DesignCon paper and tutorial explain de-embedding
Engineers at FPGA manufacturer Xilinx had a measurement challenge: how to characterize the 28Gb/s serial links in their devices. At such a data rate, even very short connections can cause signal degradation and high-frequency content in the signal will be lost. To make matters worse, the data streams must run on differential pairs and even the slightest mismatch in the signal and return paths will cause skew. A paper at DesignCon 2014 explained the problem, offered a new kind of solution, and a hands-on tutorial gave engineers a chance to use simulation and measurement software tools that solved the problem.
It's impossible to get equipment to directly measure high-speed signals because you need a system of interconnects such as a fixture probes, cables, and connectors. The interconnection between the test equipment and the device under test (DUT), however, behaves like a low-pass filter and degrades the signal. For you to effectively evaluate the FPGA transceiver (DUT), you must be able to view the original signal as it appears at the package pin. Using software to remove the effects of interconnects lets you view the original signal. It's a process called de-embedding where you characterize a channel, then apply an inverse filter in software on the captured signal.
Figure 1 shows the evaluation board that Xilinx used to characterize the FPGA. The transmission channel consists of traces, vias, connectors, and cables.
Figure 1. Xilinx Virtex-7 VC7222 Characterization Board set-up for measuring the 28 Gb/s GTZ Transceivers at the package using de-embedding features on an Agilent 86100D DCA-X with 86108B precision sampling oscilloscope.
De-embedding software has been an oscilloscope option for several years. But, acquiring the data that models a transmission channel path can be challenging and expensive. Yes, you can characterize a transmission line on a board using a probing technique, but not everyone can afford the skilled operator and equipment. Another technique uses test structures (often called "coupons") on the board, which emulate the channel.
The test structure is twice the transmission channel path, mirrored around the center reference plane, to enable simple coax connections to a VNA (vector network analyzer) for S-parameter measurements. The S-parameters for this test structure, aka 2X-THRU, can then be mathematically divided in half using AFR (Automatic Fixture Removal). The result is an S-parameter behavioral model of the single transmission channel path. A third method uses measurement-based simulation with multiple transmission line segments so you can modify the length of the transmission channel. The simulated model is flexible and can be adjusted to create an S-parameter model for all of the transmission channel paths when multiple transceivers with multiple path lengths are available.
A new technique called 1–PORT AFR uses the reflection from a short or open at the end of the test fixture in the channel to remove unwanted effects, thus allowing accurate characterization of the real world channel. The 1-port AFR refers to either a single-ended or differential connection to only one end of the in-situ transmission channel and eliminates the need for probes or a 2X-THRU calibration structure. The DesignCon paper, De-Mystifying the 28 Gb/s PCB Channel: Design to Measurement explains the process in detail. Figure 2 shows the transmission channel components. The technique demonstrated at DesignCon relies on measuring the differential path in question with an open circuit at the far end—at the balls of the BGA (Ball Grid Array). Information is obtained using the reflected signal. The technique produces the full 16 S-parameters needed to make the behavioral model. (There are actually two VNA ports used in the measurement because the signal path is differential.)
Figure 2. 28 Gb/s data signals from a Xilinx FPGA travel through the package, PCB traces, connectors, and cables before reaching test equipment.
During the presentation of the paper on January 31, Agilent's Heidi Barnes explained the problem with a measurement-based model: bandwidth. That is, you can't measure the infinite frequency response of the channel to capture all of the spectral content in a transmitted bit. So, how much bandwidth do you need? The effects of limited bandwidth can distort the bits and when you look at them with an oscilloscope through de-embedding, the eye diagram will look more distorted than it actually is.
"Without the complete spectral content," said Barnes, "a transformation from the frequency domain to the time domain can produce ripple in a square wave. It's called the Gibbs Phenomenon." The ripple isn’t actually there, but it appears in an oscilloscope measurement from the de-embedding process if the S-parameter model has insufficient bandwidth. That ripple causes an inaccurate eye opening. Barnes then compared eye opening results from the transmission channel S-parameter model with four different bandwidths.
Agilent's Rob Sleigh then showed the audience how to implement the de-embedding technique - using Agilent N1010A FlexDCA software running off-line with saved time domain waveforms from the Xilinx 28Gb/s GTZ transceiver. In the tutorial given on January 28, participants (including me) were able to see the results of the models. Figure 3 is a screen image I captured during the tutorial. It compares four eye diagrams: the signal as measured by the oscilloscope directly (D5A, top trace) and three de-embedded waveforms (F4, F5, and F6, second, third, and fourth traces, respectively). The D5A waveform is the signal before de-embedding the transmission channel path, and then followed by de-embedding with the probe (F4), 2X-THRU AFR (F5), and simulated ADS (F6) behavioral models. All of the transmission channel behavioral models are able to show the pre-emphasis peaking that exists at the package output of the transceiver.
Figure 3. Direct DUT measurement at the output of the fixture (D5A, top) vs. de-embedded signals showing what the 25 Gb/s signal would look like at the balls of the device (three lower waveforms generated using different S-parameter models).
The tutorial concept was a first for DesignCon. Everyone in the audience had a laptop running Agilent N1930B PLTS software for multi-port network-analyzer measurements. They also had Agilent ADS for simulated models and N1010A FlexDCA for in-situ time-domain measurements. I like the idea of making a paper into a hands-on tutorial, though it might come off better if the paper can be presented first. Given the structure of most conferences where tutorials precede technical papers, that was impossible.