RTX 4090 vs RTX 5090 for local AI
Today the RTX 4090 is cheaper: $2,083.48 against $4,920.99 for the RTX 5090. Per gigabyte of memory the RTX 4090 wins, $86 against $153. For Qwen3.8 27B · Q4_K_M the RTX 5090 is faster, about 71 against 40 tokens per second. The RTX 5090 holds more, 32 GB against 24 GB, and fits 11 of our 28 reference models at Q4_K_M where the RTX 4090 fits 11.
| NVIDIA GeForce RTX 4090 | NVIDIA GeForce RTX 5090 | |
|---|---|---|
| Lowest price today, delivered | $2,083.48 on eBay | $4,920.99 on eBay |
| Price per GB | $86/GB | $153/GB |
| Memory | 24 GB GDDR6X | 32 GB GDDR7 |
| Memory bandwidth | 1,008 GB/s | 1,792 GB/s |
| FP16 compute, dense | 165 TFLOPS | 210 TFLOPS |
| Reference models that fit at Q4_K_M | 11 of 28 | 11 of 28 |
| Power | 450 W | 575 W |
| Form factor | Triple slot, full height | Dual slot (Founders Edition), partner cards up to 4 slots |
| Interface | PCIe 4 | PCIe 5 |
| Software platform | cuda | cuda |
| Launched | 2022 | 2025 |
| Launch price | $1,599 | $1,999 |
Qwen3.8 27B · Q4_K_M by quantization
| Quant | RTX 4090 | RTX 5090 |
|---|---|---|
| FP16 | does not fit | does not fit |
| Q8_0 | does not fit | ~41 tok/s |
| Q6_K | ~30 tok/s, 5K ctx | ~53 tok/s |
| Q5_K_M | ~34 tok/s | ~61 tok/s |
| Q4_K_M | ~40 tok/s | ~71 tok/s |
| Q3_K_M | ~50 tok/s | ~88 tok/s |
| IQ2_M | ~71 tok/s | ~130 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.