Liabe Brave
Member
Okay, understood. And your point about rounding is quite true.FYI, the difference in the ratio between 204/192 and 218/204 is pretty small (much smaller than 0.4%). This mysterious overhead is likely consistently applied to both numbers, but the rounding to GB/s gives some error.
The point I was trying to make about read/write on the same cycle was that there should be no reason for a turnaround time because if you can read/write at the same time, then there shouldn't be a reason to turn around to begin with for calculating maximum theoretical throughput. You just read/write every cycle.
I don't think that affects my main point, though, which is that it doesn't look to me like Microsoft is lying or fudging the numbers. They seem to be calculated the same way before and after a clock change, and in my estimation it's very unlikely Microsoft would flat-out present voodoo numbers at Hotchips, with all the egg that might end up on their faces.
It looks to me like they're honestly presenting the true theoretical speed of the eSRAM, at 204 GB/s.