bad apple, but actually running on a trainium chip

this is a real profiling trace of the insts and DMAs in a kernel running on a neuroncore. one engine emits a steady stream of insts that other engines use like a clock. this kernel “executes” all frames, one after another

i work on a team that cares deeply about understanding chips to maximize performance. some of that work enables this precise drawing via insts running on the chip :)

that perfetto layout looks awfully similar to the VLIW takehome 🤔🤔🤔🤔🤔🤔

Hell yeah this rules

wow

this is good content

peak

O_o

10/10 no notes

I find this too funny