Benchmark: Windows 11 vs Lubuntu 26.04 on Llama.cpp (RTX 5080 + i9-14900KF). I didn't expect the gap to be this big.
Lubuntu 26.04 outperforms Windows 11 by over 2x on CPU-heavy AI workloads.
A Reddit user running Llama.cpp b8929 on identical hardware—Intel Core i9-14900KF, 64GB DDR5, RTX 5080 16GB VRAM—compared Windows 11 25H2 against Lubuntu 26.04. They tested five models including Gemma 4, Qwen3.5, and OSS-120B across fully offloaded GPU and hybrid CPU/GPU configurations.
Results show Linux consistently outperforming Windows. For fully GPU-offloaded models, prompt processing was 6-21% faster on Linux. On hybrid models using CPU threads (-t 8 -tb 8), Linux obliterated Windows by 109-143% in prompt speed—Qwen3.5-35B hit 742 t/s on Linux vs just 305 on Windows. Generation speeds were 4-8% faster across the board. The user noted their Windows build was a fresh install with no bloatware, making the gap even more striking. The findings suggest significant OS-level overhead in Windows for AI inference tasks, especially those relying on CPU parallelism.
- Linux was 109-143% faster than Windows for prompt processing on hybrid CPU/GPU models with Llama.cpp
- Fully GPU-offloaded models saw 6-21% prompt speed gains on Linux, with generation 4-8% faster
- Tested on RTX 5080, i9-14900KF, 64GB DDR5 with identical Llama.cpp build and CUDA versions
Why It Matters
AI practitioners running local models can gain 2x+ performance on CPU-heavy tasks by switching to Linux.