Lubuntu outperforms on llama.cpp

Performance was compared by running llama.cpp b8929 under identical conditions on Windows 11 25H2 and Lubuntu 26.04.

The environment was i9-14900KF, RTX 5080 16GB, 64GB DDR5 6800, and CUDA 13.1, with Windows using the official prebuilt and Linux built directly with CMake.

Token generation speed was overall 4~8% faster on Lubuntu.

• Gemma-4-E4B-it: 111.7 t/s → 116.7 t/s (+4.4%)
• GPT-OSS-20B: 195.8 t/s → 206.2 t/s (+5.3%)
• Qwen3.6-27B: 43.8 t/s → 46.0 t/s (+5.0%)

Prompt processing showed a bigger gap. Linux led on fully GPU-offloaded models, and the gap widened significantly in the CPU/GPU hybrid range.

• Gemma-4-E4B-it: 6,232 t/s → 7,587 t/s (+21.7%)
• Qwen3.5-35B-A3B: 305 t/s → 742 t/s (+143.2%)
• GPT-OSS-120B: 310 t/s → 649 t/s (+109.3%)

The author concluded that while the felt difference in reading speed alone isn't a strong reason to switch OS, Linux's advantage is clear in prompt evaluation and hybrid execution.