Strange power numbers w/ ASAP7

[evmanz]

Hello everyone!

I am running OpenRoad flow for my design and I see some strange numbers for reported power.

Here is a snippet of two cells on the timing path:

      Cap      Slew     Delay      Time              Description
-------------------------------------------------------------------------------
                                          _09534_ (net)
            39.1912    0.5733 2766.8992 ^ _35759_/A (XNOR2x1_ASAP7_75t_R)
   1.6660   23.6009   27.2220 2794.1211 ^ _35759_/Y (XNOR2x1_ASAP7_75t_R)
                                          _09851_ (net)
            23.6031    0.1195 2794.2407 ^ _35823_/A (XNOR2x1_ASAP7_75t_R)
   0.9752   15.4225   22.0917 2816.3323 ^ _35823_/Y (XNOR2x1_ASAP7_75t_R)
                                          _09914_ (net)

They are both the same type, both with reasonable input transition and output capacitance.
However, both have insane power consumption (more than 1mW each!):

   Internal  Switching    Leakage      Total
      Power      Power      Power      Power (Watts)
--------------------------------------------
   9.74e-04   4.72e-04   1.67e-10   1.45e-03 _35823_
   6.39e-04   5.19e-04   1.67e-10   1.16e-03 _35759_

What is strange is that the second cell in the timing path (35823) consumes more power than its driver (35759) while having better input transition & output capacitance.

I checked .lib[1], and for those operating conditions the cells should consume about 0.24 - 0.5 pW (depending on rise or fall transition, and the state of other pins), which is VERY far from the reported values.

As a result, I have designs with 100k cells, which consume 2kW of power, which makes zero sense.
The designs have small WNS (say -10ps), and TNS at most -100ps after PnR. Length of the timing path above is roughly 100 cells.

What do I miss?

[maliberty]

Can you provide a test case?

[evmanz]

@oharboe

Thank you. I used your instructions, rebuilt everything, and successfully loaded the db :)

…and I see the same issue.

For example, we have two XNOR2x2_ASAP7_75t_R cells 33728 and 24028 with 4 orders of magnitude difference in power consumption:

openroad> report_power -instance _33728_
   Internal  Switching    Leakage      Total
      Power      Power      Power      Power (Watts)
--------------------------------------------
   1.82e-03   2.32e-03   1.71e-10   4.14e-03 _33728_

and

openroad> report_power -instance _24028_
   Internal  Switching    Leakage      Total
      Power      Power      Power      Power (Watts)
--------------------------------------------
   3.06e-07   4.18e-07   1.73e-10   7.24e-07 _24028_

While both instances have similar caps and transition times XNOR2x2_ASAP7_75t_R

  Cap    Slew   Delay    Time   Description
-----------------------------------------------------------------------
  4.35   18.99   38.21  389.27 v _24028_/Y (XNOR2x2_ASAP7_75t_R)

and

    Cap    Slew   Delay    Time   Description
-----------------------------------------------------------------------
   4.06   18.19   36.55 2356.02 v _33728_/Y (XNOR2x2_ASAP7_75t_R)

Why does it happen?

[jjcherry56]

Most likely because the activities are vastly different (see get_property pin activity] which shows up in the switching power. If you post a coherent testcase (ie, with make issue) I will look at it, but I lack @oharboe's patience tinker to make your testcase run.

[evmanz]

Most likely because the activities are vastly different (see get_property pin activity] which shows up in the switching power.

To get mW power on a single cell?

If you post a coherent testcase (ie, with make issue) I will look at it, but I lack @oharboe's patience tinker to make your testcase run.

There is no make issue in the environment in OpenRoad-flow-scripts

You can run my test case by:

1. unzipping the archive
2. make run

it will build the OpenRoad flow docker with all the toolchains (which are all parts of the default OpenRoad-flow-scripts flow), and then will run PnR on the design.

EDIT:

I made an archive with make detail_route_issue: Here is the archive

[evmanz]

see get_property pin activity

This is what I get for the cell with 0.17 mW power.

get_property -object pin  _34197_/Y activity
2.30819e+11 0.500 propagated

Is this number the activity factor -> 2.30819e+11?

[evmanz]

Hi @jjcherry56

Did you have a chance to look into the power calculation?

[oharboe]

@evmanz I'm curious: did you use groups to be able to get multiple histograms or do they have some other purpose?

[evmanz]

@oharboe

I'm curious: did you use groups to be able to get multiple histograms or do they have some other purpose?

I am mostly used Synopsys tools where defining more specific path groups helps with timing closure during PnR (separation of concerns of sorts). I assumed the same happens with OpenRoad as well.

[maliberty]

I've been chasing a similar issue that resulted in parallaxsw/OpenSTA#160. I'm going to wait until that is resolved to see if it fixes this as well.

[evmanz]

Thank you for the update Matt.

I think this is it. I can't think of anything else on top of my head that can mess power beyond activity factors.