I'm planning a CachyOS gaming PC that can also run a local assistant when I'm not gaming. My current N97 system can perform iGPU inference, but it's very slow and limited. Ideally, I'd like to run a 12B model quantized to Q4, which I estimate may need around 16GB of VRAM. I'm fine unloading the AI model or switching to a much smaller one while gaming, since gaming is the main priority.
My target budget is around $1,500, though I can go higher if necessary. I'm buying in the US and will connect the system to a TV over HDMI. I need a Steam-compatible controller, while my existing keyboard and mouse can be reused. Most maintenance will probably be done over SSH.
I'm considering AMD for its generally smooth Linux gaming support, but I know NVIDIA has stronger AI software support through CUDA. I'm open to either option and would like to know whether CUDA is worth prioritizing for this kind of local inference workload.
1 Answer
Given the budget, AMD is probably the more practical choice. NVIDIA has an advantage for inference because CUDA and related tools tend to work more reliably, but for a local assistant where you’re already comfortable waiting longer than you would with a cloud service, ROCm or Vulkan may be adequate. A Radeon card in the 16GB class could provide a good compromise for gaming and occasional inference.
The main challenge is balancing the GPU with the rest of the system. A high-end GPU could consume roughly half the budget, while 32GB of RAM is useful both for gaming and for offloading model layers when needed. Once you account for the CPU, motherboard, power supply, storage, and case, a strict $1,500 ceiling may be difficult.
That’s helpful. I’m willing to go above $1,500 if needed. Would it make more sense to choose the 16GB-class GPU and 32GB of RAM first, then build the rest of the system around those priorities?