Database Mart Ollama 0.5.11 benchmark, RTX A4000 16GB

Database Mart hosted-GPU report, page footer dated 10/07/2026. One NVIDIA RTX A4000 (the page calls it 'Quadro RTX A4000'; compute capability 8.6, Ampere, 6,144 CUDA cores, 16 GB GDDR6 is the RTX A4000), run as a Windows VPS. Engine Ollama 0.5.11 (llama.cpp backend; the llama.cpp commit is not printed). Quantization is the default Ollama pull, labelled 4-bit, exact GGUF quant not printed, so format_id is null. Prompt is a single short question and output length is model-determined, so only the eval-rate column (per-request decode, batch size 1) is recorded. The gemma2:27b row is recorded WITH its GPU-utilization figure of 1% and CPU rate of 70-86%, because the 16 GB weight does not fit in 16 GB VRAM and the number is a CPU-spill measurement, not a GPU decode rate.