[Jetson-TK1] GPU (gk20a) driver and power management

Hello,

Does anyone have any idea what these files do (gk20a driver sysfs files):

root@jetson:/sys/devices/platform/host1x/gk20a.0# ls
blcg_enable
elcg_enable
slcg_enable
clockgate_delay
elpg_enable
railgate_delay

For example, the last file keeps the power rail active for x milliseconds after the last GPU activity.

Also, does anyone have any knowledge of how the GK20A (or GPUs in general) behave in terms of power usage? Are NVIDIA’s drivers doing some sort of aggressive power management? The reason I ask is that I am building a power model for the Tegra K1, but the resulting model has very poor accuracy.

For those interested: I try to model the GPU according to the (CPU-specific) formulae:

Power = aCV^2f + V*I_{leak}, where aC is a workload specific variable, V is the GPU voltage, f is the GPU frequency, and I_{leak} is the leakage current of the GPU.

As with most modern mobile processors, the Tegra K1 has sophisticated power management. One that may be a harder part to model is the Dynamic Voltage and Scaling (DVFS) subsystem.

There was a good talk “Evaluation of the Jetson TK1 Development Board for Power and Performance” by Jee Choi at GTC 2015. You can find the talk here:

http://on-demand.gputechconf.com/gtc/2015/video/S5407.html

and the slides for the talk here:

http://on-demand.gputechconf.com/gtc/2015/presentation/S5407-Jee-Choi.pdf

Obviously it depends on what you’re trying to measure. Low power applications can run in about 3-5 watts on the board, flat out it can draw 17. DVFS allows the GPU clock to run from 72MHz ~ 852 MHz with 15 steps in between. GPU Memory 12.75MHz ~ 924 MHz in 12 steps. So you can model a particular group of settings by writing to the appropriate files.

Hi Kangalow. Thanks for the links, this is really interesting. I have actually managed to model accurately the dynamic / static power of the CPU clusters. However the methodology I use for modelling the CPU does not work for the GPU…

Will have a look at those slides thank you :)

Hello!

I’m doing some research on TegraK1 GPU. It would be very nice if someone has Mr. Jee Choi presentation pdf and/or video. The links are broken now. There is any chance to find this presentation?

There is a way to find Tegra K1 power management architecture description?

Thank you very much in advance!

The link to the PDF appears to be dead, but the video still works, at least from my computer and location.

http://on-demand.gputechconf.com/gtc/2015/video/S5407.html

Thank you! The video worked for me in the end on Ubuntu after installing flash player.

Hi,

I am also trying to analyze some workloads for power on tx1. I have a question on tegra dvfs. I know we can set the gpu frequencies, cpu frequencies and memory frequencies using commands. My question is how to put back everything in default mode i.e. where the device frequencies are selected automatically at runtime.?