cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 
Highlighted
Observer
Observer
1,212 Views
Registered: ‎08-10-2018

Aurora 64/66 channel_up & Soft Errors with VCU108 QSFP transceiver

Hello all,

I am having trouble getting Aurora 64/66b core working reliably with a QSFP+ transceiver on a VCU108 Eval board.

Hardware Target: VCU108 (GTY) -> QSFP Transceiver -> (Lane 1 TX -> Lane 1 RX via Optical loopback).  I also have the Eval kit included QSFP copper loopback adapter which I am using as a baseline.

Tested IBERT example design in hardware

    1. Copper Loopback works fine, 0 Bit errors over minutes, 2d eyescan looks good
    2. Optical loopback works fine, 0 bit errors over minutes, 2d eyescan looks great (Validates optical path)

 

Tested aurora example design in hardware

Aurora (v11.2 rev4) configured: 1 Lane, Duplex, Streaming, INIT_CLK = 125Mhz board clock (default otherwise)

  1. Copper loopback works fine
    1. Data_err_count = 0 indefinitely
    2. channel_up stays 1
    3. lane_up stays 1
  2. Optical loopback has issues
    1. Data_err_count is ff within first second
    2. Channel_up de-asserted periodically once every ~3 seconds (very short low), is high otherwise
    3. Soft_errors are present
    4. Lane_up is always 1
  3. Setting Near-side loopback modes 1 or 2 with the optical loopback plugged in work fine, no issues as in 2.b above.  

 

Adding attenuation to the optical path results in higher channel_up de-assert rate.

Configuring aurora to 25Gbps (vs default of 10.25Gbps above ) results in dramatically higher channel_up de-assert rate and soft errors.

 

Your help appreciated,

Will

Tags (3)
0 Kudos
5 Replies
Highlighted
Xilinx Employee
Xilinx Employee
1,166 Views
Registered: ‎03-30-2016

Hello Will @wferry

I found the issue you reported here is very similar to this issue reported here:
https://forums.xilinx.com/t5/Networking-and-Connectivity/Aurora-64b-66b-lane-down/td-p/905033

Same board similar usecase. Electrical connectivity works perfect but failed on optical connectivity.
Set equalizer setting to LPM shows a better result compared to DFE, but you still find some issue.

 

I am wondering if you can modify your optical module to have less emphasis setting.

GTY LPM could be over-equalizing the input signal , that might trigger some soft_error.

Thanks and regards
Leo

0 Kudos
Highlighted
Observer
Observer
1,149 Views
Registered: ‎08-10-2018

 Hi Leo @karnanl

 

Thank you for your suggestion!

 

I would like to try this, but I am unsure the most straight-forward / easiest way to go about establishing i2c communication to the QSFP port on the VCU108.

 

The 'System Controller' console does not appear to offer generic i2c interface, sadly.

In the user guide, I see that the i2c bus is connected to both the FPGA and the Zync-7000 system controller.

Could you point me to any examples/resources on the easiest way to establish control to the i2c on the VCU108?

 

Thanks kindly,

Will

0 Kudos
Highlighted
Observer
Observer
1,099 Views
Registered: ‎08-10-2018

Hi Leo @karnanl

I implemented i2c control and verified: the QSFP module defaults to output emphasis setting of 0.  So we can not take it any lower.

An important point to note:

The iBERT design (at 25Gbps) , using the same GTY, optical transceiver (default settings) & optical path as with aurora, shows 0 bit errors over a course of multiple minutes.  This is with DFE enabled (by default).  (Serial Link data attached)

So the issue appears to be aurora-specific.

Any other ideas?

Thanks,

Will

serial_link_ch0.png
0 Kudos
Highlighted
Observer
Observer
998 Views
Registered: ‎08-10-2018

Hi Leo @karnanl

I added attenuation to the optical path and found performance to be worse - the channel_up de-asserted more frequently.

Will

0 Kudos
Highlighted
Xilinx Employee
Xilinx Employee
924 Views
Registered: ‎03-30-2016

Hello @wferry

1. I tried to generate the default Aurora 64B66B default IP setting using 2018.1. (XCI is attached. I assumed that you are using exactly the same configuration)
   Please let me know if my understanding is not correct.
   
   It seems that default setting is using DFE equalizer setting. (Please see below)
   W_Default_is_LPM.png

2. You have reported that you changed the default DFE setting to LPM setting.
   How did you change DFE to LPM setting ?? Could you please describe more detailed steps ?
   ( I am expecting that you set the RXLPMEN=1 and do DRP write necessary to the GTY , and reset the GTY as described in UG578 Chapter 4 "Switching Between LPM and DFE Modes at
Run Time")

3. I also looked at your EyeScan result here : https://forums.xilinx.com/t5/Networking-and-Connectivity/Aurora-64-66-Eyescan-Puzzler-Optical-vs-copper-loop-back/td-p/917518?lightbox-message-images-917518=48673i093DC16FA3D2C010
   I assumed the line-rate is 10.3125G. Could you please describe the equalizer setting you are using ?
   W_LPM_DFE_EYESCAN.png
4. BTW, You mentioned that you use optical loopback-module  Do you know if the loop-back module has AC coupling inside ??
   My concerns is without AC coupling, you will need to make the input signal to have a good DC balance.
   Aurora Training pattern is keep sending the same "Not Ready" signal during training, this is maybe less random than PRBS-31 used by IBERT.
   Because the DFE needs random input data to works.

5. Could you please use the working optical module with very short fiber-cable ?
    Standard optical module should have AC coupling inside the module.
   
Thanks & regards
Leo

0 Kudos