RF Receiver: Noise Figure, IIP3 & Yield Analysis

The Fundamental Cascade Tradeoff

Every RF receiver designer knows the Friis formula: the first stage dominates the cascaded noise figure (NF), so you should put the best (lowest NF) amplifier first and make its gain as high as possible. The math is elegant — almost deceptively simple.

What the formula doesn't immediately reveal is the tension it creates with linearity. High gain in early stages amplifies signals before they reach the linearity-limited components downstream. The IIP3 cascade formula — 1/IIP3_total = Σ G_cumul/IIP3_i — shows the opposite dependency: each stage's IIP3 contribution gets amplified by all the gain that precedes it. Add a 20 dB LNA up front and suddenly your mixer's IIP3 must work against 100× the input signal power. You've improved noise performance but potentially crippled linearity.

This is the classic receiver design knife-edge. You can't just throw gain at the front end and call it done.

This post walks through a Ku-band receiver design using the RF Cascade Analyzer, showing how to navigate this tradeoff in practice. More importantly, we'll see why the nominal design — which looks perfectly fine on paper — fails a manufacturing yield requirement when you actually run the Monte Carlo analysis. Most engineers skip this step and regret it later when production units start failing acceptance tests.

The Reference Chain

The receiver chain we're analyzing is a 6-stage Ku-band front-end for a VSAT application. Nothing exotic, just a representative design you'd see in a satellite ground terminal:

The LNA provides the initial low-noise amplification — 15 dB gain with 1.5 dB NF is typical for a decent Ku-band part. The bandpass filter follows to reject out-of-band interference before downconversion. The mixer has conversion loss (−7 dB) and relatively poor noise figure (8 dB), which is normal for mixers. After mixing down to IF, we have an IF amplifier providing 20 dB of gain, another filter for selectivity, and finally an ADC driver to interface with the digitizer.

Paste this JSON into the tool with NF spec = 6 dB, gain spec = 28 dB, IIP3 spec = −8 dBm. These specs are reasonable for a VSAT application — 6 dB system NF is fairly relaxed, 28 dB gain is moderate, and −8 dBm IIP3 is tight but achievable.

Reading the Cascade Table

After clicking Run Analysis, the cascade table shows cumulative metrics at each stage. This is where you see how the system performance evolves as the signal propagates through:

System NF of 2.5 dB looks excellent — well within the 6 dB spec with 3.5 dB of margin. You could probably use a much worse LNA and still meet the requirement. But look at the IIP3 column. The input-referred IIP3 starts at −5 dBm (just the LNA) and degrades as we add stages. By the time we reach the ADC driver, system IIP3 has dropped to −8.0 dBm. That just barely meets the −8 dBm specification with essentially zero margin.

This should immediately raise a red flag. A nominal design sitting right on the spec limit is asking for trouble in production.

NF Sensitivity Analysis

The sensitivity bar chart reveals something you probably already suspected from Friis: the LNA contributes 89% of the system NF. The BPF adds about 5%, and everything downstream contributes less than 5% combined. This is Friis in action — 13.5 dB of gain before the mixer suppresses the mixer's 8 dB NF contribution to less than 0.1 dB system impact.

The practical implication is clear: if you need to reduce system NF below 2.5 dB, you must improve the LNA. Nothing else matters. Swapping in a better mixer with 6 dB NF instead of 8 dB? You'd save maybe 0.05 dB of system NF. Not worth the BOM cost. Conversely, if cost pressure requires using a worse mixer — say, 12 dB NF instead of 8 dB — the impact is negligible. The gain preceding it buries the contribution.

This is why experienced designers obsess over the first stage and often treat everything after the first 15–20 dB of gain as relatively forgiving from a noise perspective. You've already won or lost the noise battle by then.

Why the IIP3 Is Dominated by the IF Amplifier

Wait, didn't we just say the first stage dominates? That's true for noise, but linearity tells a different story.

The Friis IIP3 cascade table (from the tool's system summary) shows the contributions:

• LNA: contributes 72% of 1/IIP3_total (15 dBm IIP3 seen from the output, but −5 dBm referred to input)
• Mixer: contributes 18% (12 dBm IIP3, but 6.5 dB of gain in front)
• IF Amp: contributes 9% (10 dBm IIP3, but 6.5 dB of gain in front)

The LNA still dominates, but not as overwhelmingly as it did for noise. It contributes 72% rather than 89%. Why? Because its IIP3 of −5 dBm is input-referred — it has no gain ahead of it to suppress its contribution. The mixer and IF amp each have 6.5 dB of gain preceding them, which means the IIP3 formula weighs their contributions by about 4.5× in linear terms. But their own IIP3 values are much higher (+12 dBm and +10 dBm respectively), so the net effect is more moderate.

Here's the key insight: to improve system IIP3, the highest-leverage fix is improving the LNA's IIP3. A 3 dB improvement to the LNA IIP3 (from −5 to −2 dBm) improves system IIP3 by approximately 2.5 dB. That confirms LNA dominance, but notice it's not a 1:1 improvement like you might naively expect. The other stages are contributing enough that you don't get the full 3 dB back.

If you improved the IF amp's IIP3 by 3 dB instead, you'd see maybe 0.3 dB system improvement. That's why the sensitivity analysis matters — it tells you where your engineering effort actually pays off.

The Monte Carlo Surprise

So far, the nominal metrics all pass. NF is 2.5 dB against a 6 dB spec. Gain is 30.5 dB against a 28 dB spec. IIP3 is −8.0 dBm against a −8 dBm spec (okay, that one's tight). On paper, you'd sign off on this design and send it to production.

But then you run the Monte Carlo analysis with realistic component tolerances: gain ±0.5 dB σ, NF ±0.3 dB σ, IIP3 ±2 dB σ. These aren't pessimistic numbers — they're typical datasheet tolerances for commercial RF components. Run 50,000 trials and look at what comes back:

• NF yield (≤6 dB): 99.8% — easily passing, as expected with 3.5 dB margin
• Gain yield (≥28 dB): 94.2% — passing but tighter than you might have expected given 2.5 dB nominal margin
• IIP3 yield (≥−8 dBm): 52.3% — failing badly
• Overall yield: 51.8%

Only half of manufactured units meet all three specs simultaneously. You've just designed a coin-flip receiver.

The problem is the IIP3 tolerance. With ±2 dB σ on each stage's IIP3, and the LNA sitting near the boundary at −5 dBm nominal, the distribution of system IIP3 spans from roughly −11 dBm to −5 dBm. The −8 dBm spec sits near the median of this distribution — exactly half the units fail. This is what happens when you design to nominal values without accounting for the statistical reality of component variation.

The NF yield is fine because you had 3.5 dB of margin and NF tolerances are tight (±0.3 dB σ). The gain yield is decent because ±0.5 dB tolerances on six stages don't accumulate too badly. But IIP3 tolerances are large (±2 dB σ is typical for active components), the spec is tight, and you had zero margin nominally. Recipe for disaster.

The Fix

Three options appear immediately, each with different cost and risk tradeoffs:

Option 1: Tighten the LNA IIP3 spec. Require the LNA's IIP3 to be −3 dBm minimum instead of −5 dBm typical. In statistical terms, you're asking for −3 dBm at p5 (5th percentile) rather than accepting −5 dBm as the mean. This shifts the system IIP3 distribution up by roughly 2 dB, raising IIP3 yield to approximately 88% and overall yield to something acceptable.

The downside? You're now specifying a part at the edge of its distribution, which means either paying more for a premium-binned component or accepting lower supplier yield (which they'll pass along as higher cost anyway). But it works.

Option 2: Relax the system IIP3 specification. If the −8 dBm requirement was derived with some conservatism — maybe the link budget analysis assumed worst-case interference that's unlikely in practice — the actual minimum acceptable IIP3 might be −10 dBm. At a −10 dBm spec, IIP3 yield rises to 82% and overall yield jumps to 80%. Much better.

This is often the right answer if you can negotiate it with the system architect. Specs tend to accumulate margin-on-margin as they flow down from system to subsystem to component, and sometimes you can recover some of that conservatism when you see the actual statistical distribution.

Option 3: Redesign the first stage. Replace the LNA + BPF combination with an integrated front-end component that achieves −1 dBm IIP3. Some modern integrated solutions offer this, though you'll pay for it. System IIP3 improves to approximately −3 dBm nominal, and yield rises above 95%. You've bought margin with money, which is sometimes the cleanest solution.

The Monte Carlo analysis makes the right intervention obvious in a way that nominal analysis never can. Without running the statistics, you'd ship this design, discover the yield problem in production, and then scramble for a fix under time pressure. Ask me how I know.

Key Rules From This Analysis

A few lessons crystallize from this exercise:

Write component specs against the p5 Monte Carlo curve, not nominal values. A component sitting at its nominal IIP3 is at the median of its distribution — half the production units will be worse. If your system spec requires nominal component performance, you've designed a 50% yield product. Specify components at their p5 or p10 values (5th or 10th percentile) to achieve acceptable system yield. Yes, this costs more. That's the price of actually meeting specs in production. IIP3 yield requires more margin than NF yield. IIP3 tolerances (±2 dB σ is typical) are much larger than NF tolerances (±0.3 dB σ), and IIP3 specs are typically tighter relative to nominal margin because linearity is harder to achieve than low noise. If you have 1 dB of nominal margin on IIP3, you probably don't have enough. If you have 1 dB of nominal margin on NF, you're likely fine. The statistics are different. The sensitivity analysis tells you where to spend BOM budget. When the analysis shows 89% NF contribution from the LNA, that means a better mixer buys you nothing for noise performance. Save the money. When it shows 72% IIP3 contribution from the LNA, that means a more linear LNA directly improves system linearity. That's where the budget should go. Don't waste money improving components that contribute 2% to the system metric you're struggling to meet.

The cascade analyzer turns these tradeoffs from vague intuition into quantitative decisions. Use it early in the design cycle, not after you've already committed to a BOM and discovered the yield problem in production. Your future self will thank you.