Power measurement on Metis

Hi ​@Aimar.gnz! 

I think axmonitor might be what you’re looking for. It's Voyager’s system monitoring tool that can show core temperature, utilisation, and on supported hardware, power metrics as well.

There's a tutorial in the SDK docs here: https://github.com/axelera-ai-hub/voyager-sdk/blob/HEAD/docs/tutorials/axmonitor.md

Let me know how you get on!

Hi ​@Spanner , 

The thing is power metrics are only available on 4-Metis PCIe cards, and I only have one, so I was wondering whether there was an alternative way.
I am using a smart plug that gives actual power metrics, but my AIPU is not recognised in the Metis compute board.

Error is:
 The Metis card is visible on PCIe, but during boot the driver reads an invalid firmware header:
      - Invalid hdrv magic 0
      - Fail to get hdrv area

Really stuck with this.

Hello ​@Aimar.gnz ,

Regarding the messages that you see (- Invalid hdrv magic 0 - Fail to get hdrv area):

These messages are not an issue and the driver and Metis card are working as expected. 

We are aware that these messages can be confusing to the user, as they do not mean there are any issues. Therefore, I have raised this internally and we will remove these traces in next releases.

To sum up, do not worry about those messages, everything is working well.

Best,

Victor

Hi, we have been trying the new v1.6 of the voyager SDK.
We are currently using PCIe version of the accelerator, and theres no data of power consumption on the axmonitor.

Is this normal? Does this only works for m2 version?

Kindest regards

Hi ​@rozero ,

Likely the other way round, if I recall. AxMonitor's power readings need on-board power telemetry, which the 1-chip PCIe Metis doesn't expose. The 4-chip PCIe card does, and the new M.2 Max will.

So if you've got the 1-chip PCIe, no power data is likely expected behaviour. If it's the 4-chip and you're still seeing nothing, that's worth digging into. 👍

• ​@Aimar.gnz ,
  
  I have been exploring how to measure the power for M.2 modules which do not have on-board telemetry enabled.  I have documented a solution that works on my personal web site (blog):
  
  - https://mariobergeron.com/posts/edge-ai-power-p03-ina228/
  
  I hope you find this useful !
  
  Note that this article currently only mentions Hailo-8 and compares results with Hailo’s on-board telemetry, as a way of validating the methodology.  However, it also works with Axelera Metis.
  
  I have not yet published an article specific to Axelera Metis, since I have not been able to reproduce the ResNet-50 benchmark of 1756 FPS yet … work in progress … 
  
  Cheers !

Thank you for replying Spanner.
Just to fully confirm, this PCIe accelerator is a 1-chip then?
 

I may have just confused AI Cores with Chips :/
Best

That’s right, yep. The Metis PCIe has one chip, with four cores.

Although there is also the four chip PCIe card, which, in total, has 16 cores. If that makes sense? 😄 It’s like four of the single chip cards all bundled together.

Ty Spanner!.