cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 
699 Views
Registered: ‎05-22-2018

What is the dependency among lane data rate, PRBS pattern, TX DIFF swing and BER for IBERT test on Kintex Ultrascale?

Jump to solution

Hi everyone,

 

I am working with kintex ultrascale development board and vivado v2018.2. I went through the 14th and 15th chapters in UG908 document, pg152 document and created an example project to test the serial I/O links (I have the hardware setup for 2 links) using PRBS 7-bit, 9-bit and 31-bit patterns (IBERT test). Following are my settings for the IBERT core.

Line rate: 12.5 Gbps, data width: 40, refclk: 625MHz.

Protocol selection: custom 1/12.5 Gbps

sysclk frequency 250MHz

 

However I see dependency of BER based on the PRBS pattern selected for both TX and RX patterns. Following are my observations.

1) For PRBS 7-bit pattern for both TX and RX at the default TX DIFF swing of 950mV, I observed that the BER was in the order of E-12.

2) When I change to PRBS 9-bit, at the default TX DIFF swing of 950 mV, after performing the BERT reset which is necessary, the BER jumped to the orders of E-2 and even E-1. When the TX DIFF swing was reduced to 530mV and 250mV and BERT reset, the BER again went into the orders of E-11. 

3) I proceeded to using PRBS 31-bit pattern. In a document I came across, related to the development board I am using, it was told that the MGTs for that board do not support 12.5 Gbps for PRBS-31bit pattern. Also the guidelines to test out the setup with PRBS 31-bit pattern was such that the TX DIFF swing value was different for different links (950mV and 950mV were changed to 530mV and 250mV)

4) When I make changes to the BERT core to support 11.25 Gbps to test out PRBS 31-bit pattern, I see when I auto detect links, the DFE is enabled which was not the case for PRBS 7-bit and 9-bit patterns.

 

From the above observations, my queries regarding the experiment are as follows

1) Why is it that the BER is dependant on the PRBS bit pattern used? Increasing the randomness of the 1's and 0's sent over the serial link should not generally affect the BER right?

2) How does decreasing the voltage swing improving the BER? From what I know, higher swing should actually prevent bit errors?

3) What is the meaning of the MGT not supporting 12.5 Gbps for PRBS 31-bit pattern? What is the dependency of the PRBS pattern used for the test, on the line rate?

4) Also, two different TX DIFF swing values (530mV and 250mV) were needed to be given to the two serial I/O links being used to obtain the same BER order (E-11). Why was using two different values required here?

5) Why is the DFE being auto-enabled when using the PRBS 31-bit pattern?

 

Can someone suggest answers for any or all of the queries above? Also if there has been any misunderstanding in any of my interpretations above, please do suggest the correct interpretations too.

 

Thanks in advance,

 

-Chan100

0 Kudos
Reply
1 Solution

Accepted Solutions
Moderator
Moderator
442 Views
Registered: ‎07-30-2007

 

 1) At 11.5Gbps PRBS-31 will have a poorer BER compared to PRBS-9. 

Yes

2) To improve BER of PRBS-31 at 11.5Gbps an equalizer has to be introduced (at both Tx and Rx ends?)

Yes, usually, but it still depends on the channel loss.  A very short channel might not have enough loss to require equalization.  Also you may get away with equalizing only the receiver but both ends is best.

3) Moving to 12.5Gbps, ISI (inter-symbol-interference) will be higher for PRBS-31 compared to PRBS-9. So compared to 11.5Gbps, BER will be respectively poorer for both PRBS-9 and PRBS-31. 

Yes

4) So at 12.5Gbps equalizer has to definitely be used. 

Still depends on what the channel loss is at the nyquist frequency.

5) However equalizer works better with PRBS-31 compared to PRBS-9. 

Yes

1) One of the objectives of encoding is to increase the number of transitions to aid in clock recovery. 

Yes

2) However increasing transitions increases ISI and hence higher grade of encoding (i.e, 128b/130b to 64b/66b to 8b/10b) requires a more complex equalizer. 

The better the equalizer the more channel loss you can handle

3) Does this mean than at the same data rate i.e., 11.5Gbps 64b/66b encoded PRBS-9 will have poorer BER compared to 8b/10b encoded PRBS-9. 

In theory PRBS7 or 9 can still use the DFE but 8B10B can't.  So 64/66 would work better.  I have had cases as low as 6 to 8 Gbps where getting a good BER required changing from 8B10B to 64/66 so that DFE could be used.  Most of the time this is a moot point since the protocol usually defines the encoding and 8B/10B wastes 20% of your throughput.

The reason why DFE doesn't work well on 8B10B is that a string of data that is the same, say all zeros, does not update all of the adapting DFE taps properly.  One work around is to take the 8B10B data pre-scramble it with, say, an LFSR-15 and then use 8B10B encoding.  8B10B can work with DFE in that manner.




----------------------------------------------------------------------------
Don't forget to reply, kudo, and accept as solution
----------------------------------------------------------------------------


View solution in original post

9 Replies
628 Views
Registered: ‎05-22-2018

Awaiting a reply,

 

Thanks in advance,

 

-Chan100

0 Kudos
Reply
Moderator
Moderator
596 Views
Registered: ‎07-30-2007

1) Why is it that the BER is dependant on the PRBS bit pattern used? Increasing the randomness of the 1's and 0's sent over the serial link should not generally affect the BER right?

The longest string of solid 1's or zeros is the same as the PRBS number.  PRBS-7 can have a run length of 7 0's or 1's but no longer than that.  PRBS-31 has run lengths as long as 31 bits.  This means that the longer PRBS patterns are more susceptible to intersymbol interference.  You can reference https://www.xilinx.com/publications/archives/books/serialio.pdf page 42/43 on this.

2) How does decreasing the voltage swing improving the BER? From what I know, higher swing should actually prevent bit errors?

There is an AGC and equalization it just depends.  We see this sometimes when large amplitude inputs are over equalized.  This is because the channel is designed to compensate large insertion losses.  If it becomes a problem LPM mode can be used and you can set the channel to use minimum equalization in extreme cases with high amplitudes and very short channels.

3) What is the meaning of the MGT not supporting 12.5 Gbps for PRBS 31-bit pattern? What is the dependency of the PRBS pattern used for the test, on the line rate?

?  Each block has its own limitations and it is possible that this is one.  It would probably be dependent on the speed grade.  Where do you see this.

4) Also, two different TX DIFF swing values (530mV and 250mV) were needed to be given to the two serial I/O links being used to obtain the same BER order (E-11). Why was using two different values required here?

Same answer as above.

5) Why is the DFE being auto-enabled when using the PRBS 31-bit pattern?

It would be most normal to use DFE for PRBS-31.  It mimics the properties of 64/66 which is used for higher line rates which tend to use DFE because higher line rates have more Insertion loss which would bring DFE into play more often.

 




----------------------------------------------------------------------------
Don't forget to reply, kudo, and accept as solution
----------------------------------------------------------------------------


0 Kudos
Reply
572 Views
Registered: ‎05-22-2018

Hi @roym ,

 

Thanks for the elaboration.

3) What is the meaning of the MGT not supporting 12.5 Gbps for PRBS 31-bit pattern? What is the dependency of the PRBS pattern used for the test, on the line rate?

?  Each block has its own limitations and it is possible that this is one.  It would probably be dependent on the speed grade.  Where do you see this.

So I saw this in a document titled 'Avnet Kintex UltraScale Board IBERT Design'. Pg 17 of the document mentions this point.

I have attached the document in question to this post.

Also when you say each block has its own limitations, are you implying that within a device, of particular speed grade, different banks of MGT have different dependencies? This would be one of the most important considerations while building a system, if it were true. Can you further clarify on this?

 

4) Also, two different TX DIFF swing values (530mV and 250mV) were needed to be given to the two serial I/O links being used to obtain the same BER order (E-11). Why was using two different values required here?

Same answer as above.

This too has been mentioned in the same page of the same document. However it was not implied they need to work with those different swing values but we happened to observe high BERs for default voltage swings (which was around 950 mV for both MGTs of SMA connectors) that got better with the suggested values of voltage swing.

 

-krishnachandrasekhar100


Tags (4)
0 Kudos
Reply
Moderator
Moderator
529 Views
Registered: ‎07-30-2007

I was talking about the sub blocks of the Transceiver each having its own timing specs.  I know in earlier generations the 8b/10B encoder stopped working around 8 or 9 Gbps.  It was never a problem because no protocol used it at those rates.  I don't really think that is the situation PRBS generator but it is possible. 




----------------------------------------------------------------------------
Don't forget to reply, kudo, and accept as solution
----------------------------------------------------------------------------


0 Kudos
Reply
505 Views
Registered: ‎05-22-2018

Hi @roym ,

 

Unfortunately, I could not understand what you meant. I ask about different MGT banks having different dependencies because let us say if the MGT sources to be used are of FMC connector, for a configured JESD mode, I cannot have different BERs for different serial I/O lanes.  If you mean the encoding type is what is of concern here, could it be that the BER can improve if 64b/66b encoding is used?

Taking these points into consideration, can you further share your views about this?

 

Thanks in advance,

 

 

-krishnachandrasekhar100

0 Kudos
Reply
Moderator
Moderator
493 Views
Registered: ‎07-30-2007

Not different bank but different sub blocks of the GT have their own timing limitations.  All the GTH's in a given device will operate the same way.  

What I see in the paper is that 12.5 Gbps cannot use PRBS-9.   This is because 12.5 Gbps in that set up has a lot of insertion loss and needs DFE which won't work as well with PRBS-9.  I suspect you could get this link working at 12.5 PRBS-9 but it would be more difficult.  You might have to tune the TX emphasis and gain settings use DFE but I think it ought to work.  But as I said 12.5 Gbps is usually where 64B/66B encoding is used and it has a similar run length to PRBS-31 so that would normally be used to characterize the channel at that line rate.




----------------------------------------------------------------------------
Don't forget to reply, kudo, and accept as solution
----------------------------------------------------------------------------


0 Kudos
Reply
Explorer
Explorer
460 Views
Registered: ‎11-18-2013

Hello,

    A few more questions on the same topic:

1) At 11.5Gbps PRBS-31 will have a poorer BER compared to PRBS-9. 

2) To improve BER of PRBS-31 at 11.5Gbps an equalizer has to be introduced (at both Tx and Rx ends?)

3) Moving to 12.5Gbps, ISI (inter-symbol-interference) will be higher for PRBS-31 compared to PRBS-9. So compared to 11.5Gbps, BER will be respectively poorer for both PRBS-9 and PRBS-31. 

4) So at 12.5Gbps equalizer has to definitely be used. 

5) However equalizer works better with PRBS-31 compared to PRBS-9. 

Are the above correct?

Now bringing encoding into picture:

1) One of the objectives of encoding is to increase the number of transitions to aid in clock recovery. 

2) However increasing transitions increases ISI and hence higher grade of encoding (i.e, 128b/130b to 64b/66b to 8b/10b) requires a more complex equalizer. 

3) Does this mean than at the same data rate i.e., 11.5Gbps 64b/66b encoded PRBS-9 will have poorer BER compared to 8b/10b encoded PRBS-9. 

@roym I will be grateful if you can answer my queries point by point. 

Thanks again,

 

 

0 Kudos
Reply
Moderator
Moderator
443 Views
Registered: ‎07-30-2007

 

 1) At 11.5Gbps PRBS-31 will have a poorer BER compared to PRBS-9. 

Yes

2) To improve BER of PRBS-31 at 11.5Gbps an equalizer has to be introduced (at both Tx and Rx ends?)

Yes, usually, but it still depends on the channel loss.  A very short channel might not have enough loss to require equalization.  Also you may get away with equalizing only the receiver but both ends is best.

3) Moving to 12.5Gbps, ISI (inter-symbol-interference) will be higher for PRBS-31 compared to PRBS-9. So compared to 11.5Gbps, BER will be respectively poorer for both PRBS-9 and PRBS-31. 

Yes

4) So at 12.5Gbps equalizer has to definitely be used. 

Still depends on what the channel loss is at the nyquist frequency.

5) However equalizer works better with PRBS-31 compared to PRBS-9. 

Yes

1) One of the objectives of encoding is to increase the number of transitions to aid in clock recovery. 

Yes

2) However increasing transitions increases ISI and hence higher grade of encoding (i.e, 128b/130b to 64b/66b to 8b/10b) requires a more complex equalizer. 

The better the equalizer the more channel loss you can handle

3) Does this mean than at the same data rate i.e., 11.5Gbps 64b/66b encoded PRBS-9 will have poorer BER compared to 8b/10b encoded PRBS-9. 

In theory PRBS7 or 9 can still use the DFE but 8B10B can't.  So 64/66 would work better.  I have had cases as low as 6 to 8 Gbps where getting a good BER required changing from 8B10B to 64/66 so that DFE could be used.  Most of the time this is a moot point since the protocol usually defines the encoding and 8B/10B wastes 20% of your throughput.

The reason why DFE doesn't work well on 8B10B is that a string of data that is the same, say all zeros, does not update all of the adapting DFE taps properly.  One work around is to take the 8B10B data pre-scramble it with, say, an LFSR-15 and then use 8B10B encoding.  8B10B can work with DFE in that manner.




----------------------------------------------------------------------------
Don't forget to reply, kudo, and accept as solution
----------------------------------------------------------------------------


View solution in original post

380 Views
Registered: ‎05-22-2018

Hi @roym ,

 

Thanks for your elaboration! That brought more insight into IBERT testing for us.

 

-krishnachandrasekhar100

0 Kudos
Reply