Radio operators have long chased the same prize: pulling a readable signal out of a band that sounds like a hiss. For decades, that meant CW with a sharp ear, then RTTY, then PSK31, each one trimming a few decibels off the noise floor. Then in June 2017, Joe Taylor (K1JT) and Steve Franke (K9AN) released a protocol that pushed the threshold to a place where the signal vanishes below the noise on a normal receiver, yet still gets decoded with near certainty. That protocol is FT8, and the question worth asking is whether its weak-signal reputation holds up against the older digital modes when you put the numbers side by side.
How FT8 packs information into 50 hertz of spectrum
The architecture of FT8 is unusually compact. A transmission lasts 12.64 seconds inside a 15-second cycle, carries a 77-bit payload extended with a 14-bit CRC, and runs that 91-bit string through an LDPC(174,91) low-density parity-check encoder. The result is 174 coded bits, mapped onto 58 three-bit symbols using 8-tone frequency shift keying with Gray coding. Three Costas synchronization arrays of seven symbols each sit at the start, middle, and end of the frame, bringing the total to 79 symbols. Tone spacing is exactly 6.25 Hz, so the occupied bandwidth works out to 8 x 6.25 = 50 Hz.
That number matters. Most amateur transceivers treat a 2500 Hz audio passband as standard for SSB work, and FT8 squeezes into one-fiftieth of that space. The Costas arrays, originally developed for radar applications, give the decoder a known reference pattern that resists Doppler shift and timing drift, which is why FT8 survives multi-hop ionospheric paths where amplitude and phase are jumping around.
The decoding threshold and what -20 dB actually means
FT8 is documented as decoding signals down to -20 dB SNR measured in a 2500 Hz reference bandwidth. The QST publication that introduced the protocol gives -20 dB as the threshold for 50 percent decode probability on a non-fading AWGN channel, with several decibels of additional margin when a priori (AP) decoding is enabled. Some implementations push to -21 dB or even -24 dB with AP.
Here is where the apparent magic dissolves into arithmetic. The reference bandwidth is wider than the signal itself, so when an FT8 tone is reported at -20 dB SNR in 2500 Hz, the actual signal-to-noise ratio inside the matched filter is much higher. The relation is:
SNR_matched = SNR_2500 + 10 x log10(2500 / B_signal)
For FT8 with a 6.25 Hz symbol detection bandwidth, that correction gives roughly:
10 x log10(2500 / 6.25) = 10 x log10(400) = 26 dB
So a signal at -20 dB SNR in the 2500 Hz noise window corresponds to about +6 dB SNR inside the narrow detection filter. The signal is below noise to the ear and the waterfall, but well above noise to the math. This is the same principle that lets WSPR decode at -28 dB and JT65 at -25 dB: narrow filtering plus error correction.
Classic digital modes and their reception thresholds
Comparing modes on equal footing requires fixing the reference bandwidth. Using the standard 2500 Hz benchmark, the published thresholds line up like this:
- SSB voice requires roughly +10 dB SNR for readable copy;
- RTTY at 45 baud needs about -5 dB SNR for usable text;
- PSK31 holds copy down to about -10 dB SNR;
- CW by ear sits near -18 dB SNR, with QRSS modes reaching -26 dB;
- FT8 decodes at -20 dB SNR, with AP pushing several decibels lower;
- JT65 reaches -25 dB and is still used for moonbounce work;
- WSPR achieves -28 dB SNR, which is its purpose as a beacon system.
PSK31 occupies about 31 Hz of bandwidth and was a notable advance in the early 2000s, but its threshold is roughly 10 dB worse than FT8. RTTY, the old workhorse of digital contesting, needs 15 dB more signal than FT8 for reliable decode. CW remains competitive because a trained operator's auditory system performs surprisingly well as a matched filter, but only the slow QRSS variants approach FT8 territory.
Where the formulas come from
The theoretical floor for any communication system is governed by Shannon's capacity equation:
C = B x log2(1 + S/N)
where C is channel capacity in bits per second, B is bandwidth in hertz, and S/N is the linear signal-to-noise ratio. For FT8, the effective payload is 77 information bits transmitted in 12.64 seconds, or about 6.1 bits per second. Plugging that into Shannon's formula with the 50 Hz occupied bandwidth shows that the protocol operates comfortably above the theoretical limit, leaving headroom for the LDPC code to do its work.
The energy-per-bit to noise-power-spectral-density ratio (Eb/N0) gives another lens. The Shannon limit for an arbitrarily low bit error rate sits at Eb/N0 = -1.59 dB. Published measurements put FT8 at around +1 dB Eb/N0 for its 50 percent decode threshold, putting it within roughly 3 dB of the absolute theoretical bound. PSK31, by comparison, runs at +9 dB Eb/N0, and RTTY at +14 dB. The gap explains everything.
Real-world performance versus laboratory numbers
Bench measurements assume additive white Gaussian noise, but the ionosphere does not cooperate. Fading, Doppler spread, and selective absorption all degrade real performance. The WSJT-X documentation includes simulations on mid-latitude moderately disturbed channels, where FT8's threshold rises by perhaps 1 to 2 dB, depending on the channel model. Independent testing from operators who compared FT8 and JT65 in evening conditions with real-world noise confirms the QST values come close to lab predictions, sometimes with a small safety margin built in.
What this means in practice: on a 40-meter band saturated with thunderstorm crashes from distant fronts, FT8 will still complete contacts that PSK31 cannot even start. A station running 5 watts to a wire antenna at midnight has a realistic chance of decoding a counterpart 8,000 miles away during a propagation peak, something that would have required at least 100 watts and a directional antenna with older modes.
Trade-offs that often go unmentioned
FT8 is not a universal upgrade. The 15-second cycle structure means a complete contact takes about 90 seconds, with the operator essentially confirming an automated exchange of call signs, grid squares, and signal reports. Free-form conversation is awkward at best. The fixed message format also rules out the ragchew style that PSK31 and RTTY made enjoyable.
There is also a bandwidth-per-throughput question. FT8 carries 6.1 information bits per second across 50 Hz. PSK31 carries 50 bits per second across 31 Hz. On a busy band with strong signals, PSK31 actually delivers more throughput per hertz of spectrum. The point is that FT8 wins decisively when the signal is weak, not when it is strong. The newer FT4 variant, introduced in 2019, sacrifices 3 to 4 dB of sensitivity for a faster cycle aimed at contesting, and the experimental FT2 mode from early 2026 trades even more sensitivity for an 11-second QSO time.
Aggregated spotting data from automated reporting networks indicates roughly 60 to 70 percent of all HF digital contacts worldwide now run on FT8. The network effect has compounded the technical advantage: more stations on FT8 means more chances to find a contact, which draws still more stations to the mode.
What the next generation of weak-signal modes might offer
The mathematics of channel coding has not stopped advancing. LDPC codes used in FT8 are already close to Shannon-optimal for their block length, but longer codes with more iterations could squeeze out another decibel or two. The Q65 protocol, also in the WSJT-X family, uses 65-tone modulation and Q-ary Repeat-Accumulate (QRA) forward error correction for paths with extreme Doppler spread, and reaches sensitivity comparable to JT65 with better resistance to scattering. FST4 and FST4W extend the WSPR concept with even longer integration times.
What FT8 demonstrated, and what the radio engineering community absorbed, is that sensitivity gains are no longer about clever antennas or expensive receivers. They come from how the bits are arranged, how the redundancy is structured, and how the decoder uses every piece of available information. A laptop and a $300 transceiver now do what a half-acre antenna farm could not do twenty years ago. The signal still has to get there, but once it arrives, the math finishes the job.
Our Telegram Channel
Our YouTube Channel