1x RTX 3090 (used) vs 1x RTX 5090
The 1x RTX 5090 gives a model 32 GB to work in, against 24 GB on the 1x RTX 3090 (used). The 1x RTX 5090 has 1.9× the memory bandwidth (1,792 against 936 GB/s), which sets the ceiling on generation speed. The largest 4-bit model each holds at 8,192 tokens of context: Qwen3.6 35B-A3B on the 1x RTX 3090 (used), Nemotron 3.5 Lightning 30B-A3B on the 1x RTX 5090. Rows where the two differ are tinted.
Single GPU
1x RTX 3090 (used)
The cheapest way to 24 GB of CUDA memory: one used RTX 3090 on a current AM5 desktop. It runs 24-32B dense models at 4-bit and small MoE models entirely on the GPU.
gpt-oss-20b, 4-bit: 162 tok/s
Single GPU
1x RTX 5090
The fastest single consumer card: 32 GB of GDDR7 at 1.79 TB/s. It fits 32B dense models at 8-bit or 70B at about 3 bits.
gpt-oss-20b, 4-bit: 290 tok/s
| Specification | 1x RTX 3090 (used) | 1x RTX 5090 |
|---|---|---|
| Kind | Single GPU · DIY | Single GPU · DIY |
| Chip or GPUs | RTX 3090 | RTX 5090 |
| Memory | 24 GB VRAM + 64 GB RAM | 32 GB VRAM + 64 GB RAM |
| Memory for models | 24 GB (VRAM) | 32 GB (VRAM) |
| Memory bandwidth | 936 GB/s | 1,792 GB/s |
| FP16 compute | 71.2 TFLOPS | 209.5 TFLOPS |
| Price | ~$2,944 + unpriced parts (used GPU price not included) | ~$3,234 + unpriced parts (GPU price not included) |
| Largest at 4-bit, 8k context | Qwen3.6 35B-A3B | Nemotron 3.5 Lightning 30B-A3B |
| Software | CUDA, Vulkan | CUDA, Vulkan |
| Power | 1000 W power supply | 1000 W power supply |
What each one runs
Every model in the catalog that at least one of the two holds at 4-bit with 8,192 tokens of context. Copper speeds are measured medians; outlined ones are estimates. Rows where only one fits, or where the speed differs by more than a quarter, are tinted.
| Model | 1x RTX 3090 (used) | 1x RTX 5090 |
|---|---|---|
| Nemotron 3.5 Lightning 30B-A3B30B, 3B active · 23.7 GB file | does not fit | ~175–256 tok/sestimate |
| Qwen3.6 35B-A3B35B, 3B active · 20.8 GB file | ~122–178 tok/sestimate | 117 tok/smedian of 1 run |
| Ornith 1.5 35B-A3B35B, 3B active · 20.2 GB file | ~125–183 tok/sestimate | ~240–350 tok/sestimate |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · 17.3 GB file | ~116–169 tok/sestimate | ~222–324 tok/sestimate |
| Gemma 4 31B IT30.7B · 17.1 GB file | 35 tok/smedian of 1 run | 61 tok/smedian of 1 run |
| Granite 4.2 30B29.3B · 16.8 GB file | ~25–40 tok/sestimate | ~49–77 tok/sestimate |
| Qwen3.8 27B27B · 16.2 GB file | ~26–41 tok/sestimate | ~50–79 tok/sestimate |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · 15.9 GB file | 119 tok/smedian of 1 run | 122 tok/smedian of 2 runs |
| Qwen3.6 27B27B · 15.7 GB file | ~27–43 tok/sestimate | ~52–82 tok/sestimate |
| Devstral Small 2 24B (2512)24B · 13.3 GB file | ~32–50 tok/sestimate | ~61–96 tok/sestimate |
| gpt-oss-20b21B, 3.6B active · 10.8 GB file | 162 tok/smedian of 3 runs | 290 tok/smedian of 4 runs |
| Phi-414.7B · 8.3 GB file | ~52–81 tok/sestimate | ~99–155 tok/sestimate |
| Ministral 3 14B Reasoning (2512)13.9B · 7.7 GB file | ~56–87 tok/sestimate | ~107–167 tok/sestimate |
| Gemma 4 12B IT11.95B · 6.6 GB file | ~64–101 tok/sestimate | ~123–194 tok/sestimate |
| Qwen3.5 9B9B · 5.3 GB file | ~81–127 tok/sestimate | ~155–243 tok/sestimate |
| Granite 4.2 8B8.8B · 5.2 GB file | ~83–130 tok/sestimate | ~159–249 tok/sestimate |
| Gemma 4 E4B IT8B · 5.0 GB file | ~151–237 tok/sestimate | ~289–454 tok/sestimate |
| Qwen3 8B8.2B · 4.7 GB file | 115 tok/smedian of 1 run | 200 tok/smedian of 1 run |
| Llama 3.1 8B Instruct8B · 4.6 GB file | 122 tok/smedian of 2 runs | 200 tok/smedian of 3 runs |
| Qwen3.5 4B4B · 2.6 GB file | ~167–263 tok/sestimate | ~321–503 tok/sestimate |
Full pages: 1x RTX 3090 (used) · 1x RTX 5090 · check a different quant or context length