Multi GPU slower than One GPU several time

HI

When I use multi-gpu command, I found the training time is slower several times than that of one gpu.
```
trainer = pl.Trainer(strategy="ddp", accelerator="gpu", gpus=2, precision=16, callbacks=[logger])
```

Start by reproducing the reported comparison using the provided Trainer configuration with one GPU and with DDP on two GPUs. Record the training times and environment details, then investigate the multi-GPU execution path to identify the source of the slowdown. Done means the cause is isolated and the two-GPU configuration no longer takes several times longer than one GPU.

Written by the indexing model from the issue text.