It’s the maximum rate at which the hardware can process under perfect conditions. So if the instruction streams is formed perfectly to allow multi-issue of instructions, data in the caches at the right times, no pipeline stalls or flushes, nothing gating performance before this state in the pipeline, nothing going slower and unable to consume vertices after it in the pipeline.
All things considered GPUs are incredibly efficient, you don’t even want to think about how much of the maximum potential you aren’t seeing from even good CPUs like Core 2 or Athlon64, then you can go ahead and compare that to CPUs like Pentium 4 or Cell which are just down right horrible.