14:13 <tomeu> lynxeye1: something I have observed is that by tuning the clock speed to the minimum and feeding the HW a big enough job, I can observe how, for NN jobs, the SH is busy for the duration of the job, from the execution of the INST_ADDR state until the irq is fired

14:13 <tomeu> and for TP jobs, the SH remains busy because we don't get an IRQ (or the other way around)

14:13 <tomeu> so I guess NN and TP jobs use indeed some resource in the SH unit, and in the TP case the operation goes wrong and either it hangs, or just doesn't finish

14:14 <tomeu> wonder if the SH unit isn't used for feeding the data to the systolic array, and retrieving data from it