Well, most of our results are not 17x. But still (IMO) solid across the board!
Also, the 17x came from a pretty obscure fusion optimization that isn't called out anywhere in the documentation (we had to run the profiler to see what was actually going on). Wouldn't be surprised if whoever within AWS wrote the kernel didn't know about that optimization.
Also, the 17x came from a pretty obscure fusion optimization that isn't called out anywhere in the documentation (we had to run the profiler to see what was actually going on). Wouldn't be surprised if whoever within AWS wrote the kernel didn't know about that optimization.