With vias in vogue, where will high-speed DDR4 signals go?

yvonneGan

With vias in vogue, where will high-speed DDR4 signals go? [Copy link]

This post was last edited by yvonneGan on 2020-2-25 17:30

DesignCon 2020 article interpretation series
Author | Huang Gang (member of YiBo Technology's high-speed team)

Fans who are familiar with Mr. Gaosu's writing style know that each issue of our booklet has a different theme. So have you noticed which themes we write a lot of articles about? Now you might as well take out your long-collected booklet and read it once, and you will know (provided that... you have it). You will find that there are several topics that Mr. Gaosu always enjoys digging into, such as high-speed serial, such as Dongge's production and processing series, which is also known to everyone as... "Ruyan". Of course, you will also find that there are many articles about DDR.

Yes! Our DDR articles definitely occupy a large part of the space, mainly because there are so many application scenarios. Almost every board will have various DDR systems, from one chip to dozens of chips, from chip version to DIMM version, covering more than 80% of our boards. So in this issue of DesignCon article interpretation, Mr. Gaosuo brings you an article about DDR applications to see where the design limit of DDR is.

The article I want to share with you today is as follows. The title this time is very easy to understand. It is called " Analysis of the Impact of Via Stub on Signal Quality in DDR4 Channel".

(, downloads: 0)

The main point is definitely the impact of via stubs on DDR4. First of all, the author made some preliminary judgments and analyses on the signal quality of DDR4, such as requiring the insertion loss resonant frequency of the channel to be greater than 5 times the clock frequency. For example, according to the 3200Mbps analyzed in this article, the clock is 1.6GHz, so the required resonant frequency must be greater than 8GHz.

(, downloads: 0)

Well, let's get to the point and see how the author analyzes DDR4 channels in different situations. They mainly analyze three different scenarios. The first is the surface routing of the particle version, which definitely does not have via stubs; the second is the inner routing of the particle version, which has via stubs; the third is the inner routing of the Dimm version, which not only has via stubs, but also includes the impedance mismatch point of the Dimm strip connector . The following figure is a schematic diagram of the three different cases.

(, downloads: 0)

Since the purpose is to study the influence of via stubs on signal quality in detail, it is necessary to compare different via stub lengths. Therefore, this article analyzes a large number of stacks, with via stubs ranging from 52.7mil on 14 layers (the inner layer routing all starts from the L3 layer, and the longest via stubs on different stacks are analyzed) to 124.7mil on 28 layers. This covers almost 99% of application requirements.

(, downloads: 0)

In addition, the author also gives some parameters of the vias used and the model for 3D simulation .

(, downloads: 0)

OK, let's look at the results of the analysis.

First, in case 1, where the surface routing is performed without via stubs, the results are relatively simple and clear. At a rate of 3200Mbps, the signal quality is relatively good, and at a strict bit error rate of -16, the eye diagram still has a relatively large margin.

(, downloads: 0)

Then we analyze case 2. It can be seen that the via stub length is a critical point when it is 73.1mil. At this time, the eye diagram is just compressed at the mask of the -16th power bit error rate. If it is lower, it will not meet the bit error rate standard.

(, downloads: 0)

What about case 3, what will happen after adding a DIMM connector? Yes, I expected it to be worse, but did I not expect it to be this bad?

(, downloads: 0)

It can be seen that in the case of Dimm strip applications, stubs above 50mil are risky and cannot pass the -16th power bit error rate standard.

After analyzing the eye diagram, let's take a look at the frequency domain analysis and the comparison of insertion loss.

First, we compared the two cases of case 2's 73mil critical point and 83mil. It can be seen that although the difference is only 10mil, the loss conditions of the points marked in the red box in the figure below are quite different, with a difference of almost 30dB .

(, downloads: 0)

The comparison between the granular version and the Dimm bar application under the via stub length of 53mil is as follows. It can be seen that there is a difference of nearly 10dB in the Dimm bar mode.

(, downloads: 0)

Finally, the author gives a prediction of the success rate under different via stub conditions, which is very intuitive and clear.

(, downloads: 0)