Researcher eyes $3,000 GPU upgrade for local LLMs

VRAM capacity remains the ultimate bottleneck for local research, making raw compute speed secondary to total memory for large-model inference.

• –**VRAM is King:** For summarizing 100+ books, memory capacity is more critical than speed; while the RTX 5090's 32GB is a major leap, dual RTX 3090s (48GB) still offer a better "fit" for 70B models at high precision.
• –**PSU Underpowered:** A 750W Platinum power supply is insufficient for a flagship NVIDIA card paired with an existing secondary GPU; a 1000W+ upgrade is mandatory to avoid system crashes during heavy inference.
• –**Model Synergy:** The Qwen3-Coder-30B-A3B mentioned is a top-tier choice for this use case, offering MoE efficiency and a native 256K context window that fits comfortably in consumer VRAM.
• –**Hackintosh Viability:** Keeping the RX 5700 XT for macOS while using an NVIDIA card for WSL2/Windows is a proven strategy, but requires careful PCIe lane management to avoid bandwidth throttling on the secondary slot.