More Art Than Science

By Ann Steffora Mutschler

Not so long ago, high-end digital devices were mostly just digital. Today large SoCs contain a significant amount of analog/mixed-signal content. And given that analog circuits have certain sensitivities different from digital blocks, there is a desire to convert some of those analog signals to digital to achieve power savings and take advantage of digital verification technologies.

But knowing which signals to convert and which to leave in analog is a complex challenge, and more of an art than a science.

“The guys that know this the best are still the analog designers, said Qi Wang, technical marketing group director for low power & mixed-signal at Cadence. “They know the analog functionality and they would know what the problem of a particular circuit design style is for the better way to convert to digital circuitry. You need to keep in mind the conversion is not just for the power. It is meant for easier migration to advanced nodes or it is because of the IP requirement—people want more programmability or configuration of a particular analog module. So there are multiple angles driving this. You know it from your design requirement. If your design requirement needs various flexible, various configurations of this analog IP, you have to add more digital controls to have more digital controllability. And if your design requirement has a particular power budget that is impossible to achieve with the analog part then you have to convert some of it to digital.”

To some extent this becomes a partitioning issue.

“The engineer that’s developing the IP will look at the functionality of the blocks and say, ‘There’s a certain part of the signal processing chain that can be done digitally,’ and that’s where you would put your partition,” noted Navraj Nandra, senior director of product marketing for analog and mixed signal IP in the solutions group at Synopsys.

A good example of analog-to-digital conversion in IP is high-speed SerDes. Years ago, high-speed SerDes circuits were developed using analog techniques so the loop filter, the PLL and the clock invasive recovery were all done in analog.

“The original SerDes architectures going back quite a few years basically had a phase frequency detector followed by an analog charge pump, and then you had a resistor/capacitor to set up the loop filter for the PLL. Today in high-speed SerDes designs, we do the same thing. We’ve followed the same development trajectory. You see that you can actually use a lot of digital techniques inside your clock invasive recovery circuit to do exactly the same things that you were doing in the analog domain,” Nandra explained.

In the block diagram there is still a phase frequency detector. But after that engineers use a technique called decimation to take signals from analog and move them into the digital world.

“After the signals have been decimated, you can then do a lot of very clever things in the digital domain, including setting up a loop bandwidth for SerDes,” Nandra said. “Instead of using a resistor and capacitor, which are real analog components to set up the loop bandwidth, today you can use basically digital blocks. Once it’s in the digital domain, you can leverage all the great things about digital technology in terms of scaling, power savings, you don’t even have to worry about yield problems you potentially get with analog blocks because in the digital domain, the technologies are a lot easier to work with.”

Gene Matter, senior applications manager at DOCEA Power, agreed. “For analog circuitry there’s a boatload of stuff we used to do in analog for comparators and taking something from the analog domain (which is like a waveform or a wave) and digitizing that. And if you have the ability to take some sort of transduction of something in the analog domain and digitize it, it’s a matter of what is the area to do that. And what is the fidelity or precision you need?”

He noted that in the analog world we deal with the signal noise ratios and dB ,and these are the qualitative metrics for performance, while in the digital time domain there are a different set of qualitative metrics that are basically transforms and mathematical conversions per second operation. “If the fidelity on the left hand side (the analog) can be achieved with lower power and smaller area through the computational performance on the right hand side, what we’ve found is basically that the class of mathematical operations for transcendentals or things like trigometric functions (sine, cosine functions), vector operations for arrays and vector mathematics can be put it into a matrix and then solved mathematically. It has become very, very useful.”

To achieve this, these applications need to be parallelized, Matter explained. “You have to be able to take a conversion on the left hand side of a time-varying waveform, digitize that and convert it very, very quickly. It really lends itself to either vector mathematic operations in the floating point in vector array processing or through something we call SIMD single instruction multiple data – that you can do SIMD instructions.”

Further, where there is a category of analog functions and it’s at a very low level on the left hand side in analog and there is a corresponding digital representation on the right hand side to do those conversions, large-feature-size analog functions can be converted to digital, he continued. “The beauty about the digital part is that it’s re-purposeful. In other words, the same hardware that you could do for the signal processing to do a noise filter also can be used for amplification to do level boosting and scaling. As long as you don’t need the same recursive hardware simultaneously at the same time, you can take all these blocks that were over here on the left hand side that were analog that are serialized and take up a large area. You could decode those sequentially into the same hardware and put a sort of looping bumper thing that crunches all that stuff and shuts out the outputs as you need them. But you have to think about it different.”

When not to convert
Still, there are situations that do not warrant analog-to-digital conversion. That includes anything requiring really high fidelity or really high performance, where there are millivolts or microseconds or microcurrents, or which requires a lot of very precise measurement or very high fidelity, Nandra suggested.

That high-precision measurement would have to still be done in the analog domain. “It’s just the nature of the signal because analog signals are continuous whereas digital signals are segmented. It’s like people—you have analog people and digital people. Of course when you talk to a digital engineer what they will tell you is that if you have high enough resolution, you won’t be able to tell the difference because you’ve divided up that analog signal so much now that if you reconstruct it, it’s going to look just like the original signal,” he continued.

“Nowadays the DSP power is such that you could probably do everything in the digital domain because that partition goes quite far into the analog part, but the problem is you start consuming more power. One of the tradeoffs that you make with some of these techniques when you push everything into the digital domain is that your power consumption can sometimes go up because you’ve got processing going on. You’ve got some processor or whatever trying to figure out the algorithm and that consumes power so the circuitry becomes more complex. That’s okay because you’ve got lots of digital gates and the feature sizes are a lot smaller, so you won’t notice so much impact in terms of area. But you do get quite a bit of processing going on to make sure you are deconstructing and reconstructing the signals correctly,” Nandra added.

Also, there may be added challenges when converting from analog to digital, Matter observed. “There’s some stuff that you cannot do in digital that you can only do in analog because it makes the validation task that much more difficult. Digital validation techniques of digital circuits are still all based on discrete steps. Unfortunately analog validation is all based on curve tracers. So you have to adopt some different types of verification techniques. It requires a hybridization of these techniques to address the problem.”

In the end it is the design requirements and the analog designers’ capabilities to decide what is best to convert to achieve the best power, performance and area.