Field Programmable Analog – Technical Restraints, Part 2

Moore’s Law makes FPGAs a viable alternative for digital. Last time around, we showed that if you scale an FPGA design down to deep submicron (DSM) process nodes, an FPGA can become performance and cost competitive with 180nm ASIC solutions. In the digital world, transistor scaling delivers a lot of benefits. Scaled digital transistors are faster, lower power, and smaller than the same transistor function at a larger semiconductor process geometry.

The question is, does scaling transistors in an analog device make for a viable field programmable analog array?

In a field programmable device, transistors are used as switches to make the programmable interconnect. In a digital device, these switches contribute negatively by:

  • Increasing the size of the circuit
  • Reducing the noise margin
  • Increasing power consumption
  • Making the circuit run slower
  • Making the whole thing cost more

Additionally, as digital circuits scaled to smaller process nodes, the gate oxide thickness is reduced, which requires a lower operating voltage. In digital this wasn’t necessarily a bad thing, because a lower operating voltage means lower power consumption; and, since digital uses ones and zeroes, switching such circuits could accommodate a reduced noise margin.

Analog don’t scale like digital, baby. Yes, I know that some analog circuits can certainly take advantage of transistor scaling. Take those 80GHz ADCs for example (that was a joke). In general, a large majority of analog circuits benefit from the higher operating voltages that can be found on “more mature” thick oxide process nodes. The proof of this can be found by looking at how many analog components from the major manufacturers are still fabricated on 0.35micron and 0.18micron process nodes.

High-performance and high-precision analog circuits benefit from higher operating voltages because many of the best analog circuit topologies require the stacking of transistors (think diodes). DSM nodes, with their lower operating voltages, reduce the number of transistors that can be stacked, due to the diode drops.

Size does matter. For analog, bigger is better for the vast majority of circuits.

Why can’t I use transistors for switches in my programmable analog array? Just as in digital, the only problem with using transistors as the programmable interconnect in an analog array is that these switches make the circuit noisy, change its impedance, slow things down, consume excessive power, and make it large and expensive.

In a digital circuit, the transistor that acted as a switch was connecting two digital devices. The driving circuit only needed to drive a one or zero, and the receiving circuit simply needed to detect that one or zero. The fabric transistors in a digital FPGA could be relatively small but with large resistance.

Many useful analog circuits cannot tolerate large resistances in signal paths. Adding these large resistances coupled with the capacitance inherent in these circuits severely limits speed and analog bandwidth of such circuits.

If field programmable analog is such a bad thing, is there no way to lower the cost and simplify the development of analog ASICs? FPGAs greatly simplified the design of digital ICs and significantly reduced their development costs. Field programmable analog is too much of a performance and price negative to be viable in the marketplace.

Configurable analog is as close to programmable analog as you want to get.

The patent office makes a distinction between programmable and configurable ASIC solutions. They see programmable technologies as solutions that use active transistor switching that can be programmed in “the field.” Configurable technologies utilize vias that can be configured in “the fab.”

Via configurable analog is a better way to reduce cost and simplify the development of analog ASICs. Vias are the interconnect element used to connect different metal layers of all analog and digital ICs. Compared to the active switching transistors of an FPGA, vias are tiny. In fact, compared to just about anything in an integrated circuit, vias are tiny. Vias are small squares or plugs of low resistance material that goes between metal layers in all semiconductor devices.

A via configurable array consists of silicon-proven analog and digital resources that are overlaid with a global routing fabric. Wafers containing these arrays can be partially processed and staged at the foundry awaiting configuration. A user’s design is configured onto the array by a single mask layer change that places vias between metal layers in the global routing fabric.

  • Mask costs are amortized across many customers.
  • No full-custom layout costs.
  • It's low-risk because the underlying IP is well characterized and silicon-proven.
  • Processing time at the foundry is a couple of days versus three months.
  • Complete projects can go from kickoff to production-ready silicon in three to five months versus the 18 months typical of a fully custom development.
  • Complete respins can be accomplished, fabricated, and packaged in under four weeks.
  • Performance is excellent because the underlying IP is full-custom, spec-driven IP, and the interconnect uses standard metal and vias, not field programmable transistors.

In a nutshell, via configurable analog can reduce development cost and development time by 75 percent or more compared to traditional full-custom IC development. And via configurable mixed signal arrays have the performance and production costs that the market demands.

Today, companies in the defense, industrial, medical, consumer, and automotive markets are using via configurable mixed signal solutions in production volumes from hundreds to millions.

3 comments on “Field Programmable Analog – Technical Restraints, Part 2

  1. Mark Fortunato
    March 6, 2013

    While overall your point is well made, I would not say there is no market for programmable analog.  There is at least one company who has a decent amount of business with a line of products that have a microcontroller at the heart with field programmable analog.  While the Via configurable chips you talk about certainly will take far less time than full custom, field programmability is esentially instant gratification.  It takes minutes to go from a completed design to working silicon.

    So as usual in engineering, there are tradeoffs to be made.  Yes, field programmable parts will suffer in terms of noise, power, etc. compared to custom or via-configurable, but if it is good enough for an application, well, it is good enough.  You get no bonus points for perfoming better if the market does not need it.  So, once you have an application where a field programmable part's performance is good enough, it all comes down to cost and time to market with quantity of units built being a major factor.

    We have seen how FPGAs have developed a substantial market over the years with the only advantage being the virtually zero time from completed design to working silicon (and the related advantage of instant interim silicon) due to the field programability. 

    I expect, however, that analog wil not support nearly as large of a market because the range of products for which any particular prorammable chip will be “good enough” will be far smaller than for digital.  This is both because there are so many more critical requirements (noise, distortion, bandwidth, etc.) AND because analog simply calls for a greater range of functions (opamps, comparators, other amps, ADCs from 8 bits to 24 bits, from 100 sps to 1Gsps, DACs, references, etc., etc. etc.) all with a wide variety of specifications.

    It looks like your via-configurable technology provides “good enough” performance of the analog functions for a much wider range of applications than does programable technology.  The tricky art of having the right mix of configurable stuff which will satisfy a large enough number of applications, will be an issue whether field programmable or via-configurable.

    With all that said, it just means that, if I am right, the market size for programable analog is inherently much smaller than for FPGAs and that the different set of trade-offs makes the alternative of via-configurability potentially quite interesting and viable. (pun intended)

  2. jimfordbroadcom
    March 6, 2013

    I've seen these programmable analog chips come and go several times over the last 25 years.  What I haven't seen is a discussion of how much of the value and performance of analog is tied up in the proprietary semiconductor processes.  If you use a process that works well for say, op amps, it might not be great for making filters.  In order to produce a chip that can be used for multiple functions, the performance of any primitive is going to be compromised, to say nothing of the degradation due to the interconnect switches, as Reid pointed out.  The programmable analog I've seen had such lousy performance that nobody wanted to use it.  I'll have to take a closer look at Triad's via-configurable analog.  That could work.

  3. patrick_m
    March 19, 2013

    Hey Jim, I've seen many such techonlogies come and go too, but if I understand Reid's description correctly, the actual analog functions stay intact using proven analog IP, and only the vias change. So, in theory, the specific analog performance would stay intact. Reid, this correct?

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.