RTX 3090 vs RTX 4090 for local AI
Today the RTX 3090 is cheaper: $1,497 against $2,083.48 for the RTX 4090. Per gigabyte of memory the RTX 3090 wins, $62 against $86. For Qwen3.8 27B · Q4_K_M the RTX 4090 is faster, about 40 against 37 tokens per second.
| NVIDIA GeForce RTX 3090 | NVIDIA GeForce RTX 4090 | |
|---|---|---|
| Lowest price today, delivered | $1,497 on eBay | $2,083.48 on eBay |
| Price per GB | $62/GB | $86/GB |
| Memory | 24 GB GDDR6X | 24 GB GDDR6X |
| Memory bandwidth | 936 GB/s | 1,008 GB/s |
| FP16 compute, dense | 71 TFLOPS | 165 TFLOPS |
| Reference models that fit at Q4_K_M | 11 of 28 | 11 of 28 |
| Power | 350 W | 450 W |
| Form factor | Triple slot, full height | Triple slot, full height |
| Interface | PCIe 4 | PCIe 4 |
| Software platform | cuda | cuda |
| Launched | 2020 | 2022 |
| Launch price | $1,499 | $1,599 |
Qwen3.8 27B · Q4_K_M by quantization
| Quant | RTX 3090 | RTX 4090 |
|---|---|---|
| FP16 | does not fit | does not fit |
| Q8_0 | does not fit | does not fit |
| Q6_K | ~27 tok/s, 5K ctx | ~30 tok/s, 5K ctx |
| Q5_K_M | ~32 tok/s | ~34 tok/s |
| Q4_K_M | ~37 tok/s | ~40 tok/s |
| Q3_K_M | ~46 tok/s | ~50 tok/s |
| IQ2_M | ~66 tok/s | ~71 tok/s |
Fit means the weights plus an 8K context cache fit in memory; speed is the estimated decode rate. Prices are the cheapest live eBay listing that passed our filters, item plus shipping. How we estimate.