Signal Integrity Issues

Semiconductor Engineering sat down to discuss signal integrity with Rob Aitken, research fellow at ARM; PV Srinivas, senior director of engineering for the Place & Route Division of Mentor Graphics; and Bernard Murphy, chief technology officer at Atrenta. What follows are excerpts of that conversation.

SE: With the there’s noise from signal and power. The device is supposed to be cheap, but suddenly it’s very complex. And it’s supposed to be generic but it’s very specialized. And you’ve got I/O from multiple places. What happens to signal integrity?

Srinivas: We don’t have a lot of experience specifically with IoT devices yet. But you have to figure there will be a lot of inputs will be coming in and there will be long wires from the signals, so the impact on signal integrity could be significant.

Aitken: All of these IoT devices will need radios. Are you going to try to put the radio on one chip to lower costs or are you going to put it on a separate chip, which increases the cost but may simplify the other pieces of the SoC design. In a way, it’s a classic IP positioning problem. It’s a hard thing to do, but if you do it right it’s good stuff. And if someone else does it right and you can use that, then that’s advantageous. In the IoT space, is there a platform-like object that’s sufficiently common across big chunks of the leaf-node chips that can be provided as a whole, or as a fairly easily assembled object? So here’s a radio part, here’s an ADC and a DAC, so you can hook up whatever analog part you want, and here’s some processing capability that allows you to have the software stack to glue all of this stuff together.

SE: Then it becomes a partitioning problem, right?

Aitken: Yes, exactly. And do you have to have consistent pad rings? That will affect everything on-chip.

Murphy: A big question is what kind of radio. There seems to be an almost infinite number of possibilities, and they’re growing by the day, from countable infinity to uncountable infinity.

Aitken: But they do tend to work in a couple of bands. You can say you’re going to have a 200GHz band and a 900GHz band, and one radio that does each of those, and then you can tweak it to turn it into a Bluetooth or ZigBee radio.

Murphy: We can have a mesh radio based on that.

SE: With the IoT, we’re talking extreme low power, particularly on the edge devices. How does that affect noise?

Srinivas: The threshold voltage is relatively high compared to the operating voltage, so it changes this artificial way we design these. So the delay is 50% compared with 30%. You can measure the delay or waveform propagation as a signal effect, and characterize that.

Murphy: Is this switching clocks or power?

SE: It’s both.

Aitken: And the power switches are a big deal, because at these lower voltages they never turn on or off.

Srinivas: That’s true.

Aitken: You have to be very aware of this when you’re designing your power gates, and you have to figure out a good way to do that. We’ve built a couple of these IoT demonstrators and learned a few things about what kinds of power gates and transistors you should use. We’ve also observed that the variability is massive in these devices.

SE: Because of use cases and the technology itself, right?

Aitken: Yes. So if you use an LVT device for your clock, for example, then you have much better variability than if you use a RVT device. You can pay the extra penalty in leakage in order to achieve better control of your frequency. And all of these things work together. It goes back to what kind of library you’re going to use. We’ve observed that for a 65nm type of technology, you can look at three different realms of operation. You can look at the sub-vt 300 millivolt type of range, and at that range you’re going to get the lowest power but it’s not the lowest energy because it’s leaky. The lowest energy point is more like the 400 millivolt range, but in that range you still need black-belt designers to make this work because it’s really hard. On the other hand, if you go to the 500 to 600 millivolt range, you can get a lot of the energy benefits and it’s a little bit easier because many of the assumptions of current design still hold.

Murphy: Some of the memory guys are using 700 millivolts.

Aitken: The standard 6T SRAMs die at low voltage. You can add assist techniques to push them a little beyond that, or you can add coding, but you also can just use different bit cells. So for our 300 millivolt design we used a 10-transistor bit cell. It was really big, but it worked at low voltage.

Srinivas: Even multi-voltage techniques are included. Obviously you can turn off and on using power gating techniques. Implementation is difficult. Buffering knows whether it is on and off. But is power available in this domain or not?

Aitken: The tools are definitely getting better in that space with the 1801 standard. Everyone kind of agrees on the overall design.

Murphy: One thing you have to think about with an IoT device is ambient noise. The classic worst case is automotive, which is a horrible environment, but it’s going to apply in a lot of other cases, too.

SE: One of the new wrinkles on signal integrity is how accurate is your signal. What are you trying to see with a sensor and is that correct?

Aitken: It’s interesting to look at who has already solved that. Flash memory is not capable of solving the gigabytes we want it to in a classic every-bit-works mechanism. It’s pretty good, but it produces noisy data. The controller and the modem between the processor and the memory cleans all of that up. So the processor thinks it’s just reading bits that are good, but there is actually a huge amount of encoding and decoding. That contributes the latency of the overall system, but it definitely improves the integrity of the communication.

Murphy: This goes back to noise-tolerant communication. There have to be re-tries based on some kind of error code until you get a correct value.

Aitken: Or you can deal with areas that are somewhat error-tolerant. Audio and graphics are the classics in that space. But in real life it’s not so clear-cut. Gamers don’t like bad pixels in games. They will report them and complain about them. Error tolerance, in general, has a lot of promise, but it’s going to require a mindset change. And it will likely happen in places where there isn’t any other way forward. Supercomputing is already error-tolerant in a way. They have to tolerate chunks of their system falling apart constantly. Otherwise they’d only be working five minutes at a time.

Murphy: Sensor channels, depending on the application, probably could adapt to that, too. If I’m getting an analog signal and occasionally it does something, I can smooth it out.

SE: As we take advantage of established nodes, what changes there and how does that affect signal integrity?

Aitken: One is the devices. They add an ultra-low leakage device to enable ultra-low energy computing. The other thing that changes is the consideration and modeling of the low-voltage behavior. If a process has a 1.2 volt nominal, it has typically operated between 900 millivolts and 1.3 volts. No one cared about what the devices did below that or whether the SPICE models reflected them accurately. But if we keep releasing chips in that area, all of that suddenly matters and the foundries have to go back and characterize all that stuff. The good news is processes have been up and running for a long time, so presumably they have them under pretty good control.

Srinivas: They’ve been working with these processes for years.

SE: If you can build a platform and verify it works, where else can it be applied?

Srinivas: On one hand you have the big system companies, which keep making bigger chips. They’re going to the lowest process possible. They’re not going back to older nodes. But for IoT, they want to do chip design at the lowest possible cost with the lowest possible operating voltage. They can benefit from these things.

Murphy: Some of the markets controlled by regulation are going to take advantage of this, as well, such as medical or automotive. If you have to have an application that is proven compliant, you can show you’re already certified.