GeForce RTX 4070 vs RTX 3090
VS
Nvidia GeForce RTX 4070
AD104
Nvidia GeForce RTX 3090
GA102
We compared two discrete desktop gaming GPUs: the GeForce RTX 4070 12 GB with 46 pipelines and 5888 shaders against the 2 years and 7 months older RTX RTX 3090 24 GB that utilizes 82 pipelines and 10496 shaders. Here you will find complete details about specs, efficiency, performance tests, and more.
Review
Gaming
Performance in DirectX, OpenCL, and Vulkan games
Workstation
Perf. in 3D modeling, video editing and rendering apps
AI/ML
Capabilities for machine learning and AI-related tasks
Energy Efficiency
Power consumption efficiency in different scenarios
NanoReview Final Score
Overall video card score
Value for money (Beta)
Key differences
Reasons to consider the GeForce RTX 4070
- Manufactured using a more efficient 5 nm process technology
- Supports Nvidia DLSS 3 technology
Reasons to consider the GeForce RTX 3090
- Performs better (up to 21%) in 3DMark Steel Nomad Lite
- Shows 13% higher average frame rate in modern games at QHD resolution – 120 vs 106 FPS
- 22% higher maximum theoretical performance (35.6 vs 29.1 TFLOPS)
- Includes 12 GB more video memory
- Has 86% higher memory bandwidth: 936.2 vs 504 GB/s
- Features 144 more tensor cores for effective ML and AI workloads
- Achieves 26% more points in the GeekBench 6 Compute test (211K vs 167K)
- Has 78% more shading units (10496 vs 5888)
Gaming Performance
Frame rate comparison across popular AAA titles at different resolutionsGames
| Forza Horizon 5 | 172 115 102 79 |
186 128 117 91 |
| The Witcher 3 | 194 159 114 61 |
205 170 128 66 |
| Counter-Strike 2 | 272 205 149 81 |
275 207 155 83 |
| Far Cry 6 | 149 133 115 68 |
156 145 124 77 |
| Hogwarts Legacy | 117 97 77 39 |
128 108 86 47 |
| CoD: Modern Warfare III | 149 145 107 69 |
171 159 123 80 |
| Ghost of Tsushima | 106 88 72 43 |
120 97 85 48 |
| Cyberpunk 2077 | 129 119 75 37 |
148 132 90 43 |
| Shadow of the Tomb Raider | 211 190 145 80 |
237 226 168 91 |
| 1080p High | 167 | 181 |
| 1080p Ultra | 139 | 152 |
| 1440p Ultra | 106 | 120 |
| 4K Ultra | 62 | 70 |
| Margin of Error | Low | Low |
Benchmarks
Graphics cards’ performance in recent benchmarking apps3D Mark
Steel Nomad Lite Score
| Time Spy | 17799 | 19896 |
| Solar Bay | 82861 | 97395 |
| Port Royal | 11156 | 13633 |
| Fire Strike | 44067 | 47512 |
| Wild Life Extreme | 35472 | 43669 |
| Night Raid | 153385 | 139766 |
GeekBench 6 OpenCL
GB6 Compute Score
| Background Blur | 233 img/sec | 320.8 img/sec |
| Face Detection | 141.7 img/sec | 207.2 img/sec |
| Horizon Detection | 6.62 Gpixels/sec | 10.1 Gpixels/sec |
| Edge Detection | 9.19 Gpixels/sec | 14.4 Gpixels/sec |
| Gaussian Blur | 9.2 Gpixels/sec | 11.5 Gpixels/sec |
| Feature Matching | 1.78 Gpixels/sec | 1.48 Gpixels/sec |
| Stereo Matching | 845.4 Gpixels/sec | 984.2 Gpixels/sec |
| Particle Physics | 25043.9 FPS | 27764.9 FPS |
| API | OpenCL | OpenCL |
Cinebench 2024 GPU
Cinebench 2024 GPU
Passmark Graphics
G3D Mark Score
| G2D Mark | 1165 | 1069 |
| DirectX 11 | 243 FPS | 218 FPS |
| DirectX 12 | 103 FPS | 109 FPS |
| GPU Compute | 14570 Ops/s | 15028 Ops/s |
Blender
Blender GPU
Recent User Tests
GeForce RTX 4070
| Date | Benchmark | Result |
|---|---|---|
| 📘 2025-12-14 (kashitsu) | Cinebench 2024 | 18051 |
GeForce RTX 3090
No benchmark results yet
Artificial Intelligence Tests
Performance in machine learning and artificial intelligence tasksGeekBench 6 ML
GB6 ML Single Precision
GB6 ML Half Precision
GB6 ML Quantized
| Image Classification (SP) | 10246 | 10665 |
| Image Segmentation (HP) | 26153 | 28963 |
| Image Super Resolution (Q) | 32054 | 22697 |
| Face Detection (HP) | 51337 | 55232 |
| Pose Estimation (Q) | 124617 | 128155 |
| Text Classification (SP) | 3703 | 3152 |
| Machine Translation (HP) | 5484 | 5088 |
| Object Detection (SP) | 13134 | 13891 |
| Depth Estimation (Q) | 50548 | 37984 |
| Style Transfer (SP) | 263477 | 272000 |
| Framework | ONNX | ONNX |
| Backend | DirectML | DirectML |
Specifications
Technical specifications of GeForce RTX 4070 and RTX 3090General
| Vendor | Nvidia | Nvidia |
| Build | Discrete | Discrete |
| Released | April 13, 2023 | September 24, 2020 |
| Launch price (MSRP) | $599 | $1499 |
| Case | Desktop | Desktop |
| Purpose | Gaming | Gaming |
| Segment | Mid-range | High-end |
| Architecture | Ada Lovelace | Ampere |
| GPU Codename | AD104 | GA102 |
| Rival Equivalent | - Radeon RX 7900 XT | - Radeon RX 6900 XT |
| Successor | - GeForce RTX 5070 | - GeForce RTX 5090 |
| Recommended CPU | - Intel Core i7 14700K or above | - Intel Core i9 12900K or above |
Desktop GPU rating (28th and 20th place)
Graphics Processing Unit
| Base Clock | 1920 MHz | 1395 MHz |
| Boost Clock | 2475 MHz | 1695 MHz |
| Shading Units | 5888 | 10496 |
| Texture Mapping Units (TMUs) | 184 | 328 |
| Render Output Units (ROPs) | 64 | 112 |
| Compute Units (Pipelines) | 46 | 82 |
| Tensor Cores | 184 | 328 |
| Ray-tracing Cores | 46 | 82 |
| L1 Cache | 128KB per cluster | 128KB per cluster |
| L2 Cache | 36MB shared | 6MB shared |
| Instructions Per Cycle | 2 IPC | 2 IPC |
Raw Performance
| Pixel Fill Rate | 158 GPixel/s | 190 GPixel/s |
| Texture Fill Rate | 455 GTexel/s | 556 GTexel/s |
FLOPS (FP32)
Physical
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 |
| TGP | 200 W | 350 W |
| Manufacturing | TSMC | Samsung |
| Fabrication Process | 5 nm | 8 nm |
| Die Size | 294 mm² | 628 mm² |
| Transistor Count | 35 billion | 28 billion |
| Transistor Density | 119.05 MTr/mm² | 44.59 MTr/mm² |
Memory
| Memory Type | GDDR6X | GDDR6X |
| Memory Size | 12 GB | 24 GB |
| Memory Clock | 1313 MHz | 1219 MHz |
| Effective Memory Speed | 21000 Mbps | 19500 Mbps |
| Bus | 192-bit | 384-bit |
| ECC | No | No |
Memory Bandwidth
API
| DirectX | 12 | 12 |
| Vulkan | 1.3 | 1.3 |
| OpenGL | 4.6 | 4.6 |
| OpenCL | 3.0 | 3.0 |
| CUDA | 8.9 | 8.6 |
| Ray Tracing | Yes | Yes |
| DLSS | DLSS 3 | DLSS 2 |
| DisplayPort | 1.4a | 1.4a |
Cast your vote
Total votes: 58