Shmoo results

Autonomous Machines Jetson & Embedded Systems Jetson TK1
1

We are doing some preparation work for memory characterization on Jetson TK1 board. Here are two questions regarding to the memory characterization.

  1. According to the memory characterization application note, we have passed chapter 5.8 Run Santiy Shmoo. However, when doing chapter 5.9 Running Full Shmoo, no comments could be found in the generated excel file. So is there any version requirements for the excel software? Currently, the installed excel version in our PC is 2015.

2)Is it correct that the chapter 5 and chapter 8 are doing the same task? That is, if we go through the chapter 5 test, then it is not necessary to do the chapter 8.

Thanks!

2

Hi NICK_LIN,

  1. No version requirements, but as 2015 is so latest, you can try with 2010 which I used OK with.

  2. They are different, chapter 5 is for generating CFG/DVFS, chapter 8 is for regression verify test.

3

Thanks Trumany. We have figured out the excel problem with 2010 version.

Besides, I was wondering if it is normal to take about 9 hours to run chapter 5.9 in each frequency? It seems will take lots of time to get the shmoo results with all frequency.

Thanks!

4

Yes, full shmoo is a very long term test, 8~9 hrs for each frequency is normal.

5

Hi Trumany,

Thanks for your help!

In our custom board, we use the same schematic as Jetson Tk1 board for DDR3L, except replace 2GB memory to 4GB memory with chip Hynix_4GB_H5TC8G63AMR_PBA. Below are my questions regard to the DDR3L architecture when running shmoo.

1)In TK1 reference manual, it is said that “The Tegra K1 device supports a single physical channel of memory interface, which operates in either signle 1x32 or hybrid 2x32/1x64 configurations.”
Is it correct that the hybrid 2x32 is used for Jetson TK1 board in default?

  1. The parameter DDR3L_TOPOLOTY is set to 1Rx32 in file TK1_generatBKVfiles.bat. What does it mean? Do I need to change this parameter for our custom board?

Thanks!

6

Hi NICK_LIN,

  1. Yes, it is 2X32.

  2. 1Rx32 each rank is x32 type, it should not be changed.

7

Hi Trumany,

Is it correct that it is not necessary to add DRAM package PIN delay to the board delay parameters? The DRAM package PIN delay is described at Table 38 of TK1 Design Guide.

If it is true, than I was wondering if the DRAM package delay will be neglected or it would have been complemented somewhere by the Tools when generating the CFG file.

Thanks!

8

No need to add DRAM package PIN delay, and Table 38 does not mean DDR chip, it is the pin delay of TK1’s ddr interface balls.
Only need to follow the AppNote in shmoo package to input board delay, pin delay will be covered in tool automatically.

9

When analyzing the full shmoo results for our custom board, we found out Fail in the left side and Pass in the right side for DqIvref and DQsIvref. SO we need to adjust the center value to get margin.

However, it seems no change of CFG file setting value when input new EmcDllXformDqsOffs and EmcDllXformDqsOffs values into TK1_converter_V2.xlsx file. Could these two parameters be adjusted? How should I adjust these values?

Thanks!

10

Adjusting the center value to get margin is correct, and yes, the EmcDllXformDqsOffs and EmcDllXformDqOff will not impact value of EmcXm2DqsPadCtrl, but they can impact EmcDllXformDqsn and EmcDllXformDqn.

BTW, seems you are not using latest package which should be V1.0.2, and the analyzing file in it is TK1_converter_V3.xlsx, there is a little bit difference between V3 and V2, so please use this latest one.

11

With the Updated V1.0.2 tool, our custom board could pass 708MHz when running sanity shmoo. So we start to run full shmoo in 672MHz.

It seems the full shmoo results do not have good margin with AubocaloffsetDQDQS, DqIvref,DqMultoffs, DqsIvref,DqsMultoffs, others are all pass. We have tried lots of methods to adjust the center values for each items, however, only a little bit improvements could be achieved. It seems our custom board do not have enough margin for us to adjust.

Below are my questions

  1. Since the board have passed 708MHz sanity shmoo, does it mean it could also pass full shmoo in 672Mhz if we put more efforts in adjusting the margin?

  2. The items we failed are all related to Dqs, is it better to adjust all failed items in the same time or one item by item? Any other suggestions for us to further improve the margin?

Thanks!

12

Here are attachments of failed items.

https://drive.google.com/open?id=0B1OwKZBEYjM4a1Q2NzFZcUNfaWs

https://drive.google.com/open?id=0B1OwKZBEYjM4dmhGWEdfWWgtenc

https://drive.google.com/open?id=0B1OwKZBEYjM4aXRybVlDUzd6N0U

https://drive.google.com/open?id=0B1OwKZBEYjM4X0x0UFdNeU40ZE0

https://drive.google.com/open?id=0B1OwKZBEYjM4dmhGWEdfWWgtenc

13

Sanity passing doesn’t mean it could pass full shmoo.

Per your result, please change AutoCalOffset to ‘6’ and use TK1_converter_V3.xlsx to generate relevant register value and test again.

14

Hi Trumany,

 I want seek your help about the Full Shmoo Results. we passed 512MHz full shmoo ,but there is some C(crashed) mark in the Full Shmoo Results sheet( According to the memory characterization C means that HOST_PC does not detect DUT in RECOVERY MODE), if it will produce an effect on Memory Regression Test for L4T

Thanks!

15

Since you passed full shmoo which means each register setting has enough margin, these C will not affect later test.

16

Hi Trumany,

Thanks for you imformation.
How can I get Test Image(Tegra124_Linux_<RX.X>) For Memory Regression Test for L4T ?

17

Hi Frank,

Assume it is your own board, if so you need to build these system image based on your board.
For Jetson Tk1 board, you can download image from [url]https://developer.nvidia.com/embedded/downloads[/url]