In local single-GPU inference scenarios, the bottleneck is typically memory bandwidth; when a single verification pass can confirm multiple draft tokens simultaneously, the average number of memory reads per token is greatly reduced, which is the physical source of the speedup.