GeForce RTX 5060 Ti faces 3090 wall

Short version: the software stack will keep improving, but it won't make the memory bus disappear. FP4-aware kernels can narrow the gap, yet the 3090's wider memory system and extra VRAM still matter most for single-user local inference.

• –Blackwell really does add FP4 and FP6 tensor support, so the user's instinct about future optimization is directionally right.
• –GGUF or q4 alone does not guarantee FP4 execution; the runtime has to have a matching kernel path, and attention is only one piece of inference.
• –The GeForce RTX 3090's 24GB GDDR6X and 936 GB/sec bandwidth still buy more headroom for larger models, longer context, and fewer offloads.
• –Smaller quants and slimmer models make 16GB more viable, but context growth and MoE tradeoffs keep VRAM demand alive.
• –The GeForce RTX 5060 Ti wins on power, thermals, and buying-new peace of mind, which makes it a better efficiency buy even if it is not a 3090 replacement.