Why exactly does performance improve as the secured GPU memory increases?

[kyuheejang]

Performance appears to improve as GPU memory utilization increases.

Performance improved when using two GPUs in tensor parallel rather than using only one GPU. It seems that increasing GPU memory has a greater impact on performance than computing speed.

Is the performance improving as the gpu memory increases because the batch size increases - continuous batching, or because the number of running sequence groups increases - PagedAttention scheduling?