llama-bench and llama-cli inference differs

[afsara-ben]

When I run llama-bench I get an eval rate of 88t/s but the same prompt len and --n using llama-cli gives me 85 t/s. Is there a reason for the difference?