On Tue, Oct 16, 2012 at 8:39 AM, Jeremy Bennett
<[email protected]> wrote:
> It's not truly FLOSS (restriction on commercial use), but Wattch source
> code is available for non-commercial/non-profit use.
>
>         http://www.eecs.harvard.edu/~dbrooks/wattch-form.html
>
> Another case of someone writing their own "open source" license, rather
> than using a tried and tested one!
>
> This is probably the most widely used power estimation system for
> mid/high-level analysis, certainly within the academic community. Its
> proponents claim it is accurate to 10%.

Just my two cents:

I've done lots of work with sim-wattch, and I don't trust its results.
 Modern architectures have deviated somewhat from what it simulates,
and it simulates at a very abstract level.  The way it maps the
abstract out-of-order simulator back onto actual hardware components
is what I don't trust.  It's easy to cheat or make mistakes in your
power estimation if you make changes to the simulator.  Plus, it uses
an 8-byte RISC-like instruction format, combined with all kinds of
hackery to convert the results from using an 8-byte instruction to
what you would get with a 4-byte instruction.

On top of all that, it's a tremendous pain to work with, because it
relies on a 10+ year old version of GCC that often generates incorrect
code.  You'll spend a lot of time just trying to figure out whether
the bugs are in the compiler or in the simulator.

Finally, it simulates an out-of-order core, which makes it unsuitable
for simulating the power usage of the OR1200.  You -can- tell the
simulator to issue in-order, but it still simulates all the structures
that an OoO core would have, so your power estimate won't be
realistic.

> Accurate power estimation by simulation is provided by commercial tools
> operating at the gate level. It is critically tied into knowledge of the
> target silicon process. These tools run extremely slowly (seconds per
> cycle, rather than cycles per second for big chips), but apparently give
> very good estimates.

This is the way to go.  It's slow, and lots of work, but I don't
believe anyone would sincerely question the estimates you'd get this
way, provided you simulate for a real process, and use an up-to-date
simulator.

Unfortunately, there is no way to do this with open source tools, and
probably will never be.

-Pete
_______________________________________________
OpenRISC mailing list
[email protected]
http://lists.openrisc.net/listinfo/openrisc

Reply via email to