ROCm / ROC-smi

ROC System Management Interface
https://github.com/RadeonOpenCompute/ROC-smi/blob/master/README.md
179 stars 55 forks source link

Option to show current voltage #58

Closed SandboChang closed 5 years ago

SandboChang commented 5 years ago

Hi,

With the command

watch rocm-smi

I can see the temp/clkfreq/power consumption and so on, but it seems there isn't a way to see the current voltage yet. Recently I am trying to undervolt my Vega, --setslevel SCLKLEVEL SCLK SVOLT option seems to successfully set the P6/P7 to a reduced voltage and I could see a temperature drop. But it would be nice to see if the voltage is sitting at an expected level.

Would it be available at some point?

jlgreathouse commented 5 years ago

I'll leave the feature request for rocm-smi to @kentrussell . However, I will note that you can read the voltage of your AMD GPU (in millivolts) by reading the sysfs file /sys/class/hwmon/hwmonN/in0_input. N, in this case, can change depending on your GPU's number.

So if you have a single GPU, for instance, you could (for now) run the following to supplement the output of rocm-smi with a reading of the voltage: watch "rocm-smi; cat /sys/class/hwmon/hwmon0/in0_input"

kentrussell commented 5 years ago

As always, I'll say it here: "watch rocm-smi" is a terrible thing to do. It's designed to be a user-friendly interface, not as an efficient system monitor (see https://github.com/RadeonOpenCompute/ROC-smi/issues/38 ) . I'd recommend only watching the fields that you want (use "watch rocm-smi -P -t -c" to only watch Power/Temperature/Clocks) . I'll see if we can get voltage into 2.3 .

SandboChang commented 5 years ago

Interesting, as I actually tried to do some crazy by doing watch -n 0.1 rocm-smi, which from #38 should have run into error, but it actually works very well even with the three cards I have.

Yes, at the end I probably don’t need to watch everything, but the capability of watching voltage as I tune it will be great.

kentrussell commented 5 years ago

Also, I am adding a --showvoltage flag that will hopefully make it into 2.3 .

kentrussell commented 5 years ago

In 2.3 release