| Author |
Message |
Ron N.
Guest
|
 Posted:
Wed Dec 07, 2005 1:16 am Post subject:
Re: symmetric transpose form fir filter |
|
|
abariska@student.ethz.ch wrote:
| Quote: | Jerry Avins wrote:
abariska@student.ethz.ch wrote:
I'm not talking about chosing IIR over FIR, as this would be part of
the spec. Or perhaps the real art is to bend the specs to achieve
minimum power consumption?
Implementation specs are second-rate at best. The spec should be on
performance.
I don't want to argue about the definition of "specs", but about power
consumption. Even if you could choose between filter types in an
implemenation, most certainly that type which is faster to compute
also consumes less power.
|
This certainly depends on the system implementation. On many
circuits and systems the relative number of cycles taken by code
sequences and relative mAH of energy consumed by the circuit
executing those code sequences are not identical. IC designers
spend a lot of time on power optimization these days, but some
functional units almost intrinsically require more power to cycle
than others in the same technology.
IMHO. YMMV.
--
rhn A.T nicholson d.O.t C-o-M |
|
| Back to top |
|
 |
Steve Underwood
Guest
|
| Posted:
Wed Dec 07, 2005 8:08 am Post subject:
Re: symmetric transpose form fir filter |
|
|
abariska@student.ethz.ch wrote:
| Quote: | Jerry Avins wrote:
abariska@student.ethz.ch wrote:
...
I'm not talking about chosing IIR over FIR, as this would be part of
the spec. Or perhaps the real art is to bend the specs to achieve
minimum power consumption?
Implementation specs are second-rate at best. The spec should be on
performance.
I don't want to argue about the definition of "specs", but about power
consumption. Even if you could choose between filter types in an
implemenation, most certainly that type which is faster to compute also
consumes less power.
Regards,
Andor
Simplifying the algorithm is an obvious starting point, but simplifying |
for power is not always the same as simplifying for speed. A few
examples of things that people seldom look at, unless they have power
problems and someone kicks them:
- Do you look carefully at power management issues when choosing
devices, and do you really understand the tradeoffs when choosing? For
example, a super low power mode that takes an age to wake up from (many
do) may be completely useless, if your application is bursty in nature.
Wise choices in this area can have a big effect on average battery drain.
- You have a fairly slow requirement. Your processor has DMA
capabilities. Are you using DMA whenever possible? As well as speeding
things up, DMA takes less power than moving thing through software. In
extreme cases you might be able to keep the CPU asleep while some data
block DMAs in or out at a leisurely pace with near zero consumption. If
the device you are using has special processing blocks, like a viterbi
processor, can DMA be used to feed data in and out of those blocks? If
so, it will probably be a lower power method than using software. DMA
can be a huge power saver.
- You might have reduced the algorithm to the minimum number of
operations, but is it using the least power hungry parts of the device.
Using a shift instead of an add, for example, can be quite beneficial
(provided the architecture doesn't make you jump through hoops to get
the data in the right place). Quite often, if you shuffle the scaling of
some coefficients around, a few multiplies can be reduced to shifts.
Anyone who has done bit serial DSP should be a wizz at this. :-)
Too often people just put their code together, run it, measure current,
and say that is what is achievable. If they do look for improvement,
they tend to attack the algoroithmic complexity. Looking at the
algorithmic fit to the qualities of the device being used often offers
bigger potential for improvement.
Regards,
Steve |
|
| Back to top |
|
|
Rick Lyons
Guest
|
| Posted:
Thu Dec 08, 2005 8:29 am Post subject:
Re: symmetric transpose form fir filter |
|
|
On Tue, 06 Dec 2005 09:14:48 +0800, Steve Underwood <steveu@dis.org>
wrote:
| Quote: | Rick Lyons wrote:
(snipped) |
| Quote: |
Ron, you are quite rignt. During the execution of MAC instructions the
DSP power consumption is several times higher then when executing a
general program flow. This is the fact; I observed it with ADSP-21xx.
Vladimir Vassilevsky
Hi Vladimir,
Now that is interesting,... very interesting.
Thanks,
[-Rick-]
Not only interesting. It should be fairly obvious. Its interesting to
spot the people in this group that never work on portable equipment.
Lines like "MACs are really fast these days, so we can use an FIR
instead of an IIR" are a dead giveaway. :-) The software in portable
devices very much affects their battery life, right down to choosing the
least complex instructions. Designing software for the absolute minimum
power consumption is an interesting art.
Regards,
Steve
|
Hi Steve,
yep, you're right. The systems I worked on were in the
range of 50 to 20,000 watts. The power to my systems
was limited only by the size of the circuit breakers
in the laboratory's electric panel.
I've never built any battery-powered systems for a customer.
I know that the more transistors you have that
are switching states (on-to-off & off-to-on) the more
power is consumed. But it never occurred to me
to think about the difference in power consumption
in performing a single addition versus a single multiply.
What would be interesting to me is if Vladimir could tell
us the following: if continous binary adds require P watts
in a DSP chip, do continuous multiplies requre 2P, 4P,
8P watts, or what.
See Ya'
[-Rick-] |
|
| Back to top |
|
|
Randy Yates
Guest
|
| Posted:
Thu Dec 08, 2005 8:47 am Post subject:
Re: symmetric transpose form fir filter |
|
|
R.Lyons@_BOGUS_ieee.org (Rick Lyons) writes:
| Quote: | On Tue, 06 Dec 2005 09:14:48 +0800, Steve Underwood <steveu@dis.org
wrote:
Rick Lyons wrote:
(snipped)
Ron, you are quite rignt. During the execution of MAC instructions the
DSP power consumption is several times higher then when executing a
general program flow. This is the fact; I observed it with ADSP-21xx.
Vladimir Vassilevsky
Hi Vladimir,
Now that is interesting,... very interesting.
Thanks,
[-Rick-]
Not only interesting. It should be fairly obvious. Its interesting to
spot the people in this group that never work on portable equipment.
Lines like "MACs are really fast these days, so we can use an FIR
instead of an IIR" are a dead giveaway. :-) The software in portable
devices very much affects their battery life, right down to choosing the
least complex instructions. Designing software for the absolute minimum
power consumption is an interesting art.
Regards,
Steve
Hi Steve,
yep, you're right. The systems I worked on were in the
range of 50 to 20,000 watts. The power to my systems
was limited only by the size of the circuit breakers
in the laboratory's electric panel.
I've never built any battery-powered systems for a customer.
I know that the more transistors you have that
are switching states (on-to-off & off-to-on) the more
power is consumed. But it never occurred to me
to think about the difference in power consumption
in performing a single addition versus a single multiply.
What would be interesting to me is if Vladimir could tell
us the following: if continous binary adds require P watts
in a DSP chip, do continuous multiplies requre 2P, 4P,
8P watts, or what.
|
Hi Rick,
TI has a spreadsheet for this for the DM64x DSP, so you can play with
the numbers there:
http://focus.ti.com/docs/apps/catalog/resources/appnoteabstract.jhtml?abstractName=spra962f
The spreadsheet inputs (for the CPU) are percent CPU utilization and
clock frequency, so you can't really compare adds to MACs, for
example. They also use some sort of non-linear relationship (as a
function of percent utilization), but it's close to linear.
Power consumption estimation on a chip like this, which is really a
group of complex subsystems, and in which the subsystems interact,
is no easy task.
--
% Randy Yates % "Bird, on the wing,
%% Fuquay-Varina, NC % goes floating by
%%% 919-577-9882 % but there's a teardrop in his eye..."
%%%% <yates@ieee.org> % 'One Summer Dream', *Face The Music*, ELO
http://home.earthlink.net/~yatescr |
|
| Back to top |
|
|
Rick Lyons
Guest
|
| Posted:
Fri Dec 09, 2005 7:39 am Post subject:
Re: symmetric transpose form fir filter |
|
|
On Thu, 08 Dec 2005 02:47:53 GMT, Randy Yates <yates@ieee.org> wrote:
(snipped)
| Quote: |
What would be interesting to me is if Vladimir could tell
us the following: if continous binary adds require P watts
in a DSP chip, do continuous multiplies requre 2P, 4P,
8P watts, or what.
Hi Rick,
TI has a spreadsheet for this for the DM64x DSP, so you can play with
the numbers there:
http://focus.ti.com/docs/apps/catalog/resources/appnoteabstract.jhtml?abstractName=spra962f
The spreadsheet inputs (for the CPU) are percent CPU utilization and
clock frequency, so you can't really compare adds to MACs, for
example. They also use some sort of non-linear relationship (as a
function of percent utilization), but it's close to linear.
Power consumption estimation on a chip like this, which is really a
group of complex subsystems, and in which the subsystems interact,
is no easy task.
--
% Randy Yates % "Bird, on the wing,
|
Hi Randy,
Thanks, I'll have a look.
[-Rick-] |
|
| Back to top |
|
|
Jerry Avins
Guest
|
| Posted:
Fri Dec 09, 2005 8:43 am Post subject:
Re: symmetric transpose form fir filter |
|
|
Rick Lyons wrote:
| Quote: | On Thu, 08 Dec 2005 02:47:53 GMT, Randy Yates <yates@ieee.org> wrote:
(snipped)
What would be interesting to me is if Vladimir could tell
us the following: if continous binary adds require P watts
in a DSP chip, do continuous multiplies requre 2P, 4P,
8P watts, or what.
Hi Rick,
TI has a spreadsheet for this for the DM64x DSP, so you can play with
the numbers there:
http://focus.ti.com/docs/apps/catalog/resources/appnoteabstract.jhtml?abstractName=spra962f
The spreadsheet inputs (for the CPU) are percent CPU utilization and
clock frequency, so you can't really compare adds to MACs, for
example. They also use some sort of non-linear relationship (as a
function of percent utilization), but it's close to linear.
Power consumption estimation on a chip like this, which is really a
group of complex subsystems, and in which the subsystems interact,
is no easy task.
--
% Randy Yates % "Bird, on the wing,
Hi Randy,
Thanks, I'll have a look.
|
I've heard of using the power vs. time curve as clues to reverse
engineering. Good encryption algorithms maintain a "power drain poker
face" while doing there thing. It's too easy to break the code if they
don't.
Jerry
--
Engineering is the art of making what you want from things you can get.
ŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻ |
|
| Back to top |
|
|
Jerry Avins
Guest
|
| Posted:
Fri Dec 09, 2005 9:16 am Post subject:
Re: symmetric transpose form fir filter |
|
|
Jerry Avins wrote:
| Quote: | Rick Lyons wrote:
On Thu, 08 Dec 2005 02:47:53 GMT, Randy Yates <yates@ieee.org> wrote:
(snipped)
What would be interesting to me is if Vladimir could tell us the
following: if continous binary adds require P watts
in a DSP chip, do continuous multiplies requre 2P, 4P, 8P watts, or
what.
Hi Rick,
TI has a spreadsheet for this for the DM64x DSP, so you can play with
the numbers there:
http://focus.ti.com/docs/apps/catalog/resources/appnoteabstract.jhtml?abstractName=spra962f
The spreadsheet inputs (for the CPU) are percent CPU utilization and
clock frequency, so you can't really compare adds to MACs, for
example. They also use some sort of non-linear relationship (as a
function of percent utilization), but it's close to linear.
Power consumption estimation on a chip like this, which is really a
group of complex subsystems, and in which the subsystems interact,
is no easy task.
--
% Randy Yates % "Bird, on the wing,
Hi Randy,
Thanks, I'll have a look.
I've heard of using the power vs. time curve as clues to reverse
engineering. Good encryption algorithms maintain a "power drain poker
face" while doing there thing. It's too easy to break the code if they
don't.
Jerry
|
their thing, dammit!
--
Engineering is the art of making what you want from things you can get.
ŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻŻ |
|
| Back to top |
|
|
Randy Yates
Guest
|
| Posted:
Fri Dec 09, 2005 9:16 am Post subject:
Re: symmetric transpose form fir filter |
|
|
Steve Underwood <steveu@dis.org> writes in some detailed about interesting power reduction
techniques:
We never tweaked at this level in our handset code at Ericsson/Sony Ericsson, but thanks
for the ideas, Steve. I guess if you were really going whole-hog for the lowest power
they could be very important.
Can you tell us what applications you have applied these techniques to?
--
% Randy Yates % "Rollin' and riding and slippin' and
%% Fuquay-Varina, NC % sliding, it's magic."
%%% 919-577-9882 %
%%%% <yates@ieee.org> % 'Living' Thing', *A New World Record*, ELO
http://home.earthlink.net/~yatescr |
|
| Back to top |
|
|
|
|
|
|
FAQ
Memberlist
Register
Profile
Log in to check your private messages
Log in