GPU / RTX-5090
NVIDIA GeForce RTX 5090
High-end local inference, quantized 30B-class models
Spec sheet
NVIDIA
- Vendor
- NVIDIA
- Architecture
- Blackwell
- Memory
- 32 GB GDDR7
- Bandwidth
- 1.79 TB/s
- Power
- 575 W
- Class
- Consumer
- Form factor
- PCIe card
- Interconnect
- PCIe 5.0
- Released
- 2025
- Workloads
- Inference · Fine-tuning
Model fit / Inference
8K ctx · batch 1 · ≥10% headroom
| Model | FP16 | INT8 | INT4 |
|---|---|---|---|
| Llama 3.2 1B | 1×3.0 GB | 1×1.6 GB | 1×0.95 GB |
| Llama 3.2 3B | 1×8.0 GB | 1×4.5 GB | 1×2.7 GB |
| Llama 3.1 8B | 1×18.7 GB | 1×9.9 GB | 1×5.5 GB |
| Llama 3.1 70B | 8×158.0 GB | 4×80.3 GB | 2×41.5 GB |
| Llama 3.3 70B | 8×158.0 GB | 4×80.3 GB | 2×41.5 GB |
| Llama 3.1 405B | —897.2 GB | —450.7 GB | 8×227.5 GB |
| Qwen2.5 1.5B | 1×3.6 GB | 1×1.9 GB | 1×1.1 GB |
| Qwen2.5 3B | 1×7.1 GB | 1×3.7 GB | 1×2.0 GB |
| Qwen2.5 7B | 1×17.2 GB | 1×8.9 GB | 1×4.7 GB |
| Qwen2.5 14B | 2×34.0 GB | 1×17.8 GB | 1×9.7 GB |
| Qwen2.5 32B | 4×73.6 GB | 2×37.9 GB | 1×20.0 GB |
| Qwen2.5 72B | 8×162.6 GB | 4×82.7 GB | 2×42.7 GB |
| Mistral 7B | 1×17.0 GB | 1×9.0 GB | 1×5.1 GB |
| Mistral NeMo 12B | 1×28.2 GB | 1×14.8 GB | 1×8.1 GB |
| Mistral Small 24B | 2×53.3 GB | 1×27.3 GB | 1×14.3 GB |
| Mixtral 8x7B | 4×103.8 GB | 2×52.4 GB | 1×26.8 GB |
| Mixtral 8x22B | —312.1 GB | 8×157.0 GB | 4×79.4 GB |
| Gemma 2 9B | 1×23.1 GB | 1×13.0 GB | 1×7.9 GB |
| Gemma 2 27B | 4×62.9 GB | 2×33.0 GB | 1×18.0 GB |
| Phi-3 Mini 3.8B | 1×11.6 GB | 1×7.4 GB | 1×5.3 GB |
| Phi-3 Medium 14B | 2×32.5 GB | 1×17.1 GB | 1×9.4 GB |
| DeepSeek R1 Distill Qwen 32B | 4×74.3 GB | 2×38.2 GB | 1×20.2 GB |
| DeepSeek R1 Distill Llama 70B | 8×158.0 GB | 4×80.3 GB | 2×41.5 GB |
| DeepSeek V3 671B | —1,477 GB | —738.7 GB | —369.6 GB |