24GB · AI Score 10.4/100 · first-party measured on 12 AI workloads
10.6 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 0.6B | 770.13 tok/s | 86 W36°CQ4_K_M | ✓ Measured |
| Llama 3.2 1B | 752.52 tok/s | 95 W42°CQ4_K_M | ✓ Measured |
| MiniCPM5 2B | 372.72 tok/s | 105 W56°CQ4_K_M | ✓ Measured |
| LFM2.5 2.6B | 385.05 tok/s | 142 W57°CQ4_K_M | ✓ Measured |
| Granite 4.1 3B | 276.62 tok/s | 143 W60°CQ4_K_M | ✓ Measured |
| Agents-A1-4B | 219.4 tok/s | 160 W33°CQ4_K_M | ✓ Measured |
| Nemotron 3 Nano 4B | 258.12 tok/s | 166 W56°CQ4_K_M | ✓ Measured |
| Qwen3 4B | 260.54 tok/s | 3.1 GB peak116 W38°C2.24 tok/WQ4_K_M | ✓ Measured |
| Spark-X2.5-4B | 245.97 tok/s | 166 W33°CQ4_K_M | ✓ Measured |
| DeepSeek Coder 7B Instruct v1.5 | 188.59 tok/s | 179 W63°CQ4_K_M | ✓ Measured |
| OLMo 3 7B Instruct | 170.75 tok/s | 186 W34°CQ4_K_M | ✓ Measured |
| OLMo 3 7B Think | 170.69 tok/s | 183 W34°CQ4_K_M | ✓ Measured |
| Qwen2-7B-Instruct | 177.4 tok/s | 189 W35°CQ4_K_M | ✓ Measured |
| Qwen2.5-7B | 183.66 tok/s | 221 W47°CQ4_K_M | ✓ Measured |
| Qwen2.5-Coder 7B | 183.73 tok/s | 221 W49°CQ4_K_M | ✓ Measured |
| Apertus-8B-Instruct | 163.09 tok/s | 191 W36°CQ4_K_M | ✓ Measured |
| Llama 3 8B | 167.69 tok/s | 199 W54°CQ4_K_M | ✓ Measured |
| Llama 3.1 8B | 171.29 tok/s | 5.1 GB peak191 W42°C0.9 tok/WQ4_K_M | ✓ Measured |
| Qwen3 8B | 164.32 tok/s | 214 W58°CQ4_K_M | ✓ Measured |
| Nemotron Nano 9B v2 | 121.89 tok/s | 206 W55°CQ4_K_M | ✓ Measured |
| Ornith 1.5 9B | 144.04 tok/s | 205 W57°CQ4_K_M | ✓ Measured |
| Gemma 4 12B | 103.57 tok/s | 219 W59°CQ4_K_M | ✓ Measured |
| Qwen2.5-Coder 14B | 95.12 tok/s | 8.5 GB peak174 W46°C0.55 tok/WQ4_K_M | ✓ Measured |
| Qwen3 14B | 96.37 tok/s | 238 W60°CQ4_K_M | ✓ Measured |
| gpt-oss-20b | 286.23 tok/s | 154 W47°CQ4_K_M | ✓ Measured |
| Gemma 4 26B A4B | 188.23 tok/s | 139 W55°CQ4_K_M | ✓ Measured |
| Qwen3.6 27B | 49.02 tok/s | 261 W65°CQ4_K_M | ✓ Measured |
| Qwen3.8 27B | 48.01 tok/s | 259 W63°CQ4_K_M | ✓ Measured |
| Qwen3 30B A3B | 259.61 tok/s | 135 W56°CQ4_K_M | ✓ Measured |
| Qwen3 30B A3B Instruct 2507 | 263.17 tok/s | 128 W59°CQ4_K_M | ✓ Measured |
| Qwen3-Coder 30B A3B | 271.04 tok/s | 146 W45°CQ4_K_M | ✓ Measured |
| Gemma 4 31B | 45.08 tok/s | 266 W65°CQ4_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 1.5 | 80.14 images/min | 325 W53°C | ✓ Measured |
| SD Turbo | 786.15 images/min | 91 W50°C | ✓ Measured |
| Stable Diffusion 2.1 | 60.58 images/min | 275 W38°C | ✓ Measured |
| LCM DreamShaper v7 | 260.94 images/min | 178 W60°C | ✓ Measured |
| SDXL Turbo | 652.11 images/min | 90 W34°C | ✓ Measured |
| SSD-1B | 15.77 images/min | 414 W68°C | ✓ Measured |
| SDXL-Lightning | 86.71 images/min | 271 W49°C | ✓ Measured |
| Z-Image Turbo | 7.24 images/min | 414 W78°C | ✓ Measured |
| Sana 1.6B | 41.47 images/min | 411 W59°C | ✓ Measured |
| Stable Diffusion XL | 16.28 images/min | 14.8 GB peak421 W53°C3.7 s/img | ✓ Measured |
| DreamShaper XL Lightning | 91.52 images/min | 306 W71°C | ✓ Measured |
| DreamShaper XL Turbo | 54.35 images/min | 412 W74°C | ✓ Measured |
| Playground v2.5 | 9.96 images/min | 396 W52°C | ✓ Measured |
| PixArt-Sigma XL | 24.27 images/min | 415 W60°C | ✓ Measured |
| Stable Diffusion 3 Medium | 13.95 images/min | 437 W73°C | ✓ Measured |
| FLUX.2 klein 4B | 42.38 images/min | 392 W50°C | ✓ Measured |
| Kolors | 9.88 images/min | 417 W70°C | ✓ Measured |
| Z-Image | 1.17 images/min | 398 W59°C | ✓ Measured |
| AuraFlow v0.3 | 3.02 images/min | 423 W61°C | ✓ 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 |
|---|---|---|---|
| Stable Video Diffusion | 2.24 clips/min | 403 W74°C | ✓ Measured |
| LTX-Video (image to video) | 4.19 clips/min | 384 W73°C | ✓ Measured |
| Wan 2.2 TI2V-5B (image to video) | 1.12 clips/min | ✓ Measured CPU offload | |
| Cosmos-Predict2 2B Video2World | 0.06 clips/min | 424 W66°C | ✓ Measured CPU offload |
| Stable Video Diffusion XT | 1.21 clips/min | 347 W68°C | ✓ Measured 2 hosts ±2% · CPU offload |
| CogVideoX-5B I2V | 0.33 clips/min | 344 W70°C | ✓ Measured 2 hosts ±1% · CPU offload |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| Wan 2.1 1.3B | 0.67 frames/s | 411 W73°C72.9 s/clip | ✓ Measured |
| AnimateDiff-Lightning | 10.63 frames/s | 389 W56°C | ✓ Measured |
| CogVideoX-2B | 0.52 frames/s | 410 W53°C96.5 s/clip | ✓ Measured 2 hosts ±2% |
| CogVideoX-5B | 0.17 frames/s | 423 W85°C287.9 s/clip | ✓ Measured CPU offload |
| LTX-Video (distilled) | 6.7 frames/s | 227 W45°C14.5 s/clip | ✓ Measured 3 hosts ±25% |
| Wan 2.2 5B (720p) | 0.43 frames/s | 18.6 GB peak366 W68°C114.5 s/clip | ✓ Measured CPU offload |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| Shap-E (image to 3D) | 1309.09 assets/hour | 362 W47°C | ✓ Measured |
| Stable Fast 3D | 9113.92 assets/hour | 157 W37°C | ✓ Measured |
| Hunyuan3D 2mini | 19.81 assets/hour | 71 W50°C | ✓ Measured |
| TRELLIS.2 Image-to-3D | 57.2 assets/hour | ✓ Measured | |
| Hunyuan3D 2.0 | 12.27 assets/hour | 68 W54°C | ✓ 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, #28 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 | 225% | 23.8 | |
| NVIDIA GeForce RTX 5090 | 210% | 22.3 | |
| NVIDIA RTX 5880 Ada Generation | 139% | 14.7 | |
| NVIDIA RTX 5000 Ada Generation | 104% | 11 | |
| NVIDIA GeForce RTX 4090 | 100% | 10.6 | |
| NVIDIA GeForce RTX 3090 Ti | 80% | 8.5 | |
| NVIDIA Titan RTX | 77% | 8.2 | |
| NVIDIA GeForce RTX 3090 | 73% | 7.7 | |
| GeForce RTX 5080 | 49% | 5.2 |
Same card, other workloads: NVIDIA GeForce RTX 4090 Gaming benchmarks
← All AI & Machine Learning GPU rankings
| Transistors | 76,300 million |
| Die size | 608.4 mm² |
| Process node | 4 nm |
| Fabricated by | TSMC |
| Transistor density | 125.4 million per mm² |
Denser than 97% of the 746 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 | 3.8 min | 24.33 Wh | all 2 stages measured |
| 60-second AI short film | 5.8 min | 21.08 Wh | all 3 stages measured |
| Full codebase review | 10.5 min | 30.47 Wh | measured |
| Animate a batch of images | 17.8 min | n/a | measured |
| Short social clips | 21.5 min | 125.78 Wh | all 3 stages measured |
Can't run: Product photo shoot (needs FLUX.1 Kontext dev), Photo restoration batch (needs FLUX.1 Kontext dev), Restore and enlarge photos (needs FLUX.1 Kontext dev), Character sheet, 12 poses (needs FLUX.1 dev), 6-panel comic page (needs FLUX.1 dev), Long-form article batch (needs Llama 3.3 70B).
This card is $1,599 to buy. The cheapest listed rate on RunPod is $0.340/hour, but that is the floor: we budget $0.408/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 3,919 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 | $298 | 5.4 years |
| 8 hours a day, working on it | 2,920 | $1,191 | 1.3 years |
| 24/7, always-on agent | 8,760 | $3,574 | 5.4 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.