RTX A4000 fits older Ollama builds

For a budget local-LLM box, this is the right kind of compromise: reuse the old server, spend on the GPU, and accept that VRAM — not platform age — will define the experience.

• –NVIDIA positions the RTX A4000 as a single-slot 16GB GDDR6 workstation GPU with 140W power draw, which makes it easier to drop into an older chassis than bulkier gaming cards
• –PCIe 3.0 is usually not the real bottleneck for Ollama once a model is loaded into VRAM; model size, quantization, and memory headroom matter more for day-to-day coding performance
• –16GB VRAM is enough for many quantized 7B to 14B class coding models, but it is not a comfortable tier for larger agents, long contexts, or heavier multitasking
• –The system's 128GB RAM gives useful room for offload and experimentation, but an older Xeon platform will still feel secondary to a newer box once the project grows past proof-of-concept stage