I'm currently running it on an RTX 3090 (street price ~$1000 USD) with a long context and getting pretty good performance.

Prefill: ~1000 tok/s

Decode: 75-100 tok/s

It'll be far faster on a 5090, but I find the above performance to be acceptable. I've seen some claims that it even works OK on an AMD RX 7900XT (~$500USD)