batch size and number of iterations

[alexander-sony]

Question:
say i want to benchmark 2 gpus, one with 24g and second with 48g of gpu mem. i guess that i should approximately double the batchsize when benchmarking the second gpu. consequently, the test will take much longer time, so the number of iterations should be decreased for the second gpu, or? i mean in a practical situation, the training would stop much earlier for the second gpu, if we would aim for the same level of validation accuracy for both gpus.