One desktop, measured limits.
This is the personal rig behind my local AI experiments. Publishing the setup makes it easier to judge whether a result is relevant to your own hardware.
Verified . A recorded snapshot, not live monitoring.
Current rig
- GPU
NVIDIA GeForce RTX 5060 Ti
16 GB VRAM; nvidia-smi reports 16,311 MiB available.
- CPU
AMD Ryzen 5 7500F
6 cores, 12 threads.
- Memory
32 GB DDR5
2 × 16 GB Kingston KF560C30-16, configured at 6000 MT/s.
- Storage
1 TB Kingston NV3 SSD
KINGSTON SNV3S1000G.
- Operating system
Windows 11 Pro
64-bit, build 26200. WSL2 Ubuntu 24.04 is used in serving experiments.
- NVIDIA driver
610.88
Driver at the time of this snapshot.
Software across the experiments
I use llama.cpp and llama-server, Ollama, and vLLM under WSL2, depending on the study. GGUF models and Qwen3.5-9B Q4_K_M recur in the small-model tests. These are tools used across different runs, rather than one fixed benchmark configuration.
The research write-ups specify the model, quantisation, context, engine and settings that produced each result. Use those study-specific records when reproducing a measurement.
What this setup changes
The 16 GB GPU limits how much of a model and its context can stay in video memory. The 32 GB system-memory budget also matters when weights spill onto the CPU. A model loading successfully does not establish that it answers fast enough or accurately enough for a task.
These are personal, clean-room experiments using my own code and runs with public datasets or synthetic tasks. Results on this one machine do not establish performance across other GPUs or real business workflows.
Setup history
: first published rig snapshot, NVIDIA driver 610.88. The earlier serving study used driver 610.74. Historical results retain their original configuration; this page describes the current rig.