Same effective batch does not mean same training time with gradient accumulation, tested on LoRA on T4 and L4 [D]

I had assumed 1 × 4, 2 × 2 and 4 × 1 will take somewhat similar time because effective batch is 4 in all cases. They did not. I ran Qwen3-1.7B with TRL and LoRA for 100 optimizer updates. GPU 1 × 4 2 × 2 4 × 1 T4 287.6s 258.8s 238.2s L4 213.02s 119.47s 124.76s Model, data, sequence length, precision