With all of the data being generated during test cycles, it’s logical to start applying appropriate analytics, artificial intelligence (AI), and machine-learning principles to act upon that data to get answers, insights, and improve processes and outcomes. It’s logical, but not easy. It turns out that there are quite a few challenges, not least of which is developing the right learning models, and then breaking down the data silos along the path through development, test, manufacturing, and in-field operation.
The benefits of tackling these problems include faster test times, more accurate channel modeling, less reliance on specific hardware, more integrated design and test processes and resources, faster scaling, and plenty of cost savings. As neural nodes learn faster and cloud-based, massively parallel computing takes hold, there’s also the possibility of finally getting to a point of predictability.
1. Tektronix and HPE collaborated at DesignCon 2018 to show how to use machine learning in combination with a real-time oscilloscope as a substitute for software-model-based channel analysis for high-speed links.
At its booth, Tektronix collaborated with Hewlett Packard Enterprise to show how to combine test instruments with machine learning as a substitute for software-model-based channel analysis (Fig. 1)
It’s always a difficult, iterative, process to use EDA tools to model high-speed channels and get correlation in the eye diagrams between the model and the measured channel or link. Beyond the physics of the actual interconnect, there’s the problem of missing models and the compute-intensive nature of the simulations.
However, by using a two-step training method and a specially developed, self-adapting machine-learning model, the team was able to get correlation on a 12-Gb/s PRBS9 channel (Fig. 2).
2. Using a two-step training process and a self-correcting, machine-learning model running on a real-time scope, the Tektronix and HPE team achieved tight correlation between the scope-embedded model and the actual measured signal.
Chris Cheng, distinguished technologist at HPE, led a team that delivered a paper at DesignCon titled, “Test Instruments as the Machine in Machine Learning.” His co-presenters included Yongjin Choi, Ph.D, master technologist at HPE, as well as Ting Zhu, Ph.D., expert engineer at HPE.
In the paper, the team elaborated upon the two-step training method, as well as the advantage of machine-learning models. These include:
- Complete IP protection
- Better accuracy than IBIS AMI
- Lots of free toolkits available
- Massively parallelizable
The two factors that make test instruments so suitable as platforms upon which to implement machine learning are that they form a measurement front end for neural-network training, and they can provide correlation for retraining inaccurate models.
With specific regard to channel optimization, the team concluded that applying machine-learning techniques means that SerDes vendors don’t need to know the exact channel topologies. Moreover, system designers don’t need to completely know the equalization components.
The concepts are interesting and Electronic Design has encouraged Tektronix and HPE to develop a feature article on the topic and its practical application to present here, so stay tuned.
Keysight Launches PathWave Toward Predictive Analytics
Also at DesignCon, a secretive team of “blue ocean” thinkers in the test and measurement software space were allowed to unfurl the culmination of their effort: PathWave. According to Brigham Asay, director of strategic planning at Keysight’s Internet Infrastructure Group, the team was set up in Atlanta, Ga., with the express goal of remaining out of the daily corporate grind so that they could spend their time researching and developing something new, and bold, and hopefully useful.
With PathWave, they seem to have come through, in spades. PathWave is a software platform that integrates design, test, measurement, and analysis (Fig. 3). In doing so, it moves data from barren silos into a fertile, connected, analytics environment where it can be used to develop greater insight, improve performance of test solutions, and get products to market faster.
3. Keysight took advantage of DesignCon for the big reveal of PathWave, a fully integrated approach to design and test that makes full use of test data to accelerate and improve processes and outcomes.
These improvements will come about through a number of PathWave features, including advanced analytics on manufacturing and test and equipment data. Using these analytics, it will be able to mine data sets, determine patterns, and predict future outcomes and trends. Key enabling factors include open APIs and scalability, using cloud-based servers.
This latter point is interesting. Asay said this also untethers test data analysis from the specific equipment at the desk or in the field, so test and measurement starts to become equipment agnostic: Just get the data (accurately, of course).
If all this sounds familiar, it should. It seems like just a few months ago that we discussed how and why Keysight should “own the IoT.” The premise of that piece was that Keysight, and other test-equipment vendors, including National Instruments, are well-versed in acquiring and handling data. It’s a critical part of the IoT, but the real benefits accrue from good analytics, and that’s clearly where Keysight is headed with PathWave.
Along the way, it won’t be abandoning VSA, X-apps, Signal Studio, or Network Analyzer. Those will still be sold standalone. Much of what Keysight has done to date has also been incorporated into PathWave (the team did check in now and again to make sure they didn’t go too far into the blue).
A Time to Rest
The test and measurement industry has had many terms to express how fast we get results from our test equipment. It used to be “time to answer,” then it was “time to insight.” Now, it may be “time to rest.” So, sit back and let the machines do the work, for a change. Just don’t get caught.