NVIDIA · GeForce RTX 40 Series
NVIDIA GeForce RTX 4090
NVIDIA GeForce RTX 4090 specs, benchmarks, and best prices. 24GB GDDR6X, 16384 CUDA cores, Ada Lovelace architecture. Previous-gen flagship.
Performance
80
AGaming
82
AValue
32
DTypical street price (used market)
$2,250–2,500
Launch MSRP $1,599 · Released Oct 2022
Prices vary with the 2026 memory shortage. As an Amazon Associate we earn from qualifying purchases.
Full Specifications
| VRAM | 24GB |
| Memory Type | GDDR6X |
| Boost Clock | 2520MHz |
| Base Clock | 2235MHz |
| Shader/Stream Processors | 16384 |
| Performance | 82.6TFLOPS |
| TDP | 450W |
| Memory Bus | 384-bit |
| Bandwidth | 1008GB/s |
| Architecture | Ada Lovelace |
| Process Node | 4nm |
| RT Cores | 128 |
| AI Upscaling | DLSS 3.5 |
| Interface | PCIe 4.0 x16 |
| Power Connector | 1x 16-pin |
| Card Length | 336mm |
| Release Date | 2022-10-12 |
Local AI on the RTX 4090
With 24GB of VRAM, here's what this card can run and how fast — estimated from memory bandwidth, the number that actually decides LLM speed.
| Model | Best quant that fits | Est. speed | Verdict |
|---|---|---|---|
| Qwen3 30B-A3B | Q5_K_M | ~129 tok/s (Instant-feeling) | Runs great |
| Qwen3-Coder 30B-A3B | Q5_K_M | ~129 tok/s (Instant-feeling) | Runs great |
| Qwen3.6 27B | Q5_K_M | ~26 tok/s (Fast) | Runs great |
| Gemma 4 26B-A4B | Q5_K_M | ~106 tok/s (Instant-feeling) | Runs great |
| Mistral Small 3.2 24B | Q5_K_M | ~29 tok/s (Fast) | Runs great |
| gpt-oss-20b | Q8_0 | ~79 tok/s (Instant-feeling) | Runs great |
| Qwen3 8B | FP16 | ~31 tok/s (Fast) | Runs great |
| Llama 3.1 8B | FP16 | ~32 tok/s (Fast) | Runs great |
8 of 8 featured models run fully in VRAM on this card. Check a custom model size →
Compare RTX 4090 With
Understand the Specs
Browse all terms →Running AI Locally
Running large language models and AI on your own hardware instead of cloud services. Performance depends on VRAM, RAM bandwidth, and quantization — not CPU speed.
GPU (Graphics Processing Unit)
A specialized processor designed for parallel computation, primarily used for rendering graphics but increasingly important for AI and compute workloads.
VRAM (Video RAM)
Dedicated memory on a graphics card used to store textures, frame buffers, and render data. More VRAM allows higher resolutions and more detailed textures.
Ray Tracing
A rendering technique that simulates how light physically behaves — bouncing, reflecting, and refracting — to produce photorealistic lighting, shadows, and reflections in games.
DLSS (Deep Learning Super Sampling)
NVIDIA's AI-powered upscaling technology that renders games at lower resolution and uses neural networks to reconstruct near-native image quality at much higher frame rates.
FSR (FidelityFX Super Resolution)
AMD's upscaling technology that boosts frame rates by rendering at lower resolution and reconstructing detail. FSR 4 uses AI on RDNA 4 hardware; older versions work on any GPU.