PE1-5.3 SDR Architectures & Digital Hardware

What this is and why it exists

Draw the ideal software radio and it is one line. An antenna, a converter, and a processor doing everything else.

Nobody builds that. Understanding exactly why is the most useful thing in this topic, because the reasons are quantitative and they explain every compromise real architectures make.

The vocabulary

• Digitisation point — where in the chain the signal is converted between analog and digital.
• Dynamic range — the ratio between the largest and smallest signals a converter can handle at once.
• Effective number of bits — the resolution a converter really achieves, after its own noise.
• Sample rate — how many conversions per second the converter performs.
• Intermediate frequency sampling — converting after one stage of mixing, rather than at the antenna.
• Direct conversion — mixing straight down to baseband before converting.
• Figure of merit — a measure combining a converter's resolution, speed and power.
• Blocker — a strong unwanted signal near the wanted one.

The mental model

The architectures differ in one decision. Where does the conversion between analog and digital happen? Push that point closer to the antenna and more of the radio is in software. Everything else follows from that choice.

Now take the ideal case seriously and see what it demands. A converter at the antenna must digitise every signal present, not only the one you want. That includes a distant weak station and a nearby powerful transmitter at the same instant. The ratio between them is enormous, so the converter needs a very large dynamic range.

It must also sample fast enough for the highest frequency present, which for a wideband antenna is very fast. And converter power consumption rises steeply with both resolution and sample rate. Ask for both at once and the power required is far beyond anything a handset could supply, and beyond what is available at any price.

That is why the ideal architecture is a useful fiction. It is not forbidden by physics. It is forbidden by converter technology and by the power budget, and stating the required numbers turns the argument from rhetoric into arithmetic.

Real architectures therefore keep some analog processing in front of the converter. A filter removes the signals you do not want, so the converter no longer has to accommodate them, and the dynamic range requirement collapses. A mixer brings the wanted signal down in frequency, so the sample rate requirement collapses too. Every practical architecture is a different choice about how much of each to do.

The derivation of minimum power consumption is the serious content here. Converter resolution, sample rate and power are related, and writing that relation down gives a floor. That converts a hand-wave about battery life into a number any proposed design can be measured against. If a design promises to beat that floor, something in it is wrong.

Handset and base station diverge at exactly this point, and for an obvious reason. One is in a cabinet on mains power with cooling available. The other is in a pocket. So a base station can afford a converter and a processing chain that a handset cannot, and can therefore digitise closer to the antenna. The list of viable digital processing options is genuinely different for the two cases. Comparing the two lists is the point of that pair of items.

Current technology limitations are worth stating plainly and worth reading with a date attached. This is the part of the subject most likely to have moved, because converter performance improves steadily. The reasoning does not change; the numbers do.

What you should now be able to explain or do

Say what distinguishes one architecture from another. Explain why a converter at the antenna needs an impossible dynamic range, using the strong-and-weak-signal argument. Say why the sample rate requirement is equally severe. Explain what filtering and mixing before the converter each relax. Use the power floor as a check on a proposed design, and say why handset and base station diverge.

Check yourself