nvprof elapsed_cycles_sm vs time in milliseconds

Why is there a difference between a. the time taken by the kernel in seconds/milliseconds displayed by nvprof in summary mode and b. elapsed_cycles_sm metric ?
I have a tesla K80 machine and I read that the Kepler architecture GPUs have only one clock domain (2x shader frequency for execution units is no longer used). The clock frequency seems to be 823500 KHz. However, elapsed_cycles_sm / 823500000 is not equal to time in seconds displayed by nvprof. I also checked time for the kernel using cudaEventRecord, which is equal to the nvprof time in seconds. What am I missing here ?