A GPU-heavy local AI setup shows where a Mac mini may fall short
The current machine uses an Aorus Elite X570 , an R9 5950X processor, an RTX 5090, 64 GB of memory, a 2 TB SSD, and a 1200W . It runs as an Ollama server that controls an OpenClaw agent on a separate mini PC in the same network. It also runs ComfyUI to create 1440p video with the LTX 2.3 base BF16 model, a distilled LoRA, and a Gemma 3 text encoder.
The current setup can produce about 7 to 9 seconds of video before hitting an OOM error. The planned upgrade is to add an RTX 5060 Ti with 16 GB of VRAM. Moving the monitor to the 5060 Ti could free about 2 GB of VRAM on the RTX 5090 because Windows display use would move to the smaller card.
For Ollama, the goal is to move from Q4 with a 192K context to Q8 with a 256K context by combining both GPUs. The expected tradeoff is about 10% to 15% less speed than the old Q4 setup on the 5090 alone, in exchange for better quality and a longer context.
Key points
- The RTX 5090 setup reaches a VRAM limit after about 7 to 9 seconds of 1440p .
- The proposed RTX 5060 Ti 16 GB upgrade is mainly about adding VRAM, not maximizing speed.
- Using the 5060 Ti for the monitor may save about 2 GB of VRAM on the RTX 5090.
- The Ollama goal is to move from Q4 192K context to Q8 256K context.
- The expected cost is 10% to 15% lower speed, with better quality and a longer context.