24GB · AI Score 10.4/100 · first-party measured on 12 AI workloads
10.4 AI Score ✓ Measured
Every number on this page is first-party: NVIDIA GeForce RTX 4090 was run on our pinned 12-workload AI suite on 2026-07-11, with under 0.5% run-to-run variance. On Llama 3.1 8B (Q4_K_M) NVIDIA GeForce RTX 4090 delivers about 171.29 tokens/sec. Stepping up to Qwen3 32B it holds roughly 44.28 tok/s. Llama 3.3 70B does not fit. It needs roughly 42GB and this card has 24GB. For image generation, SDXL runs at 8.14 it/s, while FLUX.1-dev won't fit at BF16 (needs ~26GB). 4 of the 12 workloads won't fit on 24GB at the tested precision, Llama 3.3 70B, FLUX.1-dev, FLUX.1 Kontext, Qwen-Image-Edit. We publish those as hard gates rather than quietly dropping to a smaller quant. NVIDIA GeForce RTX 4090 isn't a retail purchase for most people. It's rented by the hour. You can run this exact card on RunPod.
Best consumer efficiency I measured, period: 2.24 tokens/watt on Qwen3 4B. It'll take its full 450W on SDXL and stay under 68°C. The one thing people get wrong: 24GB still can't hold Llama 3.3 70B: that model wants ~46GB, and no consumer card changes that. Quick note on the setup: all my AI benchmarking was done on rented cloud GPUs, I used all three of Vast.ai, RunPod and Modal depending on which had the card, and they all have their pros and cons. Same pinned harness on every run, and everything here got double-checked before it went up.
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| Qwen3 4B | 260.54 tok/s | 3.1 GB peak116 W38°C2.24 tok/WQ4_K_M | ✓ Measured |
| Llama 3.1 8B | 171.29 tok/s | 5.1 GB peak191 W42°C0.9 tok/WQ4_K_M | ✓ Measured |
| Qwen2.5-Coder 14B | 95.12 tok/s | 8.5 GB peak174 W46°C0.55 tok/WQ4_K_M | ✓ Measured |
| Qwen3 32B | 44.28 tok/s | 18.9 GB peak164 W47°C0.27 tok/WQ4_K_M | ✓ Measured |
| Llama 3.3 70B | ✕ Won't fit needs ~46 GB | VRAM-gated at this precision | ✓ Measured |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| Stable Diffusion XL | 16.28 images/min | 14.8 GB peak421 W53°C3.7 s/img | ✓ Measured |
| Z-Image Turbo | 7.43 images/min | 23.3 GB peak425 W58°C8.1 s/img | ✓ Measured |
| FLUX.1 dev | ✕ Won't fit needs ~26 GB | VRAM-gated at this precision | ✓ Measured |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| FLUX.1 Kontext dev | ✕ Won't fit needs ~26 GB | VRAM-gated at this precision | ✓ Measured |
| Qwen-Image-Edit | ✕ Won't fit needs ~42 GB | VRAM-gated at this precision | ✓ Measured |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| LTX-Video (distilled) | 4.7 frames/s | 9.4 GB peak269 W55°C20.6 s/clip | ✓ Measured |
| Wan 2.2 5B (720p) | 0.43 frames/s | 18.6 GB peak366 W68°C114.5 s/clip | ✓ Measured |
| Architecture | Ada Lovelace |
| CUDA cores | 16,384 |
| VRAM | 24GB GDDR6X |
| Memory bus | 384-bit |
| Memory bandwidth | 1008 GB/s |
| Boost clock | 2,520 MHz |
| TDP | 450 W |
| Process | 4nm |
| Interface | PCIe 4.0 x16 |
| Release date | 2022-10-12 |
| Launch MSRP | $1,599 |
NVIDIA GeForce RTX 4090 scores 10.4/100, #29 of 102. It ran 8 of 12; 4 exceeded its 24GB. Every figure here is our own measurement.
100% = this card, AI & Machine Learning headline metric (AI Score). #5 of 61 desktop cards in this vertical.
| GPU | Relative | % | AI Score |
|---|---|---|---|
| NVIDIA RTX 6000 Ada Generation | 229% | 23.8 | |
| NVIDIA GeForce RTX 5090 | 214% | 22.3 | |
| NVIDIA RTX 5880 Ada Generation | 157% | 16.3 | |
| NVIDIA RTX 5000 Ada Generation | 106% | 11 | |
| NVIDIA GeForce RTX 4090 | 100% | 10.4 | |
| NVIDIA GeForce RTX 3090 Ti | 82% | 8.5 | |
| NVIDIA Titan RTX | 79% | 8.2 | |
| NVIDIA GeForce RTX 3090 | 74% | 7.7 | |
| AMD Radeon RX 7900 XTX | 64% | 6.7 |
Same card, other workloads: NVIDIA GeForce RTX 4090 Gaming benchmarks
← All AI & Machine Learning GPU rankings
| Transistors | 76,300 million |
| Die size | 608.5 mm² |
| Process node | 4 nm |
| Fabricated by | TSMC |
| Transistor density | 125.4 million per mm² |
Denser than 99% of the 76 cards we have silicon data for. Density is the clearest measure of what a process node bought: a card that gained it without growing the die got its speed from the fab rather than the architecture.
Silicon figures from Wikipedia (CC BY-SA 4.0). Benchmarks on this page are our own. Compare every chip.
Whole-job timings, composed from our measured per-model results on this card.
| Workflow | Time | Energy | Basis |
|---|---|---|---|
| 24-frame storyboard | 5 min | 24.32 Wh | all 2 stages measured |
| 60-second AI short film | 7.9 min | 30.54 Wh | all 3 stages measured |
| Full codebase review | 10.5 min | 30.47 Wh | measured |
| 10 short social clips | 22.7 min | 125.77 Wh | all 3 stages measured |
Can't run: 40-product photo shoot (needs FLUX.1 Kontext dev), 6-panel comic page (needs FLUX.1 dev), 20 long-form articles (needs Llama 3.3 70B), Character sheet, 12 poses (needs FLUX.1 dev), 100-photo restoration batch (needs FLUX.1 Kontext dev), 100-photo restore and enlarge (needs FLUX.1 Kontext dev).
This card is $1,599 to buy. The cheapest listed rate on Vast.ai is $0.155/hour, but that is the floor: we budget $0.186/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 8,597 GPU-hours. Below it you are paying for idle silicon.
| How you would use it | GPU-hours a year | Rental cost a year | Time to break even |
|---|---|---|---|
| 2 hours a day, hobby | 730 | $136 | 11.8 years |
| 8 hours a day, working on it | 2,920 | $543 | 2.9 years |
| 24/7, always-on agent | 8,760 | $1,629 | 11.8 months |
At hobby usage this card is very unlikely to pay for itself before it is superseded. Rent it. Rental figures include a 20% premium over the cheapest listed rate. Ignores electricity, resale and the fact that a rented card can be a newer one tomorrow.
Cheapest of the RunPod and Vast on-demand rates we see, sampled daily. Spot and interruptible pricing runs lower.