bowman.brian

Not sure that they just have bags of it laying around waiting for tests.

Most GPUs already do some sort of “inexact” processing by using special computational units that saturate (can’t overflow) and by taking other shortcuts with difficult computations like inverse and square root.

If I had to take a guess, the easiest way to create this “inexact” chip would be to take a normal pipe-lined processor and throw out the hazard detection and stall logic necessary to ensure data coherency. This would ensure you never have to delay to allow previous instructions to complete and never flush the pipeline due to incorrect branch prediction or other ideas. Fewer wasted cycles = more efficient. Not sure what they are getting at with “pruning” … Some modern processors turn off components that are unneeded for awhile (like a floating point unit) but this one suggests they remove some things altogether. Maybe getting rid of the stuff I mentioned counts as pruning but this seems like a kindergarten approach to comp. arch (simply ignoring anything difficult and claiming performance increases).

Good job helping out constructively.

If we do indeed look at the data this post was about we see that the dividing line down the middle separates about 32 models from the rest. This is 50% market share. The article mentions almost 4000 device models reporting so that is ……………

0.8% I think this number is still worth something. That less than a percent of android phones hold 50% of the android market.

Maybe the poster above just rounded and messed up the decimal. The overarching point is still salient.