RTX 5070 Ti, 3090 Split Local LLM Buyers

For local LLMs, VRAM headroom usually matters more than a small generational speed gap once context windows stretch into six figures.

• –Two-GPU setups rarely behave like a clean pooled-memory system, so the effective ceiling is still constrained by per-card allocation and sharding strategy.
• –A 3090's 24GB is the safer path for 32B dense models plus very long contexts; KV cache growth can eat the 16GB card fast.
• –The 5070 Ti is the cleaner buy if you value new-in-box reliability, lower risk, and Blackwell-era tensor features more than absolute headroom.
• –For 120B MoE workloads at 30+ tps, the 3090 path is more likely to avoid constant offload compromises, especially as context scales.
• –If your real target is Q4/IQ4 experimentation rather than near-saturated 70B+ throughput, the 5070 Ti + 4070 combo may be enough with careful model choice.