RTX 3090 vs Tesla P40 for local AI
Today the Tesla P40 is cheaper: $273 against $1,497 for the RTX 3090. Per gigabyte of memory the Tesla P40 wins, $11 against $62. For Qwen3.8 27B · Q4_K_M the RTX 3090 is faster, about 37 against 14 tokens per second.
| NVIDIA GeForce RTX 3090 | NVIDIA Tesla P40 | |
|---|---|---|
| Lowest price today, delivered | $1,497 on eBay | $273 on eBay |
| Price per GB | $62/GB | $11/GB |
| Memory | 24 GB GDDR6X | 24 GB GDDR5 |
| Memory bandwidth | 936 GB/s | 346 GB/s |
| FP16 compute, dense | 71 TFLOPS | n/a |
| Reference models that fit at Q4_K_M | 11 of 28 | 11 of 28 |
| Power | 350 W | 250 W |
| Form factor | Triple slot, full height | Dual slot, passive, needs a fan |
| Interface | PCIe 4 | PCIe 3 |
| Software platform | cuda | cuda |
| Launched | 2020 | 2016 |
| Launch price | $1,499 | n/a |
Qwen3.8 27B · Q4_K_M by quantization
| Quant | RTX 3090 | Tesla P40 |
|---|---|---|
| FP16 | does not fit | does not fit |
| Q8_0 | does not fit | does not fit |
| Q6_K | ~27 tok/s, 5K ctx | ~10 tok/s, 5K ctx |
| Q5_K_M | ~32 tok/s | ~12 tok/s |
| Q4_K_M | ~37 tok/s | ~14 tok/s |
| Q3_K_M | ~46 tok/s | ~17 tok/s |
| IQ2_M | ~66 tok/s | ~24 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.