5070 Ti vs 3090: VRAM wins for LLMs

Raw VRAM capacity remains the ultimate bottleneck for local LLM inference, making the older RTX 3090 a better value for developers targeting 30B+ models. The 24GB VRAM on the 3090 enables larger models and deeper context windows that 16GB cards cannot accommodate, though the 5070 Ti's 5th Gen Tensor Cores and GDDR7 offer higher tokens-per-second for 7B-14B models using low-precision FP4. While Blackwell's larger L2 cache mitigates the bandwidth disadvantage of its 256-bit bus, the used 3090 remains the price-to-parameter leader for local inference.