I am currently running 2x RTX 5060 ti and happened across some good sales for additional ones coinciding with a really good sale of a highend Z890 motherboard (replacing my B860 board) that could support quad GPUs (with 2 M.2 adapters, ending with running 1 GPU at 5.0 x8 and the rest at 5.0 x4, all via CPU lanes).
2x 5060 ti 16gb discounted is about the same price (~960€) as 1 used 3090 (most i can find are actuall ~1000€).
I am wondering how such a quad 5060 setup compares to dual RTX 3090 in prefill and generation speed (on higher quality quants of Qwen 3.6 27B for example, like int8 / fp8)?
RTX 5060 ti can easily OC memory (+3000Mhz), providing close to 500gb/s bandwidth, so looking at bandwidth per GB, its pretty close overall, and looking at FP8 TFLOPs the 5060 also comes out ahead. However, tensor parallelism is not exactly perfect scaling so I am curious where it ends up.
[link] [comments]