wireless communication
RSSI explained: why -90 dBm can beat -80 dBm
A signal-strength reading is a power meter, not a quality meter. Here is the arithmetic that decides which of two nodes has the better link.
Two meters, forty nodes, one reading that sent a crew up a tower
The call came from a water utility: forty meters in a basement car park, one gateway on the roof of the pump house. Commissioning was three weeks late. The hand-held survey tool showed every meter between -95 and -70 dBm, which on paper is comfortable for a 470 MHz link. Two meters were losing a third of their packets. The two with the strongest readings, -80 dBm and -78 dBm, were the worst of the lot.
The crew replaced the antenna on the gateway. Nothing changed. They raised the gateway by four metres. Nothing changed. The reading that looked healthiest was the one that hid the fault.
That is the whole problem with received signal strength: it is a power meter, not a quality meter. It tells you how much energy is arriving in the channel. It says nothing about whether any of that energy is the packet you are waiting for. On a shared 470–510 MHz band, with five hundred 200 kHz channels and neighbours you did not install, the two are frequently different things.
This article walks through the arithmetic behind that difference, using the sensitivity and power figures our own radios publish, so you can decide what a reading means instead of ranking nodes by it.
What an RSSI number actually is
An RSSI value is the total power the receiver sees inside its detection bandwidth, sampled over a short window, usually during the preamble of a packet. Thermal noise, man-made noise, the neighbour's carrier on the same channel, the switching hash from a nearby VFD, and — somewhere in there — your signal. All of it summed, then reported as one integer in dBm.
Two consequences follow immediately.
First, a strong reading can be manufactured by interference. A co-channel transmitter 6 dB above your packet raises the meter by roughly 7 dB while doing nothing for your decode margin. In fact it does the opposite: the wanted signal has to compete with a stronger carrier inside the same passband, and the detector sees a worse SNR than the number suggests.
Second, the value is close to meaningless as a cross-device comparison. Different silicon, different bandwidth settings, different calibration offsets, different points in the receive chain where the sample is taken. The literature is blunt about this. A 2015 paper on PHY-layer link quality indicators notes that the RSSI signal "can be used to flag activity or inactivity in the medium, but it cannot distinguish between different transmitters or interferers, and as a result, the RSSI signal cannot by itself be used to track the states of multiple links" (arXiv:1508.04524). A 2026 study across fifteen outdoor environments and four low-power radios reaches the same conclusion from the deployment side: RSSI is "highly sensitive to environmental and internal factors", and the variation is poorly correlated with actual link quality (arXiv:2603.21107).
So the useful question is not "which node reads higher". It is "how far is this node above its own decode threshold, and what is the noise floor underneath the reading".
Sensitivity is a function of data rate, and nobody prints the table
Here is the part that turns intuition upside down. Receiver sensitivity is not a property of a radio. It is a property of a radio at a given data rate and modulation. Halve the rate and you halve the noise bandwidth you have to sit in, which buys roughly 3 dB of sensitivity. Quarter it and you buy 6 dB. The relationship is close to pure arithmetic:
Pmin(rate) = Pmin(ref) - 10 · log10(rate / rate_ref)Take the -116 dBm sensitivity that TP5603-D01L10, TP5510 and TP5500 publish at 19.2 kbps, and scale it across the rates the TP5603 actually supports:
Data rate | Computed sensitivity | RSSI needed for a 10 dB fade margin |
|---|---|---|
1.2 kbps | -128.0 dBm | -118.0 dBm |
2.4 kbps | -125.0 dBm | -115.0 dBm |
4.8 kbps | -122.0 dBm | -112.0 dBm |
9.6 kbps | -119.0 dBm | -109.0 dBm |
19.2 kbps | -116.0 dBm | -106.0 dBm |
76.8 kbps | -110.0 dBm | -100.0 dBm |
Two footnotes, both important. The anchor -116 dBm @ 19.2 kbps is a datasheet value; the rest of the column is arithmetic from it, under the assumption that the receiver's noise figure and its required Eb/N0 do not change with rate. That assumption is standard for narrowband FSK, but it is an assumption — expect a dB or two of error at the extremes, where filters and synthesiser settling start to matter. The 10 dB column is a site-planning convention, not a specification.
Now the two nodes from the car park, read properly:
Node | Rate | RSSI reading | Threshold at that rate | Margin |
|---|---|---|---|---|
A | 76.8 kbps | -80 dBm | -110.0 dBm | 30 dB |
B | 2.4 kbps | -90 dBm | -125.0 dBm | 35 dB |
Node B reads 10 dB weaker and has 5 dB more headroom. Its packets are slower, so it sits in a narrower slice of spectrum and needs less energy per bit to decide what a symbol was. If both nodes carry the same four-byte meter reading, node B is the more robust engineering choice — and it will lose on any spreadsheet that ranks links by RSSI.
That is the first, and most common, reason a -90 dBm node outperforms a -80 dBm node: the two numbers were never on the same scale.
De-embedding: turning a reading into something you can act on
If RSSI is total power, the wanted signal can be recovered arithmetically provided you know the noise floor. Measure it on an idle channel — same channel, same bandwidth, no packets in flight — then subtract powers:
S = 10 · log10( 10^(RSSI/10) - 10^(N/10) )
SNR = S - NWith an assumed idle floor of -104 dBm in 19.2 kbps mode, the same reading maps to wildly different realities:
RSSI reading | Implied signal | Implied SNR | Against a 15.2 dB decode threshold |
|---|---|---|---|
-100 dBm | -102.2 dBm | 1.8 dB | 13.4 dB short |
-95 dBm | -95.6 dBm | 8.4 dB | 6.8 dB short |
-90 dBm | -90.2 dBm | 13.8 dB | 1.4 dB short |
-86 dBm | -86.1 dBm | 17.9 dB | 2.7 dB spare |
-80 dBm | -80.0 dBm | 24.0 dB | 8.8 dB spare |
Where does 15.2 dB come from? From the datasheet anchor itself. Thermal noise at 290 K is -174 dBm/Hz; at 19.2 kbps the noise in the information bandwidth is -174 + 10·log10(19200) = -131.2 dBm. The published sensitivity of -116 dBm is 15.2 dB above that, which is the receiver noise figure plus the Eb/N0 the demodulator needs. Every row of the table is that single number applied to a different input.
Notice the -90 dBm row. At 19.2 kbps, with this noise floor, it is 1.4 dB below the threshold — the reading looks respectable and the link is dead. At 2.4 kbps the same -90 dBm reading is 23 dB above a noise floor that dropped to -113 dBm with the narrower bandwidth, and the link has 7.8 dB to spare. Same number, opposite verdicts, decided entirely by context the meter refuses to show you.

What our own hardware publishes, and what it does not
It is worth being explicit about which numbers here are facts and which are arithmetic:
Model | Published sensitivity | Transmit power | Rates | Form factor |
|---|---|---|---|---|
-116 dBm | 17 dBm, configurable | 1.2 – 76.8 kbps | DIN-rail DTU, RS485, ≤1 W | |
-116 dBm @ 19.2 kbps, 470 MHz | 26 ± 1 dBm | 2.4 / 19.2 / 76.8 kbps | IP65 outdoor, 100 terminals | |
-116 dBm @ 19.2 kbps | 26 ± 1 dBm | 2.4 / 19.2 / 76.8 kbps | IP65 outdoor | |
-117 dBm | 17 dBm | UL 2.4 / 76.8, DL 19.2 / 76.8 kbps | 44-pin LCC module, FSK |
Note what is absent: no RSSI accuracy figure, no register scaling, no statement of whether the reported value is measured before or after the LNA. That is normal for this class of hardware and it is exactly why the reading should be treated as an ordinal indicator — trend, not truth — unless you have characterised the specific radio in front of you.
The TP1107 module, at -117 dBm, illustrates the point about comparisons. One dB better than a TP5510 on the datasheet, and that difference is within the unit-to-unit spread of either device. Do not build a design decision on it.
Four failure modes that hide behind a strong reading
The car park incident was not exotic. It was one of four patterns that account for most "strong signal, bad link" tickets.
Symptom | What the meter shows | What is actually happening | First check |
|---|---|---|---|
Packet loss on the node with the best reading | High RSSI, unstable, jumps several dB packet to packet | A co-channel interferer inside the passband; the reading is the neighbour, not your packet | Scan the channel with the node silent; compare idle floor to busy floor |
Whole site degrades when one node transmits | RSSI rises on other nodes' reports | Front-end blocking / AGC desense from a strong nearby carrier, out of band but inside the preselector | Move the aggressor or lower its power; re-measure PER, not RSSI |
Good readings indoors, poor at the far end of a run | RSSI within 2–3 dB of threshold | Link is marginal by design; a 6 dB fade removes it half the time | Compute margin per rate, not per reading |
Reading varies 8–10 dB between consecutive packets | Any value, any time | Rayleigh fading; a single sample is one draw from a distribution, not the mean | Log 100 packets, use the 10th percentile |
The first row is the one that breaks people's intuition. An interferer does not have to be strong in absolute terms to ruin a link that is already thin. In the worked example below, a node with 2.8 dB of margin is destroyed by an interferer sitting below its own noise floor.

Worked example: 40 meters, 800 metres, two nodes
Take the basement again and put numbers on it. Assumptions are marked; the product figures are from the table above.
Site. Gateway is a TP5510 at 26 dBm into a 2 dBi antenna with 1.5 dB of feeder loss. Nodes are TP5603 DTUs at 17 dBm. Longest hop is 800 m through two concrete walls, assumed 15 dB total wall loss. Free-space loss at 470 MHz over 800 m is 84.0 dB:
FSPL = 20·log10(800) + 20·log10(470) - 27.55 = 84.0 dB
Path = 17 + 0 - 15 - 84.0 = -82.0 dBm at the gateway antennaA -82 dBm uplink is 34 dB above the 19.2 kbps threshold. That is the kind of number that makes a commissioning report look finished. Now split the fleet.
Node A runs at 76.8 kbps because someone wanted faster polling. Its idle floor at that rate is -98 dBm (the -104 dBm reference shifted by +6 dB for the wider bandwidth — assumption). It reports -80 dBm, and a neighbour's carrier on the same channel is 6 dB stronger than its own signal. De-embedding the wanted signal:
S = -80 - 10·log10(1 + 10^0.6) = -87.0 dBm
SNR = -87.0 - (-98.0) = 11.0 dBThe threshold is 15.2 dB. Node A is 4.1 dB short and will drop a large share of frames — while showing the best number in the fleet. Even on a clean channel it would have only 2.8 dB of margin, which a single fade event consumes.
Node B runs at 2.4 kbps. Floor -113 dBm, reading -90 dBm, no neighbour:
S = 10·log10(10^-9 - 10^-11.3) = -90.0 dBm
SNR = -90.0 - (-113.0) = 23.0 dB7.8 dB of margin. It tolerates a co-channel interferer up to 7 dB above its noise floor (-106 dBm) before the link degrades. Node A tolerates almost nothing: at 2.8 dB of margin, an interferer just 0.4 dB below its noise floor is enough.

The engineering conclusion is uncomfortable but simple: the polling rate decision, not the antenna, decided which nodes were reliable. Dropping node A from 76.8 to 4.8 kbps would have bought 12 dB of sensitivity — more than any practical antenna change — at the cost of slower polling. If you are sizing a network, that trade belongs in the plan; see the throughput side of the same trade in TPUNB at 100 nodes.
A short routine for the next site visit
Do these five things in order. They take twenty minutes and they replace the RSSI league table with something you can defend.
- Measure the idle floor per rate. Silence the nodes, read the channel in each data rate you plan to use, write the numbers down. Every later reading is judged against these, not against zero.
- Fix the rate before comparing nodes. Group the fleet by data rate. A reading from a 2.4 kbps node and one from a 76.8 kbps node are not comparable and never will be.
- Log at least 100 packets per node and use a low percentile. Single readings in a multipath environment scatter by 8–10 dB; the 10th percentile is a far better proxy for what the application experiences than the mean, and the mean is better than one sample.
- Record PER over the same window. RSSI is a hypothesis; packet error rate is the verdict. A node at -95 dBm with 0.1% PER is a working node. A node at -78 dBm with 30% PER is a broken one, whatever the meter says.
- Check the physics before touching the radio. Antenna height and feeder loss move the numbers far more than most firmware settings — the arithmetic is in antenna height beats transmit power, and the budget side is in how far 470 MHz actually reaches.
One code block worth keeping in the commissioning folder — it turns a CSV of readings into margins per rate, so the survey output stops being a ranking:
import math
PMIN_REF, RATE_REF, REQ = -116.0, 19.2, 15.17 # dBm, kbps, dB
def pmin(rate):
return PMIN_REF - 10 * math.log10(RATE_REF / rate)
def snr(rssi, floor):
p = 10 ** (rssi / 10) - 10 ** (floor / 10)
return 10 * math.log10(p) - floor if p > 0 else float("-inf")
for node, rate, rssi, floor in [
("A", 76.8, -80.0, -98.0),
("B", 2.4, -90.0, -113.0),
]:
s = snr(rssi, floor)
print(f"{node}: pmin {pmin(rate):6.1f} snr {s:5.1f} "
f"margin above threshold {s - REQ:+5.1f} dB")Run it on the car park data and node A prints -4.1, node B prints +7.8. That is the whole article in two lines of output.
What this does not mean
Two caveats, in the interest of not overcorrecting.
RSSI is still the cheapest early-warning signal you have. A node whose reading drifts down 6 dB over a wet season is telling you something is corroding, and no other metric will surface that as cheaply. Use it as a trend line, not a threshold.
And the noise floor is not a constant you measure once. Man-made noise at VHF/UHF is separately modelled from thermal noise in ITU-R P.372 precisely because it varies by site and by hour — a pump house at 14:00 and the same pump house at 02:00 are different radio environments. Re-measure when the site's duty cycle changes, not just when you install.
FAQ
Why can a -90 dBm reading be better than -80 dBm?
Because the two readings are probably not on the same scale. Sensitivity depends on data rate: at 2.4 kbps a receiver needs roughly 15 dB less signal than at 76.8 kbps. A -90 dBm node at 2.4 kbps sits 35 dB above its threshold, while a -80 dBm node at 76.8 kbps sits 30 dB above its own. Add in-band interference to the higher-rate node and the gap widens.
Can I subtract the noise floor to get a real SNR?
Yes, if the noise floor was measured on an idle channel in the same bandwidth and at the same rate: S = 10log10(10^(RSSI/10) - 10^(N/10)). The catch is that this only removes noise power. Interference from another transmitter is not noise, it is another signal, and the subtraction cannot separate it from yours — which is why a high reading with heavy interference still yields a low usable SNR.
Does a higher transmit power fix a weak reading?
It moves the number and often does not move the packet error rate, because the limit is frequently at the other end: the node's own receiver, its antenna, or interference. Doubling transmit power buys 3 dB; going from 76.8 to 2.4 kbps buys 15 dB of sensitivity without touching the power budget or the regulatory limit. If the link is uplink-limited, transmit power at the gateway is irrelevant.
How many RSSI samples should I log per node?
At least 100 packets, and judge on the 10th percentile rather than the mean. Indoor multipath at 470 MHz produces 8–10 dB of packet-to-packet scatter at a fixed position, so a single reading tells you about one fade state, not about the link. The full product range and its published sensitivity figures are on the products page.



