Forums

expectations from a DFE ?

Started by mk July 25, 2006
Hi,
I've been experimenting with adaptive filters for channel equalization
and I have a problem with decision feedback case. Initially I started
with a T/2 fractional linear equalizer adapted with LMS and then
converted this to delayed LMS with two delays. I am currently not
adding any noise just trying to compensate for the ISI generated by
the channel although there is quantization error as the input is
quantized with an ADC model (albeit a perfect one, no non-linearities
yet). Finally I added a feedback filter to my implementation. The
issue is that I am only seeing 1 dB MSE enhancement with the DFE in a
serial equalizer (ie serial LE vs serial DFE) and 0.6dB for one cycle
delayed LE vs DFE. The drop in enhancement for the delayed version I
understand because it can no longer cancel the first postcursor ISI
but I am not sure how to quantify whether 1dB enhancement is what I
should be getting in the serial case. Any ideas on how to work out
what a DFE should give me in this case?

Thanks.
"mk" <kal*@dspia.*comdelete> wrote in message
news:r55bc2th5chu8251mu8t4g2rqnfrn92opm@4ax.com...
> Hi, > I've been experimenting with adaptive filters for channel equalization > and I have a problem with decision feedback case. Initially I started > with a T/2 fractional linear equalizer adapted with LMS and then > converted this to delayed LMS with two delays. I am currently not > adding any noise just trying to compensate for the ISI generated by > the channel although there is quantization error as the input is > quantized with an ADC model (albeit a perfect one, no non-linearities > yet). Finally I added a feedback filter to my implementation. The > issue is that I am only seeing 1 dB MSE enhancement with the DFE in a > serial equalizer (ie serial LE vs serial DFE) and 0.6dB for one cycle > delayed LE vs DFE. The drop in enhancement for the delayed version I > understand because it can no longer cancel the first postcursor ISI > but I am not sure how to quantify whether 1dB enhancement is what I > should be getting in the serial case. Any ideas on how to work out > what a DFE should give me in this case? > > Thanks.
I'm not familiar with the term "serial" dfe? How does that differ from a regular dfe? Also, are you saying the LE has 1 dB better performance than DFE, i.e. the constellation quality/average euclidean distance to the ideal symbol is 1 db better? I would expect this since I think DFE only cancels post-cursor ISI (i.e. delays of main signal) so the pre-cursor ISI created by the symbol filter wouldn't be compensated for. Were you running LE and DFE at the same time or just one or the other. -Clark
On Wed, 26 Jul 2006 02:03:58 GMT, "Anonymous" <someone@microsoft.com>
wrote:

> >"mk" <kal*@dspia.*comdelete> wrote in message >news:r55bc2th5chu8251mu8t4g2rqnfrn92opm@4ax.com... >> Hi, >> I've been experimenting with adaptive filters for channel equalization >> and I have a problem with decision feedback case. Initially I started >> with a T/2 fractional linear equalizer adapted with LMS and then >> converted this to delayed LMS with two delays. I am currently not >> adding any noise just trying to compensate for the ISI generated by >> the channel although there is quantization error as the input is >> quantized with an ADC model (albeit a perfect one, no non-linearities >> yet). Finally I added a feedback filter to my implementation. The >> issue is that I am only seeing 1 dB MSE enhancement with the DFE in a >> serial equalizer (ie serial LE vs serial DFE) and 0.6dB for one cycle >> delayed LE vs DFE. The drop in enhancement for the delayed version I >> understand because it can no longer cancel the first postcursor ISI >> but I am not sure how to quantify whether 1dB enhancement is what I >> should be getting in the serial case. Any ideas on how to work out >> what a DFE should give me in this case? >> >> Thanks. > >I'm not familiar with the term "serial" dfe? How does that differ from a >regular dfe? >
I am experimenting with pipelined DFEs so I used serial as a replacement for non-pipelined which is probably what you mean by regular DFE.
>Also, are you saying the LE has 1 dB better performance than DFE, i.e. the >constellation quality/average euclidean distance to the ideal symbol is 1 db >better? I would expect this since I think DFE only cancels post-cursor ISI >(i.e. delays of main signal) so the pre-cursor ISI created by the symbol >filter wouldn't be compensated for.
Actually it's the reverse. My DFE has a feedforward section and a feedback section and the LE has only the feedforward section which is the same size as that in DFE.
>Were you running LE and DFE at the same time or just one or the other.
Assuming you use the DFE only as the feedback portion of the equalizer, I am running LE vs LE+DFE but I think my people use the term DFE to include both FF & FB filter portions.
> >Were you running LE and DFE at the same time or just one or the other. > > Assuming you use the DFE only as the feedback portion of the > equalizer, I am running LE vs LE+DFE but I think my people use the > term DFE to include both FF & FB filter portions.
I see, so basically when you turn on the feedback section you're only getting 1 dB improvement and you were expecting better? This would agree with my experience on the system I worked on. I think the feedback section only helps substantially on particular channel responses. I got the impression that the feedback section worked best against basically an IIR like channel response, i.e. you only need a couple taps to generate an infiniite (or very long) impulse response so only a few FB taps are needed to cancel the same. So my guess is your simulation is fine, it's just that the signal you are feeding it doesn't have enough ISI to need the FB section, i.e. the FF taps are doing well enough that the FB taps have little to contribute. -Clark