First you should note that no AFQMC development has happened for some time. However we do make an effort to keep the implementation fully alive and tested. See e.g.
https://doi.org/10.1021/acs.jctc.0c00262 and earlier works by Fionn Malone and Miguel Morales for performance data. Any HPC flavored NVIDIA GPU should be good (V100, A100, H100...) and highly performant. Memory size tends to be the bottleneck for science with large basis sets and large walker counts/batch sizes, so if you have a choice, get the GPUs with larger memory. There is also HIP support for AMD GPUs. Try to get a test drive before buying. Other AFQMC codes with GPU support should have very similar demands - AFQMC maps well to batched BLAS operations, whatever language it is written in.