COSMOS lets you build a complete private mobile network — a base station, a core network and one or more phones — out of software, servers and radios that you reserve by the hour. This section teaches you how, one step at a time, on three different software platforms. You do not need to know any of them in advance: start with a tutorial that needs no radio at all, and add one new piece of hardware at each step.
Every platform here uses the same test network settings, so what you learn on one carries over to the others: test network code (PLMN) 00101, UE address pool 192.168.100.0/22 with internet access, and the same commercial 5G modem as the phone.
A platform is the software that implements the base station (and usually comes with a software phone and a default core). Pick one, then follow its learning path. If you are unsure, start with OCUDU or Duranta: both are open source, both have a tutorial that runs on any COSMOS server with no radio, and both go all the way to O-RAN radio units.
| Platform | What it is | Radio needed? | Domains | Start here |
|---|---|---|---|---|
| OCUDU | Open-source 5G base station (CU and DU in one program) built on the srsRAN Project, written in C++ and configured in YAML. Comes with a software UE, the Open5GS core, a built-in E2 agent for RAN controllers, and separate O-RAN M-Plane tools. | Not for the first tutorial. Later tutorials use a USRP N310 or an O-RAN radio unit. | any (no-radio tutorials); sb1 (over the air); osc, sb1, rrail (O-RAN 7.2) |
ocudu-soft |
| Duranta | Open-source 5G base station and software UE, a fork of OpenAirInterface (OAI), written in C. Its base station can manage an O-RAN radio unit itself over the M-Plane, and it can simulate a satellite link. Runs with the OAI CN5G core. | Not for the first tutorial or the satellite tutorial. Later tutorials use a USRP N310 or an O-RAN radio unit. | any (no-radio tutorials); sb1 (over the air); osc, sb1, rrail (O-RAN 7.2) |
duranta-soft |
| Amarisoft | Commercial 4G and 5G stack: base station, a combined 4G/5G core, and a UE simulator (lteue) that runs many UEs on one radio. Closed source, with a floating licence shared by the whole testbed. |
Yes, always. Two USRP N310 radios, one for the base station and one for the simulated UEs. | osc + sb1 |
amarisoft |
How they differ, in one line each:
lteue is also the test UE the other platforms are checked against.GPU physical layer: NVIDIA Aerial (cuPHY/cuBB) — the 5G physical layer on a GH200 GPU, with full stacks on top: OCUDU on Aerial (one gNB, three O-RAN 7.2 cells, 1.2 Gbit/s aggregate UDP downlink) and Duranta on Aerial.
Commercial O-RAN platform on Kubernetes: BubbleRAN — a whole 5G network (OAI RAN, OAI-CN or Open5GS core, FlexRIC near-RT RIC, ODIN non-RT RIC) described as Kubernetes objects and run by BubbleRAN's operators on a dedicated rrail cluster, with the LiteOn O-RU over the air (TCP DL up to 262, UDP DL 562 Mbit/s with a Quectel modem) and an RF-simulator network for xApp and rApp work. The hands-on tutorial is for rrail group members.
The best result measured on each platform so far. The per-platform pages have every modality on each server platform, including the gaps: 2× Xeon Gold 6226, 2× Xeon Gold 6126 and a single-socket AMD EPYC 9355P, plus an NVIDIA GH200 (Arm Grace CPU + Hopper GPU) that runs the physical layer on the GPU with NVIDIA Aerial. On the EPYC, Duranta's soft-only cell runs at about 3× the Xeons (DL 447 / UL 201 Mbit/s).
UDP is the rate the receiver got at the highest offered load (the cell's ceiling; the loss at that load is the price). TCP is the best single iperf3 run over 1, 4 and 8 parallel streams. All rows use one UE.
| Platform | Configuration | UE | RAN server CPU platform | DL UDP, Mbit/s | DL TCP, Mbit/s | UL UDP, Mbit/s | UL TCP, Mbit/s | Date · image |
|---|---|---|---|---|---|---|---|---|
| OCUDU (srsRAN) | O-RAN 7.2 split, LiteOn FlexFi O-RU, n78, 100 MHz, 4T4R | Quectel RM520N-GL on USB 3 | 1× AMD EPYC 9355P, 32 cores / 32 threads | 711 (0.8 % loss; 714 at 4.8 %) | 150 | 166 | 147 | 2026-10-23 · ocudu-20261020 (26.10) |
| OCUDU (srsRAN) + ACC100 | as above, Intel ACC100 LDPC offload | SIMCom SIM8200 on USB 3 | 1× AMD EPYC 9355P, 32 cores / 32 threads | 572 (lossless up to 314) | 214 | 72.9 | 75.9 | 2026-09-29 · ocudu-20260928 |
| OCUDU (srsRAN) + NVIDIA Aerial | O-RAN 7.2 split, Benetel RAN650 O-RU, n78, 100 MHz, 4T4R, layer 1 on the GPU | Quectel modem on USB 3 | NVIDIA GH200: Grace, 72 Arm Neoverse V2 cores / 72 threads, + Hopper GPU (Aerial L1) | 590 (700 with TDD 7 DL / 2 UL) | 474 | not measured | 144 | 2026-11-09 · aerial-26-2-gh200 (Aerial 26-2) |
| Duranta (OAI) | O-RAN 7.2 split, Benetel RAN650 O-RU, n78, 100 MHz, 4 DL layers | Quectel RM520N-GL on USB 3 | 1× AMD EPYC 9355P, 32 cores / 32 threads | 882 (0.4 % loss) | 233 | 63.6 | 63.5 | 2026-10-25 · duranta-w43-20261024 (w43; w42: 846 / 222 / 64.1 / 65.0) |
| Duranta (OAI) + ACC100 | as above, 1 DL layer, Intel ACC100 LDPC offload | SIMCom SIM8200 on USB 3 | 2× Intel Xeon Gold 6126 @ 2.60 GHz, 24 cores / 48 threads | 367 | 221 | 59.2 | 52.2 | 2026-09-27 · duranta-20260927 |
| Duranta (OAI) + NVIDIA Aerial | O-RAN 7.2 split, Foxconn RPQN-7801 O-RU, n78, 100 MHz, layer 1 on the GPU | 5G modem on USB 3 | NVIDIA GH200: Grace, 72 Arm Neoverse V2 cores / 72 threads, + Hopper GPU (Aerial L1) | 690 | 528 | not measured | 105 | 2026-11-04 · aerial-26-2-gh200 (Aerial 26-2) |
| Amarisoft | O-RAN 7.2 split, LiteOn FlexFi O-RU (trx_s72), n78, 60 MHz, 2 DL layers |
Quectel RM520N-GL on USB 3 | 2× Intel Xeon Gold 6126 @ 2.60 GHz, 24 cores / 48 threads | not measured | 344–349 | not measured | 12 | 2026-10-11 · amarisoft-20261009 (Amarisoft 2026-09-11) |
One gNB, several cells: OCUDU on the GH200 serves three O-RUs from one gNB, all loaded at once: 1190 Mbit/s UDP downlink in total (1365 with TDD 7 DL / 2 UL), TCP 650–727 DL / 248–265 UL (details). The table above keeps one UE per row.
Single-UE TCP stops near 150–233 Mbit/s on the 100 MHz x86 rows, whatever the cell's UDP ceiling, so compare those cells by UDP. The two ACC100 rows used a SIMCom SIM8200, the others a Quectel RM520N-GL. Single-stream TCP peaks of 248 (OCUDU, 2026-09-09) and 268 Mbit/s (Duranta, 2026-09-07) were measured with a Quectel RM520N-GL and no UDP run.
CPU decides Duranta's layers. On the 2× Xeon Gold 6126 servers Duranta runs one DL layer (UDP 348, TCP 208, UL 61 / 53 on duranta-20260927); multi-layer needs more real-time CPU than that server leaves, and duranta.ndz picks the layer count from the host automatically. OCUDU runs 2 layers on the same Xeon.
ACC100 moves LDPC encoding and decoding onto an Intel ACC100 card in the same server. At one UE the throughput is set by the radio, so the gain shows as CPU (Duranta's DU uses about 0.5 core less during downlink). Both stacks run a 40-minute full-load ACC100 soak without a single error on the Supermicro AMD EPYC server. ⚠️ Not on the HPE DL380 Gen10 servers (2× Xeon Gold 6126): the card's slot there raises uncorrectable PCIe errors under load, so ocudu-gnb --acc100 refuses to start on HPE hardware.
The Amarisoft row uses a 60 MHz carrier and the others 100 MHz: its licence caps cell bandwidth × antenna ports at 120 MHz, so it runs 30 MHz with 4 ports, 60 MHz with 2 or 100 MHz with 1. On the same LiteOn rig and CPU platform (2026-10-10/11, TCP): 30 MHz 4T4R DL 259–277 / UL 10 with a SIMCom on USB 3 (UDP DL 323); 100 MHz 1T1R DL 216 / UL 17, or DL 150 / UL 34 with the uplink-heavier DDSUU pattern. Over the air with USRPs (n41, 40 MHz, one lteue UE) Amarisoft measured 76 / 21 Mbit/s on 2026-09-13; see the 7.2 tutorial and the results page. On the AMD EPYC (2026-10-23, the same 60 MHz 2-port LiteOn cell, Quectel RM520N-GL) Amarisoft measured UDP DL 360 (lossless), TCP DL up to 307 and UL 16–22 Mbit/s.
OCUDU on the two O-RUs (AMD EPYC, 2026-10-23, Quectel RM520N-GL): the LiteOn row above has the higher UDP and the highest uplink of any stack on any O-RU (UDP 166, TCP 147 Mbit/s). The Benetel gives the better single-UE TCP downlink: DL TCP 206–212, UDP 483–591, UL 74–83 Mbit/s. Both builds measured the same: OCUDU main gave DL TCP 203–212 on the Benetel and UL UDP 127–161 on the LiteOn. Every round is on the OCUDU trend page.
How each build performed over time is on the trend pages: OCUDU · Duranta · Amarisoft.
Each step adds exactly one new thing that can fail on its own, so when something breaks you know where to look. Do the steps in order the first time.
| Step | What is new | OCUDU | Duranta | Amarisoft |
|---|---|---|---|---|
| 0. Foundations | the concepts | 5G SA fundamentals · O-RAN and the 7.2x fronthaul | same | same |
| 1. No radio | the whole network in software on one server | ocudu-soft | duranta-soft | — (Amarisoft has no radio-less mode) |
| 2. Over the air | a real radio (USRP) and a commercial 5G modem | ocudu-ota | duranta-ota | amarisoft (1 to 64 simulated UEs) |
| 3. O-RAN 7.2 | a separate radio unit on a fronthaul network | ocudu-72 | duranta-72 | — |
| 4. Beyond | managing a radio unit, or a satellite link | ocudu-mplane (M-Plane, simulated radio unit) | duranta-ntn (satellite, no radio) | — |
Three ways of connecting the base station to the UE appear in these steps. They are called modalities:
Each tutorial is installed on every COSMOS console as a package, so there is nothing to download. Ask whether it can run, then run it:
cosmos-tutorial list # installed tutorials and versions
cosmos-tutorial check <name> # RUNNABLE / NOT RUNNABLE, with the reasons; changes nothing
cosmos-tutorial run <name> # Ansible mode
cosmos-tutorial run <name> --mode manual # the same commands the page walks through
Each tutorial page also has a Jupyter notebook for the same experiment, and a parity run that executes all three modes and checks they agree. Every run writes its logs and results to ~/cosmos-runs/<name>/ on the console.
cosmos-tutorial run <name> --mode teardown). A base station left running blocks the next user's radio.Author(s): COSMOS team · Tags: 5g, 4g, lte, sa, oran, cellular, course, index
Part of the COSMOS Tutorials collection.