r2plus1d - strange relationship with inference batch size and inference speed

I'm seeing a weird relationship between inference batch size and inference speed that I can't explain. I have an r2plus1d resnet18 model that I trained. Input size is 8x32x32x3. I've been running some tests with fake data (random arrays) to measure inference speed as a function of batch size. For some reason, when the batch size increases from 6 to 7, the throughput suddenly drops significantly. It then slowly increases with the batch size until going from batch size of 24 to 25, at which point the throughput nearly doubles. I'm running this on a machine with one RTX 2080 with no other major processes running at the same time. I can run the experiment multiple times and I see the same phenomenon. I'm using Torch 1.9.1+cu102 and Torchvision 0.10.1+cu102. If anyone has any insight that might explain what I'm seeing, I'd appreciate it.

No repository file, test, or entry point is named. Start by reproducing the reported r2plus1d ResNet-18 benchmark with fake data across batch sizes 6–25 on the stated Torch and Torchvision versions, then compare timings and profiling results. Done means identifying a reproducible explanation for the throughput changes or documenting the cause.