RetinaNet detector framework and choosing suitable hardware for detection of 3D structures.

[DSRajesh]

Dear Team
Thank you for developing MONAI. It is a fantastic toolkit. A quick question about the RetinaNet detector framework and choosing suitable hardware for detection of 3D structures.
We are looking at nodule detection ~5-15mm in diameter from CT scans (typical resolution of 0.6mm inplane and .75 to 1.5mm slice thickness).

1. Does RetinaNet detector support multi-GPU training. (a) Both GPUs in the same server(b) GPUs across different servers.
2. GPU texture memory Would 1 server with (a) Single A6000 (NVIDIA RTX A6000: 48GB memory, 10752 CUDA cores, 336 Tensor cores, NVLink older gen) GPUvs(b) Two RTX 4090 (24 GB 16384 cores each) be better ?
   A single A6000 has about 2/3rds the number of CUDA cores as a single RTX 4090. So it is slower. However it has twice the GPU memory. A single A6000 also costs twice as much as an RTX4090.
   I guess the question is that is it ok to split the GPU memory across 2 GPUs. (Related to Question 1), as far as RetinaNet detector training is concerned.

Thanks