Dual RTX 5080 vs 3090: VRAM vs Speed Dilemma

The "VRAM tax" is reaching a breaking point for local AI hobbyists where 4-year-old hardware remains more viable than current-gen flagships.

• –16GB is the new "minimum viable" for 7B-9B models with long context, but a dual-5080 setup (32GB) still hits a hard ceiling before the dual-3090 (48GB) sweet spot.
• –Blackwell's native FP4/FP8 support offers massive throughput gains, yet these are irrelevant if the model weights cannot fit in VRAM without extreme quantization.
• –Mixing architectures (5080 + 3090) works technically but bottlenecks system speed to the slowest card, negating the 5080's premium price tag.
• –Massive price gouging on the RTX 5090 ($3,900 in some regions) is driving a "frankenstein" multi-GPU market for mid-range cards.