Local LLM Build Hits VRAM Reality

The right instinct here is to optimize for VRAM first, because local inference is usually constrained by how much model you can keep on the GPU, not by whether the CPU is flashy. For this kind of build, a used 24GB card is much more compelling than a newer 16GB card, and the CPU choice matters far less than the poster thinks.

• –A Ryzen 7 9700X is already plenty for a local inference box; LLM serving is usually GPU-bound, not CPU-bound.
• –DDR5 is nice, but not worth sacrificing GPU budget for unless the platform choice already forces it; the practical win is more usable model capacity, not theoretical RAM bandwidth.
• –A used RTX 3090 Ti’s 24GB VRAM is the strongest option in the budget range if the card is healthy and priced well.
• –The “buy a 5060 Ti, then trade up” plan adds friction and risk; if the end goal is 24GB VRAM, it is usually better to buy for that target directly.
• –For a 30B-class model, system RAM matters for offload and context handling, but it is a secondary lever compared with raw VRAM capacity.