-10-31. lambda.ai/...0-vs-rtx-3090-deep-learning-benchmark
4. ^1 ^2 ^3 ^4"You can now train a 70b language model at home." Answer.AI, 2024-03-06. answer.ai/...2024-03-06-fsdp-qlora
5. ^1 ^2 ^3 ^4"PowerInfer: Fast Large Language Model Serving with a Consumer-grade GPU" (Song, Mi, Xie and Chen). arXiv, 2023-12-16. arxiv.org/...2312.12456
6. ^1 ^2 ^3"Pricing for GPU Instances, Storage, and Serverless" (RunPod pricing page, archived snapshot). RunPod via Internet Archive, 2024-07-27. web.archive.org/...pricing
7. ^1 ^2 ^3 ^4 ^5"Nvidia RTX 4090 Subject to China Export Restrictions Starting November 17." Tom's Hardware, 2023-11-02. tomshardware.com/...trictions-starting-november-17
8. ^1 ^2"Nvidia launches China-specific RTX 4090D Dragon GPU, sanctions-compliant model has fewer cores and lower power draw." Tom's Hardware, 2023-12-28. tomshardware.com/...wer-cores-and-lower-power-draw
9. ^1 ^2 ^3 ^4 ^5"Nvidia's RTX 40-series is coming to an end." Digital Trends (via Yahoo Tech), 2024-11-12. tech.yahoo.com/...a-rtx-40-series-coming-135841361
10. ^1 ^2"NVIDIA Blackwell GeForce RTX 50 Series Opens New World of AI Computer Graphics." NVIDIA Newsroom, 2025-01-06. nvidianews.nvidia.com/...d-of-ai-computer-graphics
11. ^1 ^2"Nvidia kills off NVLink on RTX 4090." Windows Central, 2022-09-21. windowscentral.com/...kills-off-nvlink-on-rtx-4090
12. ^1 ^2 ^3"Nvidia Confirms GeForce Cards Lack P2P, Pushing Expensive Pro Cards." Tom's Hardware, 2023-02-22. tomshardware.com/...geforce-cards-lack-p2p-support
13. ^1 ^2 ^3"GeForce RTX 4090 Graphics Card" (product page and full specifications). NVIDIA, accessed 2026-10-15. nvidia.com/...rtx-4090
14. ^1 ^2"NVIDIA RTX Blackwell GPU Architecture" (whitepaper, version 1.1). NVIDIA, 2025-05-12. images.nvidia.com/...lackwell-gpu-architecture.pdf
15. ^1 ^2 ^3"NVIDIA H100 Tensor Core GPU Architecture" (whitepaper, version 1.04). NVIDIA, 2023-05-04. dam-cdn.nvd.orangelogic.com/...0117r17n8042d73.pdf
16. ^1 ^2"OpenAI's New Open Models Accelerated Locally on NVIDIA GeForce RTX and RTX PRO GPUs." NVIDIA Blog, 2025-08-05. blogs.nvidia.com/...rtx-ai-garage-openai-oss
17. ^"Hardware support: Nvidia." Ollama documentation, accessed 2026-10-15. docs.ollama.com/gpu
18. ^"FP8 W8A8." vLLM documentation, accessed 2026-10-15. docs.vllm.ai/...fp8
19. ^"QLoRA: Efficient Finetuning of Quantized LLMs" (Dettmers, Pagnoni, Holtzman and Zettlemoyer). arXiv, 2023-05-23. arxiv.org/...2305.14314
20. ^"TinyBox AI accelerator now available starting at $15k, available in AMD 7900XTX and Nvidia RTX 4090 variants." Tom's Hardware, 2024-08-27. tomshardware.com/...-7900xtx-and-rtx-4090-variants
21. ^1 ^2"Stable Diffusion Benchmarks: 45 Nvidia, AMD, and Intel GPUs Compared." Tom's Hardware, 2023-12-15. tomshardware.com/...stable-diffusion-benchmarks
22. ^1 ^2"Llama.cpp AI Performance With The GeForce RTX 5090 Review." Phoronix, 2025-01-27. phoronix.com/...nvidia-rtx5090-llama-cpp
23. ^"NVIDIA Corporation Form 8-K" (current report on the October 2023 export rules). U.S. Securities and Exchange Commission, 2023-10-17. sec.gov/...nvda-20231017
24. ^1 ^2 ^3"The US government banned Nvidia's fastest gaming GPU from China: chipmaker pulls RTX 4090 listings due to AI concerns, but leaves RTX 6000 Ada." Tom's Hardware, 2023-11-19. tomshardware.com/...s-in-china-but-rtx6000-remains
25. ^1 ^2"Explainer: The Commerce Department's October 2023 Export Control Update." Center for Security and Emerging Technology, 2023-12-04. cset.georgetown.edu/...bis-2023-update-explainer
26. ^1 ^2"Single Chinese factory reportedly repurposed over 4,000 Nvidia RTX gaming cards into AI accelerators in December." Tom's Hardware, 2024-01-12. tomshardware.com/...to-ai-accelerators-in-december
27. ^"Sidestepping GPU ban, Chinese factories dismantle and transform Nvidia RTX 4090 gaming cards into AI accelerators." Tom's Hardware, 2023-11-24. tomshardware.com/...-blowers-to-old-rtx-4090-cards
28. ^1 ^2"Nvidia's top gaming GPU prices are out of control in Asia: RTX 4090 price hikes up to 60%." Tom's Hardware, 2024-02-03. tomshardware.com/...ces-are-out-of-control-in-asia
29. ^"Chinese data centers refurbing and selling Nvidia RTX 4090D GPUs due to overcapacity: 48GB models sell for up to $5,500." Tom's Hardware, 2025-04-23. tomshardware.com/...models-sell-for-up-to-usd5-500
30. ^1 ^2"It's Cyber Monday, but prices of Nvidia RTX 4090 GPUs are skyrocketing: what gives?" TechRadar, 2023-11-27. techradar.com/...-gpus-are-skyrocketing-what-gives
31. ^"These are the best GPU 'deals' based on real-world scalper pricing and our FPS per dollar test results." Tom's Hardware, 2025-03-25. tomshardware.com/...ices-cost-up-to-twice-the-msrp
32. ^"RTX 4090 GPU Rental" (Vast.ai pricing page, archived snapshot). Vast.ai via Internet Archive, 2024-08-12. web.archive.org/...RTX-4090
33. ^1 ^2"Salad GPU Cloud Pricing" (SaladCloud pricing page, archived snapshot). SaladCloud via Internet Archive, 2025-06-10. web.archive.org/...pricing
34. ^"NVIDIA Driver License Agreement" (License for Customer Use of GeForce Software). NVIDIA, accessed 2026-10-15. nvidia.com/...geforce-license
35. ^1 ^2 ^3"Nvidia: Using cheap GeForce, Titan GPUs in servers? Haha, nope!" The Register, 2018-01-03. theregister.com/...nvidia_server_gpus
36. ^"Pricing" (TensorDock pricing page, archived snapshot). TensorDock via Internet Archive, 2023-12-05. web.archive.org/...pricing
37. ^"GPU Cloud Pricing" (Runpod pricing page, archived snapshot). Runpod via Internet Archive, 2026-10-12. web.archive.org/...pricing
38. ^"Nvidia GeForce RTX 4090 Review: Queen of the Castle." Tom's Hardware, 2022-10-11. tomshardware.com/...nvidia-geforce-rtx-4090-review
39. ^1 ^2"Nvidia's Melting 16-Pin Connectors on RTX 4090: All We Know." Tom's Hardware, 2022-11-16. tomshardware.com/...pu-power-connector-all-we-know
40. ^1 ^2"16-Pin Power Connector Gets A Much-Needed Revision, Meet The New 12V-2x6 Connector." Tom's Hardware, 2023-07-03. tomshardware.com/...meet-the-new-12v-2x6-connector