Tesla P40 vs Instinct MI50 16GB for local AI
Today the Instinct MI50 16GB is cheaper: $228.78 against $273 for the Tesla P40. Per gigabyte of memory the Tesla P40 wins, $11 against $14. Only the Tesla P40 fits Qwen3.8 27B · Q4_K_M with 8K of context. The Tesla P40 holds more, 24 GB against 16 GB, and fits 11 of our 28 reference models at Q4_K_M where the Instinct MI50 16GB fits 3.
| NVIDIA Tesla P40 | AMD Radeon Instinct MI50 16GB | |
|---|---|---|
| Lowest price today, delivered | $273 on eBay | $228.78 on eBay |
| Price per GB | $11/GB | $14/GB |
| Memory | 24 GB GDDR5 | 16 GB HBM2 |
| Memory bandwidth | 346 GB/s | 1,024 GB/s |
| FP16 compute, dense | n/a | 26.5 TFLOPS |
| Reference models that fit at Q4_K_M | 11 of 28 | 3 of 28 |
| Power | 250 W | 300 W |
| Form factor | Dual slot, passive, needs a fan | Dual slot, passive, needs a fan |
| Interface | PCIe 3 | PCIe 4 |
| Software platform | cuda | rocm |
| Launched | 2016 | 2018 |
| Launch price | n/a | n/a |
Qwen3.8 27B · Q4_K_M by quantization
| Quant | Tesla P40 | Instinct MI50 16GB |
|---|---|---|
| FP16 | does not fit | does not fit |
| Q8_0 | does not fit | does not fit |
| Q6_K | ~10 tok/s, 5K ctx | does not fit |
| Q5_K_M | ~12 tok/s | does not fit |
| Q4_K_M | ~14 tok/s | does not fit |
| Q3_K_M | ~17 tok/s | ~39 tok/s |
| IQ2_M | ~24 tok/s | ~56 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.