NVIDIA GeForce RTX 4090 vs NVIDIA GeForce RTX 5090, AI & Machine Learning Comparison
NVIDIA GeForce RTX 5090 wins 59 of 61 benchmarks, averaging 43.9% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Wan 2.2 TI2V-5B (image to video), where NVIDIA GeForce RTX 5090 leads by 104% (1.122 vs 2.284 clips/min); the closest fight is Stable Diffusion 2.1 (8% apart); VRAM decides part of this one: NVIDIA GeForce RTX 5090 runs 69 of our 12 AI workloads while the other card runs 63, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA GeForce RTX 4090 | NVIDIA GeForce RTX 5090 | Difference |
|---|---|---|---|
| Qwen3 0.6B tok/s | 770.13 | 868.61 | -11% |
| Llama 3.2 1B tok/s | 752.52 | 1060.14 | -29% |
| MiniCPM5 2B tok/s | 372.72 | 507.31 | -27% |
| LFM2.5 2.6B tok/s | 385.05 | 566.57 | -32% |
| Granite 4.1 3B tok/s | 276.62 | 375.37 | -26% |
| Agents-A1-4B tok/s | 219.4 | 328.96 | -33% |
| Nemotron 3 Nano 4B tok/s | 258.12 | 410.94 | -37% |
| Qwen3 4B tok/s | 260.54 | 375.6 | -31% |
| Spark-X2.5-4B tok/s | 245.97 | 352.29 | -30% |
| DeepSeek Coder 7B Instruct v1.5 tok/s | 188.59 | 291.38 | -35% |
| OLMo 3 7B Instruct tok/s | 170.75 | 266.68 | -36% |
| OLMo 3 7B Think tok/s | 170.69 | 266.72 | -36% |
| Qwen2-7B-Instruct tok/s | 177.4 | 277.64 | -36% |
| Qwen2.5-7B tok/s | 183.66 | 284.6 | -35% |
| Qwen2.5-Coder 7B tok/s | 183.73 | 284.59 | -35% |
| Apertus-8B-Instruct tok/s | 163.09 | 257.69 | -37% |
| Llama 3 8B tok/s | 167.69 | 259.18 | -35% |
| Llama 3.1 8B tok/s | 171.29 | 268.14 | -36% |
| Qwen3 8B tok/s | 164.32 | 243.91 | -33% |
| Nemotron Nano 9B v2 tok/s | 121.89 | 202.68 | -40% |
| Ornith 1.5 9B tok/s | 144.04 | 225.79 | -36% |
| Gemma 4 12B tok/s | 103.57 | 151.54 | -32% |
| Qwen2.5-Coder 14B tok/s | 95.12 | 149.75 | -36% |
| Qwen3 14B tok/s | 96.37 | 143.11 | -33% |
| gpt-oss-20b tok/s | 286.23 | 415.12 | -31% |
| Gemma 4 26B A4B tok/s | 188.23 | 264.69 | -29% |
| Qwen3.6 27B tok/s | 49.02 | 76.6 | -36% |
| Qwen3.8 27B tok/s | 48.01 | 75.74 | -37% |
| Qwen3 30B A3B tok/s | 259.61 | 344.67 | -25% |
| Qwen3 30B A3B Instruct 2507 tok/s | 263.17 | 352.33 | -25% |
| Qwen3-Coder 30B A3B tok/s | 271.04 | 365.84 | -26% |
| Gemma 4 31B tok/s | 45.08 | 71.42 | -37% |
| Qwen3 32B tok/s | 44.28 | 71.15 | -38% |
| Llama 3.3 70B tok/s | 0 | 0 | n/a |
| Stable Diffusion 1.5 images/min | 80.14 | 102.36 | -22% |
| Stable Diffusion 2.1 images/min | 60.58 | 65.53 | -8% |
| SDXL Turbo images/min | 652.11 | 902.04 | -28% |
| SSD-1B images/min | 15.77 | 20.49 | -23% |
| Z-Image Turbo images/min | 7.24 | 10.725 | -32% |
| Sana 1.6B images/min | 41.47 | 51.52 | -20% |
| Stable Diffusion XL images/min | 16.28 | 21.08 | -23% |
| DreamShaper XL Lightning images/min | 91.52 | 122.85 | -26% |
| DreamShaper XL Turbo images/min | 54.35 | 70.72 | -23% |
| Playground v2.5 images/min | 9.96 | 13.28 | -25% |
| PixArt-Sigma XL images/min | 24.27 | 30.01 | -19% |
| FLUX.2 klein 4B images/min | 42.38 | 59.33 | -29% |
| Kolors images/min | 9.88 | 13.53 | -27% |
| Z-Image images/min | 1.17 | 1.72 | -32% |
| AuraFlow v0.3 images/min | 3.02 | 4.35 | -31% |
| FLUX.1 dev images/min | 0 | 2.036 | n/a |
| Qwen-Image-Edit images/min | 0 | 0 | n/a |
| Wan 2.1 1.3B frames/s | 0.673 | 0.879 | -23% |
| CogVideoX-2B frames/s | 0.517 | 0.7 | -26% |
| CogVideoX-5B frames/s | 0.17 | 0.259 | -34% |
| LTX-Video (distilled) frames/s | 6.696 | 10.449 | -36% |
| Stable Video Diffusion clips/min | 2.235 | 2.836 | -21% |
| Wan 2.2 5B (720p) frames/s | 0.43 | 0.588 | -27% |
| LTX-Video (image to video) clips/min | 4.188 | 6.152 | -32% |
| Wan 2.2 TI2V-5B (image to video) clips/min | 1.122 | 2.284 | -51% |
| Stable Video Diffusion XT clips/min | 1.212 | 1.344 | -10% |
| CogVideoX-5B I2V clips/min | 0.328 | 0.447 | -27% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA GeForce RTX 4090 is faster on 3 of 5; NVIDIA GeForce RTX 5090 on 2.
| Workflow | NVIDIA GeForce RTX 4090 | NVIDIA GeForce RTX 5090 | Difference | Cost per run |
|---|---|---|---|---|
| 24-frame storyboard | 3.8 min | 12.5 min | NVIDIA GeForce RTX 4090 3.24x faster | $0.022 vs $0.096 |
| 60-second AI short film | 5.8 min | 9.3 min | NVIDIA GeForce RTX 4090 1.60x faster | $0.033 vs $0.072 |
| Full codebase review | 10.5 min | 6.7 min | NVIDIA GeForce RTX 5090 1.57x faster | $0.060 vs $0.052 |
| Animate a batch of images | 17.8 min | 8.8 min | NVIDIA GeForce RTX 5090 2.04x faster | $0.101 vs $0.068 |
| Short social clips | 21.5 min | 24.8 min | NVIDIA GeForce RTX 4090 1.15x faster | $0.122 vs $0.192 |
Renting by the hour, NVIDIA GeForce RTX 4090 finishes 3 of 5 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA GeForce RTX 4090 | $0.340 |
| NVIDIA GeForce RTX 5090 | $0.463 |
NVIDIA GeForce RTX 4090 is 1.36x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA GeForce RTX 4090 | NVIDIA GeForce RTX 5090 | |
|---|---|---|
| VRAM | 24GB | 32GB |
| Transistors | 76,300M | 92,200M |
| Die size | 608.4 mm² | 750 mm² |
| Process node | 4 nm | 4 nm |
| Transistor density | 125.4 M/mm² | 122.9 M/mm² |
| Architecture | Ada Lovelace | Blackwell (GB202) |
| Memory bandwidth | 1008 GB/s | 1792 GB/s |
| Boost clock | 2,520 MHz | 2,407 MHz |
| TDP | 450 W | 575 W |
| Launch MSRP | $1,599 | $1,999 |
| Release | 2022-10-12 | 2025-01-30 |
FAQ
- Which is better for ai & machine learning: NVIDIA GeForce RTX 4090 or NVIDIA GeForce RTX 5090?
- NVIDIA GeForce RTX 5090 performs better for ai & machine learning, winning 59 of 61 benchmarks in our suite with an average 43.9% advantage.
- What are the main hardware differences between NVIDIA GeForce RTX 4090 and NVIDIA GeForce RTX 5090?
- NVIDIA GeForce RTX 4090 has 24GB VRAM and a 450W TDP, while NVIDIA GeForce RTX 5090 has 32GB VRAM and a 575W TDP.
- Does VRAM matter more than speed between NVIDIA GeForce RTX 4090 and NVIDIA GeForce RTX 5090?
- For AI, yes, NVIDIA GeForce RTX 5090 fits 2 more of our 12 workloads than NVIDIA GeForce RTX 4090. A model that exceeds VRAM doesn't run slower, it doesn't run at all, so the card that fits the model wins that workload outright.
- Where is the biggest performance difference between NVIDIA GeForce RTX 4090 and NVIDIA GeForce RTX 5090?
- Wan 2.2 TI2V-5B (image to video): NVIDIA GeForce RTX 5090 leads by roughly 104% (1.122 vs 2.284 clips/min) in our testing.
NVIDIA GeForce RTX 4090 full review · NVIDIA GeForce RTX 5090 full review · All AI & Machine Learning rankings