We would like to lower the power power consumption of a Jetson TX2-based embedded product (in Max-P mode) via a software update. We have verified that we can lower the GPU clock similar to the Max-Q profile. We are not confident (without further testing) that we can lower CPU and memory clocks to the values of the Max-Q profile and still maintain the required throughput.
Our goal is to limit instantaneous peak power consumption, and if possible, reduce the variability in power consumption.
We are currently using the GPU and four of the ARM cores with disabled Denver cores, and fixed clocks for GPU, CPU and memory.
We would like to stay below 10 W peak total power consumption independent of GPU workload.
What options do we have?
Is it possible/beneficial to set different clock speeds to different CPU cores? We could conceivably use three instead of four cores. Is it possible to reduce power consumption by enabling some Denver cores and disabling ARM cores?
Are there tables that show the power consumption across the space between Max-Q and Max-P for the various clocks?
We are not sure which settings will give us the biggest gains (apart from obviously lowering GPU clocks).
We also see a huge variability in the peak power consumption when running various CUDA kernels and are not sure what causes that.
Our software for the relevant product is based on L4T 28.x, not 32.x if that matters.