The NVIDIA RTX 4090 is still the card everyone measures against. Nearly four years after launch, it remains one of the fastest consumer GPUs you can put in a workstation, and the AI boom has kept demand high long past the point where most flagships fade. This guide breaks down the full rtx 4090 picture, specs, VRAM, real 2026 pricing, and AI performance, and covers something most guides skip entirely: the 48GB modded blower edition that doubles the memory of the standard 24GB card.
Whether you are weighing a 4090 graphics card for gaming, local AI, or a dense multi-GPU rig, the sections below give you the numbers you need and the buying context that turns those numbers into a decision.
Whether you are weighing a 4090 graphics card for gaming, local AI, or a dense multi-GPU rig, the sections below give you the numbers you need and the buying context that turns those numbers into a decision.
Quick answer:
The RTX 4090 is NVIDIA's Ada Lovelace flagship: 16,384 CUDA cores, 24GB GDDR6X, 82.6 TFLOPS FP32, and a 450W TDP. It launched at $1,599 in October 2022 and sells for roughly $1,800-$3,500 in 2026. A 48GB modded version exists for AI workloads that need more memory. This is educational content, not financial advice.
What is the NVIDIA RTX 4090?
The nvidia geforce rtx 4090 is the top card in NVIDIA's 40-series, built on the Ada Lovelace architecture and the AD102 die. It packs 76 billion transistors on TSMC's 4N process, making it the most powerful consumer GPU NVIDIA produced before the RTX 5090. When people search for 4090 or nvidia 4090, this is the card they mean, a triple-slot Founders Edition and a wide range of partner designs from ASUS, MSI, Gigabyte, and others.
What made the 4090 gpu a landmark release was the jump in both raw rasterization and AI throughput: NVIDIA claimed 2-4x the performance of the RTX 3090 Ti, driven by third-gen RT cores, fourth-gen Tensor cores, and DLSS 3 with frame generation. Those same Tensor cores are why the card became a default choice for local AI long after its gaming headlines faded.
It is worth being precise about the name, because search results are full of cards that do not exist. There is no 4090 Ti, no 4090 Super, and no GTX 4090, whatever listings might suggest. The GeForce RTX 4090 Laptop GPU is a different chip entirely, closer to a desktop 4080. When this guide refers to the 4090, it means the desktop card built on the AD102 die, in either its 24GB or 48GB form.
What made the 4090 gpu a landmark release was the jump in both raw rasterization and AI throughput: NVIDIA claimed 2-4x the performance of the RTX 3090 Ti, driven by third-gen RT cores, fourth-gen Tensor cores, and DLSS 3 with frame generation. Those same Tensor cores are why the card became a default choice for local AI long after its gaming headlines faded.
It is worth being precise about the name, because search results are full of cards that do not exist. There is no 4090 Ti, no 4090 Super, and no GTX 4090, whatever listings might suggest. The GeForce RTX 4090 Laptop GPU is a different chip entirely, closer to a desktop 4080. When this guide refers to the 4090, it means the desktop card built on the AD102 die, in either its 24GB or 48GB form.
The two variants: 24GB standard vs 48GB modded
Here is where MillionMiner's coverage goes a step further than the typical spec sheet. The stock rtx 4090 ships with 24GB of GDDR6X. But a rtx 4090 48gb variant exists, a modded blower card that doubles the memory to 48GB on the same AD102 silicon. It is not an NVIDIA retail product; it is a specialist build aimed squarely at AI users who hit the 24GB ceiling. We sell both the 24GB blower and the 48GB blower, and the difference matters more than most buyers expect.
RTX 4090 specs: the full breakdown
The rtx 4090 specs read like a data-center card that wandered into the consumer aisle. Below is the specs-at-a-glance view, followed by the details that actually affect your build.
Cores, clocks, and compute
The 4090 cuda cores count is 16,384, paired with 512 fourth-gen Tensor cores and 128 third-gen RT cores. The base clock sits at 2.23 GHz with a boost of roughly 2.52 GHz. That yields 82.6 TFLOPS of FP32, the 4090 teraflops figure quoted across every benchmark, and around 1,321 AI TOPS across FP8, FP16, BF16, TF32, and INT8 formats. For anyone tracking nvfp4 in rtx 4090 discussions: that is a Blackwell-era format, so the Ada-based 4090 handles FP8 natively but not the newer NVFP4 path found on the 50-series.
Those core counts translate into real gaming and compute headroom. At 4K with the highest settings, the 4090 graphics card still posts frame rates no other consumer card matched at launch, and DLSS 3 frame generation stretches that lead further in supported titles. For creators, the dual AV1 encoders enable real-time 8K HDR capture and hardware-accelerated video export, a genuine differentiator over AMD flagship cards at the time.
Those core counts translate into real gaming and compute headroom. At 4K with the highest settings, the 4090 graphics card still posts frame rates no other consumer card matched at launch, and DLSS 3 frame generation stretches that lead further in supported titles. For creators, the dual AV1 encoders enable real-time 8K HDR capture and hardware-accelerated video export, a genuine differentiator over AMD flagship cards at the time.
Memory and bandwidth
Stock 4090 vram is 24GB of GDDR6X on a 384-bit bus. The rtx 4090 memory bandwidth comes in at roughly 1,008 GB/s, close to 1 TB/s, which is a major reason the card stays relevant for AI even after the 5090 launch. When buyers ask how much vram does a 4090 have, the answer is 24GB on the standard card and 48GB on the modded blower. That memory headroom, not clock speed, is what decides which AI models you can actually run.
Power, thermals, and dimensions
The 4090 tdp is 450W, and real rtx 4090 power consumption under sustained load lands close to that ceiling. NVIDIA recommends at least an 850W PSU, though most builders pairing it with a high-end CPU aim for 1,000W or more to cover transient spikes. The card draws power through a single 16-pin 12VHPWR connector. On 4090 wattage and 4090 dimensions: the Founders Edition measures 304 × 137 × 61 mm as a triple-slot design, while the blower variants we stock use a 2-slot form factor that fits dense multi-GPU racks where the FE simply will not.
RTX 4090 price in 2026: what you actually pay
The rtx 4090 price is one of the strangest stories in recent GPU history. This is a card that costs more today than it did at launch, unusual for any electronics product nearly four years old.
The 4090 msrp was $1,599 when it launched on the 4090 release date of October 12, 2022, alongside the limited 4090 founders edition. In 2026, though, the 4090 graphics card typically sells for roughly $1,800 to $3,500 depending on model and retailer, with street prices for new units often landing in the $2,400-$2,800 band. Used cards trade lower, but professional demand keeps a firm floor under them.
The 4090 msrp was $1,599 when it launched on the 4090 release date of October 12, 2022, alongside the limited 4090 founders edition. In 2026, though, the 4090 graphics card typically sells for roughly $1,800 to $3,500 depending on model and retailer, with street prices for new units often landing in the $2,400-$2,800 band. Used cards trade lower, but professional demand keeps a firm floor under them.
Why is the RTX 4090 so expensive?
Three forces keep the 4090 price high. First, NVIDIA ended production in late 2024, so new-in-box units come only from remaining channel inventory. Second, sustained demand from AI developers, for whom 24GB of VRAM at this price has no direct consumer-GPU rival, keeps the secondary market unusually firm. Third, thin retail supply means even small stock swings move prices. When shoppers ask how much is a 4090 or look for a 4090 for sale, the honest answer is that pricing depends heavily on timing, model tier, and whether the unit is new or used.
Regional differences are stark too. Export controls and import duties push the 4090 graphics card well above US pricing in some markets, while local availability can pull it lower in others. The card's 4090 gpu price also splits sharply between Founders Edition units, which sit near the bottom of the range where you can find them, and factory-overclocked partner cards that command a premium for cooling headroom and resale liquidity rather than meaningful performance gains.
If you need the card for AI rather than gaming, it is worth comparing the purchase against renting. Our cloud GPU pricing guide walks through the rent-versus-buy math, and for teams running sustained workloads, GPU hosting often beats both.
Regional differences are stark too. Export controls and import duties push the 4090 graphics card well above US pricing in some markets, while local availability can pull it lower in others. The card's 4090 gpu price also splits sharply between Founders Edition units, which sit near the bottom of the range where you can find them, and factory-overclocked partner cards that command a premium for cooling headroom and resale liquidity rather than meaningful performance gains.
If you need the card for AI rather than gaming, it is worth comparing the purchase against renting. Our cloud GPU pricing guide walks through the rent-versus-buy math, and for teams running sustained workloads, GPU hosting often beats both.
RTX 4090 VRAM: why 24GB vs 48GB changes everything
For AI work, VRAM is the single most important number on the spec sheet, more than clocks, more than core count. It sets a hard ceiling on the models and batch sizes you can hold in memory. This is exactly where the rtx 4090 48gb edition earns its place.
What 24GB gets you
Twenty-four gigabytes is genuinely capable. A standard 24GB card runs inference on most open-source LLMs at full precision, 7B and 13B models comfortably, and quantized versions of 70B+ models. It handles high-resolution Stable Diffusion, ComfyUI, and video generation without the out-of-memory errors that cripple smaller cards. For fine-tuning, 24GB covers models up to roughly 7-13B parameters depending on precision.
That covers a huge share of real work: chatbots, AI agents, embeddings, image generation, and rapid prototyping all fit comfortably. The 4090 vram advantage over cheaper cards is what makes this possible, since a 16GB card hits out-of-memory walls on the same tasks. For many builders, 24GB is the sweet spot where the card stops being a bottleneck and starts being a tool.
That covers a huge share of real work: chatbots, AI agents, embeddings, image generation, and rapid prototyping all fit comfortably. The 4090 vram advantage over cheaper cards is what makes this possible, since a 16GB card hits out-of-memory walls on the same tasks. For many builders, 24GB is the sweet spot where the card stops being a bottleneck and starts being a tool.
When you need 48GB
The 24GB ceiling is real, though. Larger context windows, bigger batch sizes, and fine-tuning past ~20B parameters all push against it. This is the loophole in most 4090 guides: they tell you two cards do not merge into 48GB because the 4090 has no NVLink, and that is true. But a 48gb nvidia rtx 4090 gpu sidesteps the problem entirely by putting 48GB on a single card. One board, one PCIe slot, doubled memory. For 30B+ models, denser multi-GPU rigs, and workloads where the rtx 4090 48gb price premium pays for itself in fewer cards, it is the more practical build.
RTX 4090 for AI: real-world performance
The rtx 4090 for ai story is why this card refuses to age. Its 82.6 TFLOPS of FP32 and fourth-gen Tensor cores make it one of the most capable single GPUs for a wide range of AI workloads, at a fraction of an H100's cost.
On rtx 4090 ai performance: the card handles inference for most open-source models, LLaMA 3, Mistral, Qwen, and similar, and excels at diffusion models and rapid experimentation where iteration speed matters more than maximum VRAM. Its INT8/FP8 throughput of around 1,321 TOPS is genuinely competitive with older data-center silicon. The main constraint remains memory: 24GB on the standard card, which is exactly the gap the 48GB variant closes.
If you are architecting a rig around these cards, our guides on choosing a GPU for AI and AI server design cover the fundamentals. For scaling past a single card, our walkthrough on multi-GPU AI setups covers the build side in depth, including power, cooling, and PCIe layout for dense 4090 rigs.
One reason the 4090 gpu holds its value for AI is simple economics: a single card delivers roughly 70-90% of an A100's performance in many machine-learning tasks at a fraction of the cost. That price-to-performance ratio is what pulled the 4090 out of the gaming aisle and into serious compute rigs. If you have ever wondered why AI hardware costs so much, the 4090 is the clearest case study, a consumer card priced like a professional one because the market treats it as such.
For the rtx 4090 48gb price premium to make sense, your workload has to genuinely need the extra memory. But when it does, the 48GB card replaces what would otherwise be two or three standard cards, cutting power draw, PCIe lanes, and cooling complexity in the process. That is the practical calculus behind the modded blower, and it is exactly the option most 4090 guides never mention.
On rtx 4090 ai performance: the card handles inference for most open-source models, LLaMA 3, Mistral, Qwen, and similar, and excels at diffusion models and rapid experimentation where iteration speed matters more than maximum VRAM. Its INT8/FP8 throughput of around 1,321 TOPS is genuinely competitive with older data-center silicon. The main constraint remains memory: 24GB on the standard card, which is exactly the gap the 48GB variant closes.
If you are architecting a rig around these cards, our guides on choosing a GPU for AI and AI server design cover the fundamentals. For scaling past a single card, our walkthrough on multi-GPU AI setups covers the build side in depth, including power, cooling, and PCIe layout for dense 4090 rigs.
One reason the 4090 gpu holds its value for AI is simple economics: a single card delivers roughly 70-90% of an A100's performance in many machine-learning tasks at a fraction of the cost. That price-to-performance ratio is what pulled the 4090 out of the gaming aisle and into serious compute rigs. If you have ever wondered why AI hardware costs so much, the 4090 is the clearest case study, a consumer card priced like a professional one because the market treats it as such.
For the rtx 4090 48gb price premium to make sense, your workload has to genuinely need the extra memory. But when it does, the 48GB card replaces what would otherwise be two or three standard cards, cutting power draw, PCIe lanes, and cooling complexity in the process. That is the practical calculus behind the modded blower, and it is exactly the option most 4090 guides never mention.
Ready to Start Mining?
Free worldwide DDP shipping. Professional hosting from $0.055/kWh.
RTX 4090 vs other GPUs: how it stacks up
Buyers rarely consider the rtx 4090 in isolation. The most common cross-shops are against the newer 50-series, the older 3090, AMD's flagship, and professional cards. The table below puts the key rivals side by side.
| GPU | VRAM | CUDA cores | TDP | Launch price | Best for |
|---|---|---|---|---|---|
| RTX 4090 | 24 / 48GB | 16,384 | 450W | $1,599 | 4K gaming + AI |
| RTX 5090 | 32GB | 21,760 | 575W | $1,999 | Newest flagship |
| RTX 5080 | 16GB | 10,752 | 360W | $999 | Gaming value |
| RTX 5070 Ti | 16GB | 8,960 | 300W | $749 | 1440p+ gaming |
| RTX 5070 | 12GB | 6,144 | 250W | $549 | 1440p value |
| RTX 5060 | 8GB | 3,840 | 145W | $299 | Entry 1080p |
| RTX 4080 | 16GB | 9,728 | 320W | $1,199 | Prev-gen 4K |
| RTX 3090 | 24GB | 10,496 | 350W | $1,499 | Budget 24GB AI |
| RX 7900 XTX | 24GB | 6,144 | 355W | $999 | AMD flagship |
| RTX A5000 | 24GB | 8,192 | 230W | $2,250 | Pro workstation |
| RTX 6000 Ada | 48GB | 18,176 | 300W | $6,800 | Pro AI (48GB, ECC) |
Two comparisons dominate buyer questions. The rtx 4090 vs 5090 gap comes down to whether you need 32GB and Blackwell's newer Tensor cores, the 5090 wins on paper, but availability and a higher price complicate the choice. Our full RTX 5090 specs guide covers that card in detail. The rtx 4090 vs 5080 matchup is closer on gaming but the 5080's 16GB VRAM is a real limitation for AI. And rtx 3090 vs 4090 is the budget-AI question: same 24GB, but the 4090 delivers it with far more compute and bandwidth. Cards like the RTX 5070, 5070 Ti, 5060, 4080, A5000, and RX 7900 XTX round out the field, while the RTX 6000 Ada is the closest official 48GB alternative to the modded 4090.
24GB vs 48GB: which RTX 4090 should you buy?
The choice between the two variants comes down to your workload. Here is a straightforward way to decide:
- Gaming or single-model AI dev? Choose the 24GB card. It handles 4K gaming and most inference and fine-tuning under 20B parameters comfortably, at the lower price point.
- Running 30B+ models or large batches? Choose the 48GB card. The doubled VRAM removes the memory ceiling that forces smaller batches or model splitting on the standard card.
- Building a dense multi-GPU rig? The 48GB blower wins. Its 2-slot form factor and higher per-card memory mean fewer cards, less PCIe pressure, and simpler cooling than stacking triple-slot FE units.
- On a tight budget? Start with 24GB. You can always add a second card later, though remember the 4090 has no NVLink, so two cards mean two separate 24GB pools, not one 48GB pool
How to buy an RTX 4090 without getting burned
With production ended and prices high, the 4090 for sale market attracts its share of misrepresented cards. A few checks protect you:
- Verify the silicon. A genuine 4090 reports AD102 as the GPU and 24576 MB (or 49152 MB for the 48GB mod) of dedicated memory in GPU-Z.
- Confirm power compatibility. You need a 16-pin 12VHPWR connector or a proper adapter, and a PSU rated 850W or higher, 1,000W+ for high-end CPU pairings.
- Check the form factor against your case. The triple-slot FE is large; the 2-slot blower fits tighter builds and server chassis.
- Buy from a specialist for AI cards. For the 48GB modded variant especially, source from a vendor that stands behind the hardware rather than a random marketplace listing.
You can browse both variants and the wider lineup on our AI hardware GPU shop, and see how the 4090 performs against other cards on our GPU AI benchmarks page.
RTX 4090 at a glance: key strengths
If you want the short version of why the nvidia rtx 4090 still matters in 2026, it comes down to a handful of durable advantages:
- Class-leading consumer VRAM : 24GB standard, 48GB modded, with no direct consumer rival at the price.
- Strong AI throughput : ~1,321 AI TOPS and native FP8 support for inference and mixed-precision training.
- Proven longevity : continued driver support and firm resale value years after launch.
- Rack-friendly blower option : the 2-slot 48GB card suits dense multi-GPU AI builds the FE cannot.
- Top-tier 4K and 8K gaming : still one of the fastest cards for high-refresh, ray-traced play.
How MillionMiner fits your RTX 4090 build
Most 4090 coverage stops at specs and sends you to a marketplace to fend for yourself. MillionMiner works differently, because we build and supply these cards for real compute workloads rather than one-off gaming sales. That changes what you get on both variants of the rtx 4090.
For buyers who need the standard card, our 24GB RTX 4090 blower ships in a 2-slot form factor built for dense racks, not the oversized triple-slot design that fights your chassis. For teams hitting the memory ceiling, the 48GB modded blower doubles VRAM on a single card, the exact option most vendors do not stock. Both are verified, tested, and backed for sustained AI use.
Beyond the hardware, we cover the full deployment path. If buying outright does not fit your budget, GPU hosting lets you run 4090-class compute without the upfront cost, and our data center colocation service handles power and cooling for larger fleets.You can compare real performance figures across cards on our GPU AI benchmarks before committing to either variant, and browse the wider lineup in the AI hardware GPU shop.The short version: whether you want one 4090 graphics card for a workstation or a rack of them for production inference, we supply the hardware, the hosting, and the benchmarks to back the decision, all in one place.
For buyers who need the standard card, our 24GB RTX 4090 blower ships in a 2-slot form factor built for dense racks, not the oversized triple-slot design that fights your chassis. For teams hitting the memory ceiling, the 48GB modded blower doubles VRAM on a single card, the exact option most vendors do not stock. Both are verified, tested, and backed for sustained AI use.
Beyond the hardware, we cover the full deployment path. If buying outright does not fit your budget, GPU hosting lets you run 4090-class compute without the upfront cost, and our data center colocation service handles power and cooling for larger fleets.You can compare real performance figures across cards on our GPU AI benchmarks before committing to either variant, and browse the wider lineup in the AI hardware GPU shop.The short version: whether you want one 4090 graphics card for a workstation or a rack of them for production inference, we supply the hardware, the hosting, and the benchmarks to back the decision, all in one place.
The bottom line
The nvidia geforce rtx 4090 has aged into something rare: a flagship that stays relevant because of memory and AI capability, not just gaming headlines. The standard 24GB card covers the vast majority of gaming and AI use cases, while the 48GB modded blower solves the one problem that limits the stock card, running larger models on a single board. Whichever fits your workload, both are available now through MillionMiner's AI hardware store, built and backed for serious compute.
Frequently asked questions
How much VRAM does a 4090 have?
The standard rtx 4090 has 24GB of GDDR6X memory. A modded 48GB blower variant doubles that to 48GB on the same GPU, aimed at AI workloads that exceed the 24GB ceiling.
How much is a 4090 in 2026?
New cards typically run $1,800 to $3,500 depending on model and retailer, with street prices often in the $2,400-$2,800 range. The card sells above its $1,599 launch MSRP because production ended in late 2024 and AI demand stays high. Pricing is directional and changes with supply, this is educational, not financial advice.
Is the RTX 4090 good for AI?
Yes. The rtx 4090 for ai remains one of the best single-GPU choices for inference, fine-tuning mid-size models, and diffusion workloads. Its 24GB VRAM handles most open-source LLMs; the 48GB variant extends that to larger models and batches.
When did the 4090 come out?
The 4090 release date was October 12, 2022, at a $1,599 MSRP, launching alongside the NVIDIA Ada Lovelace architecture.
What is the difference between the 24GB and 48GB RTX 4090?
Both use the same AD102 GPU and 16,384 CUDA cores. The 4090 48gb variant is a modded blower card with double the memory, letting a single card run larger models and bigger batches than the 24GB standard version.
Does the RTX 4090 support NVLink?
No. The 4090 has no NVLink, so two cards do not merge into one 48GB memory pool, each keeps its own 24GB. This is exactly why the single-card 48GB modded variant is useful for memory-bound AI work.
Is the RTX 4090 still worth buying over the RTX 5090?
It depends on your needs. The rtx 4090 vs 5090 comparison favors the 5090 on raw specs and 32GB VRAM, but the 4090 often costs less in practice and its 24GB (or 48GB modded) still covers most real workloads. For gaming and mid-size AI, the 4090 remains a strong value.
Educational content only. Hardware pricing and availability change frequently; verify current figures before purchase. This is not financial advic
