5G Systems — the radio, the network, and the intelligence in it

Projects Tracker · 5G Systems · track 4 of 5 ← Smart Devices Pi 5 home cloud →

writing   One track covering the whole stack: how a 5G radio actually moves bits, what hardware that demands, how the network is split into open interfaces, and where machine learning enters it. This used to be three separate headings — "AI-RAN", "Open RAN" and the 5G notes — which was three names for one subject.

In one line. Four layers, bottom to top: NR (what the standard says) → the front end (what the hardware must do about it) → Open RAN (how the network is split) → AI-RAN (what learns, and how fast it is allowed to). The notes live in 5G Systems Notes/; this page is the plan and the state.

Contents
  1. What is written
  2. The four layers, and why that order
  3. The next step
  4. The sources
  5. Non-terrestrial networks

What is written

The notes live in the engine folder 5G Systems Notes/, and that folder reports itself here — counted live on every page load, so this cannot fall behind what actually exists:

14 notes, 89 sections, 63,831 words in 5G Systems Notes/. 16 notes planned and not yet written. One track over six layers, plus a reference layer. Layer 1 has three deep notes — frame structure, sounding, PDSCH. The reference layer now has eleven notes: six on the radio, and five added 2026-08-28 covering the core network and the device as a protocol entity, from the v19.7.0 core snapshot.

counted live from the notes folder

The full note table is on the folder's own hub, which is also where the reading order lives.


The four layers, and why that order

The four used to be separate interests. They are not: each one only makes sense once the one below it does, and the interesting questions live where two of them meet.

Figure 1: The four layers of this track. Each rests on the one below - you cannot reason about what a front end must do until you know what the standard asks of it, and you cannot judge where learning helps in a RAN until you know what the RAN is made of.

1 · NR — what the standard says. The physical layer: how time and frequency are diced, how the network sounds the channel, how many streams it can send. Written from TS 38.211 / 38.213 / 38.214. This is where the track is now.

2 · The front end — what the hardware must do about it. Power amplifiers, filters, switches, insertion loss, antenna switching, thermal and power budgets. Written from TS 38.101-1 plus the RF books. The seam between 1 and 2 is the most valuable part of the track, because most people know one side or the other.

3 · Open RAN — how the network is split. The O-RAN Alliance's interfaces, and what it means to buy a radio unit from one vendor and a distributed unit from another.

4 · AI-RAN — where learning enters. And this is the layer that cannot be understood in isolation, because the answer is architectural: learning enters at the RAN Intelligent Controller, and the O-RAN specifications pin the timescales — the near-real-time RIC runs xApps at 10 ms – 1 s, the non-real-time RIC runs rApps above 1 s. Anything faster than 10 ms is not a RIC problem at all; it is a scheduler or a front-end control problem, which puts it back in layers 1 and 2.

Why the order is not negotiable. A note on AI-RAN written before the NR notes would be a survey — a list of things people have tried. Written after them, it can say which decision is being learned, what it costs in insertion loss or airtime, and at what timescale it must land. That difference is the whole value of doing this bottom-up.


The next step


The sources

Everything for layers 1 and 2 is already on disk. The archive reports itself here:

The Courses (RAW) folder is not reachable from this server.

All four specifications are v19.4.0 — one coherent Release 19 snapshot, so a cross-reference from one to another resolves.

The list, and what to skip

The list below is the standard works, not a reading pile — the ones the field actually cites and the specifications that define the terms in the posting. Two rules for using it: the 3GPP specs are the primary source and they are free, and a book is there to make a spec readable, never to replace it.

On "most cited". Citation counts differ by database and drift. Where a number appears below it is one I checked against the source named; everything else is listed because it is the reference practitioners in that sub-field default to, which is the more useful test.

The specifications — layers 1 and 2

Free from the 3GPP portal. These are what the job's vocabulary is defined in, and reading them is the difference between knowing the words and knowing the procedures.

SpecWhat it definesWhy it matters hereHave it?
TS 38.211Physical channels and modulationThe resource grid, and the SRS signal itselfin the archive
TS 38.214Physical layer procedures for data§6.2.1.2 is SRS antenna switching — the UE capability supportedSRS-TxPortSwitch, 1T2R / 1T4R / 2T4R. Also MIMO, CSI and rank adaptation.fetch
TS 38.213Physical layer procedures for controlPower control and timing — the constraints a control algorithm must respectfetch
TS 38.101-1UE radio transmission and reception, FR1The front-end spec. Band combinations, MSD, RF requirements — this is where "routing flexibility versus insertion loss" stops being abstract.fetch
TS 38.300NR and NG-RAN overall descriptionThe baseline every other spec assumes; already on the NTN listfetch
TR 38.811 / 38.821NTN — study and solutionsTrack 4's existing reading listfetch
Table 1: The specifications this track is written from. TS 38.214 section 6.2.1.2 is the one to read first - it is where SRS antenna switching is actually specified, and SRS antenna switching is named in the posting.

TS 38.101-1 is the one people forget. It is the single spec that governs what a front-end must physically support.

The books

BookForWhy this one
Dahlman, Parkvall & Sköld — 5G/5G-Advanced: The New Generation Wireless Access Technology, 3rd ed., Academic Press 2023 Layer 1 The practitioner default for NR. All three authors are Ericsson Research and have been in 3GPP standardisation 20+ years — it is the standard explained by the people who wrote it. The 3rd edition adds Rel-17/18: MIMO enhancements, NTN, RedCap.
Ahmadi — 5G NR: Architecture, Technology, Implementation and Operation Layer 1 Already in your archive. Denser and more implementation-facing than Dahlman; good second pass, hard first pass.
Tse & Viswanath — Fundamentals of Wireless Communication, Cambridge 2005 Layer 1 (MIMO) The theory canon for MIMO — spatial multiplexing, diversity–multiplexing tradeoff, opportunistic communication. The authors host a free PDF. Read the MIMO chapters, not the whole book.
Goldsmith — Wireless Communications, Cambridge 2005 Layer 1 The alternative to Tse for link-level material — adaptive modulation, fading, capacity. 2,094 citations (Semantic Scholar). Pick one of the two; do not read both.
Pozar — Microwave Engineering, Wiley Layer 2 The standard undergraduate-to-professional text for S-parameters, matching and insertion loss — the vocabulary of the trade-off the posting names.
Cripps — RF Power Amplifiers for Wireless Communications, 2nd ed., Artech House Layer 2 The PA reference, consistently named as the authoritative one in the RF community. Cripps' Advanced Techniques in RF Power Amplifier Design is the follow-on.
Razavi — RF Microelectronics, 2nd ed., Prentice Hall Layer 2 Transceiver architectures from the IC side — LNAs, mixers, noise figure, linearity. Read it for why a front-end is partitioned the way it is.
Sutton & Barto — Reinforcement Learning: An Introduction, 2nd ed., MIT Press Layer 4 The RL canon, free from the authors. You have already written the deep Q-learning note; this is the reference behind it.
Table 2: The standard texts by layer. Dahlman is the practitioner default for NR because its authors wrote much of the standard; Tse and Viswanath is the theory canon and is free from the authors. Two citation counts were checked and are attributed in the rightmost column.

Papers and industry sources — layers 3 and 4

O'Shea & Hoydis, "An Introduction to Deep Learning for the Physical Layer", IEEE Transactions on Cognitive Communications and Networking 3(4):563–575, 2017 — arXiv:1702.00832. 1,386 citations (SciSpace). The paper that started ML-for-PHY as a field, by treating a communications link as an autoencoder. Read it before phase 4 so you can say where your work sits relative to it.

The O-RAN Alliance specifications — and this is the source for layer 4 It enters at the RAN Intelligent Controller, and the timescales are specified: the Near-RT RIC runs xApps at 10 ms – 1 s, the Non-RT RIC runs rApps above 1 s. That split is the whole architecture of AI in the RAN, and it is a precise, quotable answer in an interview.

Vendor application notes — Rohde & Schwarz and Keysight publish the practical material on SRS antenna switching test, band combinations and front-end measurement that no textbook covers. They are written to sell instruments and are still the best free source on how these procedures behave in practice.

What to skip. Survey papers on "AI for 6G" are abundant, cited, and near-useless — they describe a landscape rather than a mechanism. One primary paper plus the O-RAN specs beats ten surveys.



Non-terrestrial networks

NTN is the part with a fixed reading list, so it is the easiest to plan:

DocumentWhat it coversRead
TR 38.811Study on NR to support non-terrestrial networks — the channel and scenarios.no
TR 38.821Solutions for NR to support NTN — the architecture options.no
3GPP Rel-17The release that first specifies NTN properly.no
3GPP Rel-185G-Advanced: the NTN enhancements on top of Rel-17.no
TS 38.300NR and NG-RAN overall description — the baseline everything else assumes.no
Table 3: The NTN reading list. The two technical reports come first — they are the study items the releases were built from.
Projects Tracker · 5G Systems · track 4 of 5 ← Smart Devices Pi 5 home cloud →