Local LLM hardware: VRAM remains primary bottleneck

Local LLM performance is now defined by VRAM capacity rather than raw compute power.

• –RTX 3070 (8GB) is the "entry tier" for 2025, capable of running 7B-8B models like Llama 3 at 4-bit quantization.
• –64GB system RAM allows for offloading, but introduces a 10x+ performance penalty that often renders cognition research unusable.
• –Tools like Hugging Face's VRAM Calculator and vramio serve as the "mins" source for planning builds without trial-and-error.
• –User misidentification of VRAM (10GB vs 8GB) highlights the confusion around consumer GPU tiers for AI workloads.