Future of RF signal processing

[goldcoin]

http://www.edn-europe.com/news/14-bit-3-gsps-rf-sampling-adc-25k-each
Will RF processing be done only digitally, say, 10 years from now ?

[coppice]

Most of the processing in radio systems is already digital, but you can't eliminate every front end analogue operation. The question is really how much today's residual analogue will still be there in 10 years.

[voltz]

My own feeling is that front end band pass filtering could be eliminated in the future and we will see real one-chip solutions with very high performance. With the exception of maybe an isolation capacitor, the entire receiver and tx exciter could be in a single device.

[coppice]

My own feeling is that front end band pass filtering could be eliminated in the future and we will see real one-chip solutions with very high performance. With the exception of maybe an isolation capacitor, the entire receiver and tx exciter could be in a single device.

That is already the case for many radios. The OP's question was how much of that will still be analogue. It varies quite a lot right now, alrhough it can be hard to figure out from published material just what is going on inside many chips.

[mark03]

And how does/will this change when energy consumption is a principal constraint?

[coppice]

And how does/will this change when energy consumption is a principal constraint?

Digital scales, which analogue rarely does. At the finest geometries this tends to mean DSP wins on power.

[djacobow]

I'm pretty new to radio, so please take my questions as newbish.

As far as I understand, an SDR needs a front end filter for at least two reasons:

The first is aliasing. I guess if you sample fast enough, it will be easy to construct a filter that blocks everything above Nyquist, but is still super small and cheap. So I get that filters can be "eliminated" in the sense that they can disappear into a few 0402 SMD devices.

But the second is that there are often signals of _widely_ varying power in your receive bandwidth. If you just have a wideband front end you have to scale the RF before the ADC to a value that avoids clipping from the strongest signal -- no matter if that is anywhere near the frequency of your signal of interest or not. You might have a signal well outside your band of interest that is 100 dB bigger than what you're hunting. With a 16b ADC you literally can't hear the little one with the gain set not to clip the big one.

How do you overcome this without some kind of front end "real" filter, ie, the kind that can block energy from getting to your ADC?

[mark03]

But the second is that there are often signals of _widely_ varying power in your receive bandwidth. If you just have a wideband front end you have to scale the RF before the ADC to a value that avoids clipping from the strongest signal -- no matter if that is anywhere near the frequency of your signal of interest or not. You might have a signal well outside your band of interest that is 100 dB bigger than what you're hunting. With a 16b ADC you literally can't hear the little one with the gain set not to clip the big one.

How do you overcome this without some kind of front end "real" filter, ie, the kind that can block energy from getting to your ADC?

The way around this in a direct-sampling architecture (no downconversion prior to the ADC), is that the individual signals of interest are typically narrowband, at least when compared to the total receiver bandwidth.  In the digital domain, you will eventually decimate by a pretty large factor, and this will give you a bunch of extra bits---specifically, one bit per 4x decimation.  Thus, the true dynamic range of the receiver can be much greater than the dynamic range of the ADC.  At first blush this may not seem possible, because after all, wouldn't the weak signal fall entirely between ADC quantization levels and just disappear?  What saves you is that "everything else" including other signals plus noise, act as dithering for the weak signal you're pulling out.

[voltz]

My own feeling is that front end band pass filtering could be eliminated in the future and we will see real one-chip solutions with very high performance. With the exception of maybe an isolation capacitor, the entire receiver and tx exciter could be in a single device.

That is already the case for many radios. The OP's question was how much of that will still be analogue. It varies quite a lot right now, alrhough it can be hard to figure out from published material just what is going on inside many chips.

Actually his question was will this all be digital in 10 years not how much will be analog, but never mind that.
My answer is yes, it will be. Even today's SDR technology uses band pass filtering in order to reject unwanted signals from other bands. With a high enough dynamic range of ADC, you dont need them. And that technology is almost already here. Now add to that transmit within the same chip and you have a full blown transceiver in a chip with no RF analog components. We're not there quite yet. Rf designers become software engineers.. Or at least work closely with them. Thats how i see it coming.

[galvanix]

Rf designers become software engineers.. Or at least work closely with them. Thats how i see it coming.

Currently applying for internship positions as an EE student with strong interest in RF & DSP and most positions I see (not only internships) at least where I'm looking are embedded DSP and SDR and software defined networking. Very few RF hardware design positions which is a shame, because when there aren't any open positions for RF engineers then there's zero chance you'll get to work on it as an intern.

[radiogeek381]

I do a lot of work with SDR.  Especially in applications that involve very weak signals.

It is not clear to me that "direct sampling" will ameliorate all problems with out-of-band
interference, or even with near-channel interference.  Dynamic range cannot correct for
intermodulation products that, I suspect, are inevitable results of the fact that even in
a high-speed ADC, the signal will pass through a semiconductor of some sort.  Where
there is a non-linear device (that is, a junction) there will be mixing products.   

Even with a 32 bit ADC, the first diode in the circuit will be happy to mix incoming signals,
and if two of them are large enough and at the right frequency, you'll see them alongside
your target signal.

RF experts: help me out here....  The next point has been eating at me for a while

I'm concerned that very wide front ends will run afoul of Mr. Boltzmann.  Thermal
noise power in a receiver is proportional to kTB where B is bandwidth.  For terrestrial systems,
the noise floor is -174dBm + 10*log10(B).  So a 40 MHz wide input has about 76 dB of excess
noise.  We buy much of that back with "coding gain" and other mechanisms, but the problem gets
worse.  Your magic front end that's 400 MHz wide has a noise floor of -88dBm.  That's a lot to
make up, and it isn't clear to me that resampling and integration is going to get us back to where
we'd be with a narrowband front end and a lower sampling rate. What am I missing here?

-----------------------

In any case, the intermod problem is likely to be determinative -- there's no avoiding the fact
that there will be really really big signals out there just waiting to beat together and swap
that tiny signal from the moose tracking collar in the woods of northern Vermont.

Sivan Toledo et al. at http://dx.doi.org/10.1016/j.icte.2017.01.002 describe a SAW filter
front end for an Ettus USRP N200/WBX combination that is apparently working for them.  This is
a case where a filter on the order of 1MHz wide is placed in front of the 40MHz wide ADC to
improve out-of-band rejection.

--------------------------------

I believe that RF design will not die. It may split into two areas (as has my own field
of computer design)

1. A segment where most designs are executed by engineers with a broad focus
who rely on CAD tools and abstract languages that free the designer
from detailed considerations.  The tools reduce the design cycle and design cost at
the expense of optimality or maximum performance.  For much of the world, maximum
performance is too much.  Nobody needs a keyfob receiver with a 0.5dB noise figure,
an IP3 of 30dB, and a 140dB dynamic range.

2. A very small segment where every little bit of performance matters. These designs
will be executed by a very small cadre of hardcore engineers who use the same
CAD tools but also occasionally resort to "flat rock" design techniques (where we
start building western civilization with a stick and a flat rock).  These engineers are
going to design the ultra high data rate links that work in the presence of noxious
interferers and the ultra low-noise receivers that gather whispers among the stellar
background noise.

[KE5FX]

I'm concerned that very wide front ends will run afoul of Mr. Boltzmann.  Thermal
noise power in a receiver is proportional to kTB where B is bandwidth.  For terrestrial systems,
the noise floor is -174dBm + 10*log10(B).  So a 40 MHz wide input has about 76 dB of excess
noise.  We buy much of that back with "coding gain" and other mechanisms, but the problem gets
worse.  Your magic front end that's 400 MHz wide has a noise floor of -88dBm.  That's a lot to
make up, and it isn't clear to me that resampling and integration is going to get us back to where
we'd be with a narrowband front end and a lower sampling rate. What am I missing here?

That -88 dBm noise floor is only a concern if your signal bandwidth is 400 MHz.  If your signal is 4 MHz wide, congrats, now you have a -88 - 10*log10(400/4) = -108 dBm noise floor.  If your signal is 4 kHz wide -- more than enough to carry toll-quailty voice or even good-quality music -- your noise floor is -138 dBm. 

"Coding gain" achieved through digital filtering and decimation is not a trick, a hack, or a marketing term to be surrounded by finger-waving quotes.  It's simply a restatement of natural law that says equal amounts of bandwidth contain equal amounts of noise.  A hypothetical ADC with a 0-degree noise figure could do just what your example suggests.  (Subject to its own nonlinearities, of course, which will ultimately not be much worse than any other front end option you'd have as a receiver designer.)

In any case, the intermod problem is likely to be determinative -- there's no avoiding the fact
that there will be really really big signals out there just waiting to beat together and swap
that tiny signal from the moose tracking collar in the woods of northern Vermont.

From a conceptual point of view, an ADC contains switches and capacitors, not PN junctions.  As long as the signal doesn't exceed the ADC's input range -- typically a volt or two peak-to-peak -- the out-of-band racket actually helps the ADC do its job.  It acts as a dither source. 

If signals larger than that are going to be present, then yes, preselection and/or attenuation will be needed... just as it would in an old-school Harris or W-J or Collins rig.

2. A very small segment where every little bit of performance matters. These designs
will be executed by a very small cadre of hardcore engineers who use the same
CAD tools but also occasionally resort to "flat rock" design techniques (where we
start building western civilization with a stick and a flat rock).  These engineers are
going to design the ultra high data rate links that work in the presence of noxious
interferers and the ultra low-noise receivers that gather whispers among the stellar
background noise.

This much is absolutely true, the radio astronomy folks are decades in front of the commercial world.  You won't see photonic LOs and SIS mixers in your iPhone for a while.

The thing about RF is that it can be modeled perfectly.  There's nothing we don't understand about it, which means there's nothing a piece of software can't understand about it.  On the commercial side of things, only small shops will continue to need Jedi-level RF expertise.  Larger shops that can afford the best simulation tools will just treat the whole RF problem as a big black box.  Software will pick the components, model the layout, predict the outcomes of performance and compliance tests with ever-increasing accuracy, and spit out the bill of materials.  Unpleasant last-minute surprises like Apple's "antennagate" will be a thing of the past. 

The same software that does all that stuff will also help the engineering team deal with SI issues in the digital parts of the circuit, because it's all the same problem.

Ultimately there will be a few people designing chips and a lot of people using them.  The latter will have no idea what's going on under the epoxy, not because they're hopeless dumbasses but because they simply don't need to.  It'll either be a depressing future or an exciting one, most likely some of both.

[coppice]

I'm concerned that very wide front ends will run afoul of Mr. Boltzmann.  Thermal
noise power in a receiver is proportional to kTB where B is bandwidth.  For terrestrial systems,
the noise floor is -174dBm + 10*log10(B).  So a 40 MHz wide input has about 76 dB of excess
noise.  We buy much of that back with "coding gain" and other mechanisms, but the problem gets
worse.  Your magic front end that's 400 MHz wide has a noise floor of -88dBm.  That's a lot to
make up, and it isn't clear to me that resampling and integration is going to get us back to where
we'd be with a narrowband front end and a lower sampling rate. What am I missing here?

You might be missing that if the actual channel is less than 400MHz wide, the digital processing should be filtering the signal down the to actual bandwidth, removing the out of band noise in that processing (as KE5FX indicated). There is no AWGN problem with a wideband front end, there is only a linearity constraint. Just how linear can you make a super wide band ADC? Can you really keep all the IM products low enough to allow downstream digital processing to do a good job?

[radiogeek381]

That -88 dBm noise floor is only a concern if your signal bandwidth is 400 MHz.  If your signal is 4 MHz wide, congrats, now you have a -88 - 10*log10(400/4) = -108 dBm noise floor.  If your signal is 4 kHz wide -- more than enough to carry toll-quailty voice or even good-quality music -- your noise floor is -138 dBm. 

Thank you.  Intuitively I should have known that, as my own SDR stuff has a 40MHz pre-ADC filter... 

From a conceptual point of view, an ADC contains switches and capacitors, not PN junctions.  As long as the signal doesn't exceed the ADC's input range -- typically a volt or two peak-to-peak -- the out-of-band racket actually helps the ADC do its job.  It acts as a dither source. 

Conceptually and ADC contains switches.  But it is likely the switches are made from some
kind of semiconductor.  And they're connected to pads.  I've worked on I/O pad designs in CMOS.
They all have a P/N junction or two (or three or four).  Some are inherent in the switches (can't
make a FET without making a diode, and most design rule sets discourage direct gate connections
because of the opportunity for punch-through).  Others are explicit (gate protection from reflections,
bad designs, or ESD).  So, I'd still be concerned that there are mixing opportunities. 

Can they be better managed than external discrete solutions?  Perhaps not.  And you certainly make
that point.

Thanks for the response. 

[mark03]

Ultimately there will be a few people designing chips and a lot of people using them.  The latter will have no idea what's going on under the epoxy, not because they're hopeless dumbasses but because they simply don't need to.  It'll either be a depressing future or an exciting one, most likely some of both.

A tad depressing for this radio amateur.  I always wanted to learn the black arts of RF design, but lack the spare time.  As a signal-processing guy, it's exhilarating that more and more of this domain is becoming accessible to us theoreticians via software, but I wonder if there will be any RF "old hands" still living by the time I retire.  Maybe my problem is just nostalgia...

[rfeecs]

Take the front end of a cell phone radio or wifi radio.  You have filters, switches, power amplifiers and possibly low noise amplifiers.  These are inherently analog.  They are not going away.

Then you have the transceiver containing oscillators, mixers, more filters and amplifiers.  This is often now already integrated onto the same chip as the baseband digital signal processor.  You could probably get rid of the transceiver and go to direct sampling.  But what is the advantage?

So the answer is no.  Ten years from now there will still be at least some analog RF.

[G0HZU]

Ultimately there will be a few people designing chips and a lot of people using them.  The latter will have no idea what's going on under the epoxy, not because they're hopeless dumbasses but because they simply don't need to.  It'll either be a depressing future or an exciting one, most likely some of both.

At my place of work we predicted this scenario in the early 1990s. I was part of a team of engineers designing early SDR type equipment for gov/mil use and I was involved in the design of the RF converter (GHz front end down to the digital IF). Back in those days we had to design wideband synthesisers without the modern ICs we have today and also the signal path design was tougher for the same reason. But as each year passed, the RFICs got better and better and the digital IF got higher in frequency and wider as the ADC/DSP became more advanced. The joke at the time was that the RF team was slowly being replaced with integrated circuits. We could see it happening but we still joked about it.

Fast forward to today and pretty much anyone who can program an MCU and layout a reasonable circuit board can produce a fairly decent RF converter using modern chipsets. It won't have the high (RF) performance of our old discrete (and very expensive) downconverters in terms of spurious free dynamic range etc but it will cost maybe 100th the price and the NRE cost/time will be very low in comparison.

They just need to read the datasheet and they don't 'need' to know what goes on inside the chip. At its peak, our RF dept had something like 18 RF engineers across all grades. Today we have 4 RF engineers (I'm one of them) but we still have dozens of DSP/SW engineers and this tells its own tale. The company mainly wants software/DSP skills these days as our RF resources dwindle away. I think the spec requirements for spurious free dynamic range and the requirements for phase noise are gradually being traded against size/power/weight/cost and this means the designs inevitably lean towards the chipset approach and the customers often seem willing to accept the tradeoff.

[KE5FX]

At my place of work we predicted this scenario in the early 1990s. I was part of a team of engineers designing early SDR type equipment for gov/mil use and I was involved in the design of the RF converter (GHz front end down to the digital IF). Back in those days we had to design wideband synthesisers without the modern ICs we have today and also the signal path design was tougher for the same reason. But as each year passed, the RFICs got better and better and the digital IF got higher in frequency and wider as the ADC/DSP became more advanced. The joke at the time was that the RF team was slowly being replaced with integrated circuits. We could see it happening but we still joked about it.

Fast forward to today and pretty much anyone who can program an MCU and layout a reasonable circuit board can produce a fairly decent RF converter using modern chipsets. It won't have the high (RF) performance of our old discrete (and very expensive) downconverters in terms of spurious free dynamic range etc but it will cost maybe 100th the price and the NRE cost/time will be very low in comparison.

They just need to read the datasheet and they don't 'need' to know what goes on inside the chip. At its peak, our RF dept had something like 18 RF engineers across all grades. Today we have 4 RF engineers (I'm one of them) but we still have dozens of DSP/SW engineers and this tells its own tale. The company mainly wants software/DSP skills these days as our RF resources dwindle away. I think the spec requirements for spurious free dynamic range and the requirements for phase noise are gradually being traded against size/power/weight/cost and this means the designs inevitably lean towards the chipset approach and the customers often seem willing to accept the tradeoff.

It's been a common-enough prediction for a long time, right up there with fusion and flying cars, but I think it's actually starting to happen.  If you ask technical questions about chips like the AD9361 over on the Analog Devices support forum, the answer will often be "Don't worry about that, just use our driver."  As far as I'm concerned that's the depressing part... but at the same time, I understand that it's a valid evolutionary path for technology to follow.  Probably the only valid path.

[goldcoin]

Will analog rf circuits like amplifiers, filters, oscillators, mixers still be necessary for RF ICs that handle very high frequencies ? 

[dmills]

The elephant in the room with a lot of this stuff is the realization that the ADC IS a mixer... And that this implies that the usual reciprocal mixing issues apply. 

Sure I can digitize a whole swathe of bandwidth, but if I do it just implies that there is a whole lot of bandwidth for a LO spur or LO phase noise (Where the LO is the sample clock) to mix with and put a tone just where I don't want it.  Good receiver design is still about stripping away as much as you can of what you don't need as early as possible, ideally before you hit the first mixer (Aka the ADC). Reduce the bandwidth into the first mixer ADC and you reduce both the possibility of overload and the risk of some noise on the clock input mixing with something in the input to wipe out your wanted signal. ADC clocks, in spite of what the data sheets say, have NO NOISE MARGIN around the switching point. 

With subsampling ADCs, clock phase noise is a huge issue as it relates directly to sampling jitter in a way that gets horribly worse as input frequency increases, low noise LO design is very much a thing, and wideband PLLs with the dividers and VCO on the same die really do not help, but sure are convenient when a slightly deaf receiver is acceptable.

The 'digital' receiver chipsets are getting good enough that you can build very cheap radios with ok performance with them, but at UHF and up where sky noise is not much of a limit, if you want a receiver that is within a few dB of the aerial thermal noise, there is still plenty of art to it.

The threshold at which you need to get the engineering on is rising, just like everything else, but when performance really matters you still need a grey beard.