Same model, quantization, 128K context and settings; GPU selection changed. Three warm samples per input, 128 output tokens, thinking off, no cached prompt tokens. Run order: GPU 0, then a fresh two-GPU control.
| Input tokens | 1 GPU tokens/s | 2 GPU tokens/s | 1 GPU reply seconds | 2 GPU reply seconds |
|---|---|---|---|---|
| 84 | 43.6 | 52.1 | 3.75 | 4.59 |
| 6,586 | 64.3 | 74.1 | 8.81 | 15.54 |
| 26,087 | 61.4 | 72.9 | 23.53 | 33.20 |
Small sequential sample with an adapting expert cache; single-GPU accuracy and long reasoning were not retested. The slow second slot’s effect was not isolated.
All nine paired speed requests and 18 retained answers · Summary, sample ranges and aggregate resources
Single-GPU exact calls · Two-GPU exact calls
On this workstation, one RTX 5060 Ti on the PCIe 5 ×8 slot was sufficient for the tested workload. It retained 84–87% of the fresh two-GPU generation rate and completed the 128-token replies 18–43% sooner. The second card on PCIe 3 ×1 offered a modest generation gain without improving those full-reply times.
This supports using one GPU for these short-reply workloads. We did not isolate the bus’s contribution or test a faster second slot. Longer reasoning-heavy replies and single-GPU accuracy were not retested.