NVIDIA RTX 3090 vs NVIDIA RTX 5090 for local LLMs
219 reports on the NVIDIA RTX 3090 and 170 on the NVIDIA RTX 5090
Which is faster for local LLMs?
On Qwen3.8 27B at 4-bit, the NVIDIA RTX 3090 has 38.4 generation tokens/s (5 runs); prompt processing 561.0 tokens/s (2 runs); the NVIDIA RTX 5090 has 80.0 generation tokens/s (7 runs). Those are community reports whose engines and settings differ, so the gap is a rough sign, not a controlled measure of the two cards. Memory bandwidth, which caps how fast a card can write, is 1.9x higher on the NVIDIA RTX 5090 (936 GB/s on the NVIDIA RTX 3090, 1,792 on the NVIDIA RTX 5090). Across 4 shared models with generation measurements on both cards, the NVIDIA RTX 3090's median is higher on 0 and the NVIDIA RTX 5090's on 4.
- Memory
- 24GB
- Memory bandwidth
- 936 GB/s
- FP16 compute
- 142 TFLOPS
- Reports
- 219
- Memory
- 32GB
- Memory bandwidth
- 1,792 GB/s
- FP16 compute
- 419 TFLOPS
- Reports
- 170
Measured on both
Generation and prompt processing speeds from plain runs (one device, one request, no speculative decoding, the whole model in the device's memory) of the same model size at the same bit class. Each side shows the engines and reported context behind its median. Context can mean a configured window or prompt depth, so these are community comparisons with differing settings. How to read these
| Model | NVIDIA RTX 3090 | NVIDIA RTX 5090 |
|---|---|---|
| Qwen3.8 27B4-bit | 38.45 runsPP 561.0 (2 runs)llama.cpp, vLLM128K to 256K context, 1 not stated | 80.07 runsllama.cpp, TensorSharp +1, 2 not stated40K to 256K context, 2 not stated |
| Gemma 4 31B4-bit | 36.5one runllama.cppcontext not stated | 64.0one runOllama32K context |
| Qwen3.6 35B · 3B active4-bit | 101.7one runllama.cpp64K context | 215.1one runllama.cppcontext not stated |
| Qwen3.8 27B5-bit | 35.0one runengine not stated64K context | 59.0one runvLLM64K context |
| Qwen3.5 35B · 3B active4-bit | -0 runsPP 980.0 (one run)llama.cpp64K context | 194.0one runllama.cppcontext not stated |
Estimated by the calculator
The quant the calculator recommends for each card and its estimated speed at a 32,768-token context with 32 GB of system RAM, from memory bandwidth and the model's shape. Where a model was also measured above, the measurement wins.
| Model | NVIDIA RTX 3090 | NVIDIA RTX 5090 |
|---|---|---|
| Qwen3.8 27B | 33.9 t/sQ4_K_M | 44.9 t/sQ6_K |
| Gemma 4 31B | 33.2 t/sQ4_K_S | 46.2 t/sQ5_K_M |
| Qwen3.6 35B · 3B active | 120.1 t/sQ4_K_S | 163.0 t/sQ5_K_M |
| Qwen3.5 35B · 3B active | 120.1 t/sQ4_K_S | 163.0 t/sQ5_K_M |
| Qwen3.8 125B · 6B active | 19.4 t/sQ2_K, experts in RAM | 20.3 t/sQ2_K, experts in RAM |
| DeepSeek V4 Flash 284B · 13B active | doesn't fit at these settings | doesn't fit at these settings |
| Qwen3.6 27B | 33.9 t/sQ4_K_M | 44.9 t/sQ6_K |
| Gemma 4 26B · 4B active | 76.4 t/sQ5_K_S | 114.5 t/sQ6_K |
| DeepSeek V4.1 Flash 552B · 16B active | doesn't fit at these settings | doesn't fit at these settings |
| GLM-5.3 320B · 18B active | doesn't fit at these settings | doesn't fit at these settings |
Change the context, RAM or model in the calculator: NVIDIA RTX 3090 · NVIDIA RTX 5090
Every report on each card, with its source: NVIDIA RTX 3090 · NVIDIA RTX 5090