Everything between the converter and the decoder, running to a deadline set by the numerology rather than by the design. An OFDM symbol arrives every few tens of microseconds whether or not the previous one has been processed, which makes low PHY a throughput problem with a fixed budget rather than an algorithm problem with a target.
Numerically controlled oscillator and mixer bring the band of interest to baseband, followed by the decimating filter chain that gets from converter rate to something the datapath can afford.
Coarse and fine timing and frequency acquisition, then tracking. Getting this wrong does not degrade performance gracefully — the chain simply does not work.
The guard interval that turns a convolutional channel into a per-subcarrier multiplication is discarded here, and any residual timing error appears as phase rotation across the band.
The transform that separates subcarriers, sized by the numerology and required to complete inside the symbol period at every bandwidth the design supports.
Pulling the reference signals, control and data out of the resource grid according to the configuration currently in force.
Estimating the channel at pilot positions, then interpolating across frequency and time. Interpolation quality is where most implementations gain or lose performance.
Undoing the channel per subcarrier, and separating spatial layers where more than one is transmitted. Cost scales sharply with layer count.
Converting equalised symbols into log-likelihood ratios for the decoder, with the quantisation of those values traded directly against coding gain.
Subcarrier spacing determines symbol duration. Everything in the chain has to fit inside it, and the budget shrinks as spacing widens.
| Subcarrier spacing | Symbol duration, approximately | What that means downstream |
|---|---|---|
| 15 kHz | 66.7 µs plus cyclic prefix | The most forgiving budget; legacy-compatible numerology |
| 30 kHz | 33.3 µs plus cyclic prefix | Common for sub-6 GHz 5G; halves every processing budget in the chain |
| 60 kHz | 16.7 µs plus cyclic prefix | Shorter slots for latency-sensitive traffic; pressure on FFT throughput |
| 120 kHz | 8.3 µs plus cyclic prefix | Millimetre wave; the transform and equaliser dominate the design |
A design that closes at 15 kHz and is expected to also serve 120 kHz is not the same design with a parameter changed.
In a low PHY implementation the resource is rarely spent where a block diagram suggests. The transform, the channel estimation memory and the equaliser dominate, and the memory usually exceeds the arithmetic.
Parallelism is dictated by the deadline. Because a symbol must complete within its period, the datapath is sized for the worst case and then either runs at that rate constantly or is time-shared across bandwidth. Both approaches are defensible; choosing between them is an early architectural decision that is expensive to revisit.
Fixed-point analysis runs alongside rather than afterwards. Word lengths grow through the transform, shrink at the equaliser and are deliberately reduced at the soft-value output where memory cost is highest, and each of those decisions is verified against performance rather than assumed safe.
The lower part of the physical layer: down-conversion, synchronisation, cyclic prefix handling, the FFT, resource mapping and beamforming. In an O-RAN 7.2x split it sits in the radio unit, with upper physical layer processing in the distributed unit.
Because subcarrier spacing sets symbol duration, and symbol duration is the processing budget. At 15 kHz a symbol lasts about 67 microseconds; at 120 kHz about 8. A design that closes comfortably at one spacing is not the same design with a parameter changed when it has to serve another.
Memory, usually. The FFT, channel estimation storage and equaliser account for most of it, and the memory typically exceeds the arithmetic. Block diagrams suggest the opposite, which is why early estimates based on operation counts tend to be optimistic.
Reference signals occupy only a fraction of the resource grid, so the channel between them has to be interpolated across frequency and time. In fast-fading or high-Doppler conditions the quality of that interpolation is where implementations differ most, and it costs memory because it needs history.
Minimum mean square error detection requires a matrix inversion per subcarrier, and the matrix dimension grows with the number of spatial layers. What is trivial for a single layer becomes the dominant cost at four, so detector choice is a resource decision rather than purely a performance one.
Frequency-domain data, usually compressed. Exactly what sits either side of that line determines both fronthaul bandwidth and which device performs which processing, which is why the split point is an architecture decision and not a configuration option.
The layers above, and the RU/DU/CU split.
TECHNOLOGYWhat consumes the soft values.
PRODUCTThe decimation chain in front.
PRODUCTWideband channelisation.
PRODUCTCFR and DPD after the transmit chain.
CASE STUDYA datapath closed in the budgeted device.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.