Part of: NVIDIA Aerial on COSMOS · OCUDU on COSMOS · 4G / 5G & O-RAN
A live 5G standalone network on an NVIDIA GH200 server:
One OCUDU process drives three cells on three different O-RUs through one Aerial L1, and all three carry traffic at the same time. The same layer 1 with the Duranta layer 2 is on Duranta on NVIDIA Aerial.
Status (2026-11-09). Works with all three COSMOS O-RUs, one at a time and all three at once from a single gNB:
- Foxconn RPQN-7801;
- Benetel RAN650;
- LiteOn FlexFi.
With all three cells loaded, the downlink reaches 1.2 Gbit/s UDP in total and about 700 Mbit/s TCP, and the GPU stays below half load. This page records the setup and the measured results.
The point of a GPU physical layer is that one server serves many cells. The PHY work for every cell and slot runs as CUDA kernels on one GPU, while the CPU runs only the layers above it. One cell proves that the software stack fits together. Several cells, each on its own radio and each carrying real traffic at the same moment, demonstrate the architecture.
On COSMOS this is what OCUDU now does:
gnb_split_6 process is configured with three cells, one per O-RU. Each cell has its own frequency, PCI and fronthaul.With a modem on every cell and traffic on all three at once:
| Part | What runs | Where |
|---|---|---|
| Core | OAI 5G core (AMF, SMF, UPF), containers | the GH200 server |
| gNB layer 2 + CU | OCUDU (gnb_split_6) with the OCUDU–Aerial L1 plugin, built for arm64 |
container on the GH200 server, 12 pinned Grace cores |
| Layer 1 (high PHY) | NVIDIA Aerial 26-2 cuphycontroller_scf (stock build, cuPHY on the Hopper GPU) |
cuBB container on the GH200 server |
| L1–L2 interface | SCF FAPI 10.04 over Aerial nvIPC shared memory (the plugin is the nvIPC client) | same host |
| Fronthaul | O-RAN 7.2x split, eCPRI over VLAN, BFP-9 compression, PTP (G.8275.1) timing | BlueField-3 NIC (NIC mode), one 100G port → O-RU switch |
| Radios | three O-RUs from the O-RU pool, n78, 100 MHz, 30 kHz SCS, 4T4R | O-RU room |
| UEs | Quectel RM520N-GL / RM510Q-GL and SIMCom SDX55 5G modems on USB 3, SA, test PLMN 001/01 | UE nodes in the same RF space |
The layer 2 has to run on the same machine as the L1 because nvIPC is shared memory. On the GH200 that is the intended design: the GPU and the 72-core Grace CPU share one cache-coherent module, so one server runs the whole gNB.
Aerial has no O-RAN M-plane, so every O-RU runs in standalone (non-M-plane) mode with a fixed configuration that matches the L1 settings. OCUDU itself carries no fronthaul settings in this split. The O-RU MAC addresses, VLAN, eAxC map and compression all live in the Aerial L1 configuration.
| Component | Version |
|---|---|
| OCUDU | branch feat/aerial-plugin-rel-26-2, commit e99b0d35f4 |
| OCUDU–Aerial plugin | aerial_l1_ocudu_plugin (OCUDU hardware-acceleration working group), commit bf16f40a61, plus one COSMOS patch (below) |
| NVIDIA Aerial | 26-2 (cuBB), stock build.aarch64 L1 |
| Core | OAI CN5G (AMF, SMF, UPF) |
| O-RU | Centre frequency | SSB NR-ARFCN | PCI |
|---|---|---|---|
| LiteOn FlexFi | 3350.01 MHz | 620640 | 1 |
| Benetel RAN650 | 3549.84 MHz | 633984 | 2 |
| Foxconn RPQN-7801 | 3750.00 MHz | 647328 | 3 |
All cells use:
Frequency plan. The three carriers are as far apart as n78 allows, and the LiteOn sits at the bottom edge, 400 MHz from the Foxconn. Its transmitter leaks much more into neighbouring channels than the other two radios: adjacent-channel leakage is −14 / −18 dBc, against −30 / −33 dBc for the Foxconn on the same capture. With the LiteOn in the middle of the band, the other two cells' modems see downlink block error rates of 15–47 % whenever the LiteOn is loaded. With this plan they stay at 0.9 % however hard the LiteOn is loaded.
All runs are on the same server and image, with a cold start each time (O-RUs, servers and modems off beforehand). All modems are on USB 3 ports. Throughput is iperf3 TCP with 8 streams (receiver line) and iperf2 UDP at a fixed offered rate.
| CPU / GPU platform | O-RU | TCP DL (Mbit/s) | TCP UL (Mbit/s) | UDP DL (Mbit/s) | Measured |
|---|---|---|---|---|---|
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | Benetel RAN650 | 474 | 144 | 590 (<1 % loss); 700 at 0.3 % with TDD 7 / 2 | 2026-11-09 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | Foxconn RPQN-7801 | 464 | 147 | 600 (0.8 % loss); 700 at 1.3 % with TDD 7 / 2 | 2026-11-09 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | LiteOn FlexFi | 190 | 26 | — | 2026-11-09 |
On both Foxconn and Benetel the modem reports CQI 14–15 at rank 4, with DL BLER of 1–5 % and ping of 16–17 ms.
| CPU / GPU platform | Traffic | DL (Mbit/s) | UL (Mbit/s) | Measured |
|---|---|---|---|---|
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | UDP fixed rate, every cell < 2 % loss | 1190 (500 + 450 + 250) | — | 2026-11-09 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | UDP, TDD 7 DL / S / 2 UL | 1365 | — | 2026-11-09 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | TCP, one modem per cell | 650–717 | 248–265 | 2026-11-09 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | TCP, four modems (two on Foxconn) | 708–727 | — | 2026-11-09 |
| Resource | Idle cells | All three cells loaded | Budget |
|---|---|---|---|
| GPU SM utilisation | 24 % | 42–49 % | 100 % |
| PDSCH per slot (GPU), p99 | 44 µs | ≤ 78 µs (max 137 µs) | 500 µs slot |
| PUSCH per slot (GPU), p99 | — | ≤ 242 µs (max 269 µs) | 500 µs slot |
| OCUDU gNB on Grace | 1.8 cores | ≤ 2.7 cores | 12 pinned cores |
| Late or dropped fronthaul packets | 0 | 0 | — |
The per-slot times did not change between one cell and three, and the GPU clock stayed fixed at 1980 MHz (about 145 W). At 1.2 Gbit/s the busiest thread of the user-plane function (UPF) of the 5G core was at 68 %, so the core, not the GPU, is the first part to saturate. There is room for more cells on this GPU.
Build. OCUDU with -DENABLE_PLUGINS=ON and the plugin submodule, target gnb_split_6, in an Ubuntu 24.04 container on the GH200 against the Aerial 26-2 SDK headers. The plugin loads libnvipc.so from the SDK at run time. The binary is build/apps/gnb_split_6/gnb.
Plugin patch for the stock L1. The stock Aerial 26-2 build.aarch64 L1 differs from what the plugin sends in two places:
DL_TTI.request;trp_scheme byte.The one-file COSMOS patch makes the plugin's wire layout match the stock L1, so no L1 rebuild is needed. Without it every PDCCH slot fails in the L1 with DL_TTI.req message size mismatch.
Runtime container: the cuBB 26-2 image plus two libraries the gNB needs. It shares the L1's IPC namespace.
Cell settings:
Three cells: one cells: list in the OCUDU configuration, one Aerial L1 configuration with three cells, and the nvIPC channel set to the single default prefix.
cu_cp.inactivity_timer: 7200. At the 120 s default, a modem that goes idle fails to come back (UE capability and context-setup timeouts), and traffic stops on all cells.
Frequency plan (above), so that the LiteOn's adjacent-channel leakage does not reach the other two cells.
O-RU preparation is the same as for the Duranta setup:
aerial-26-2-gh200 image.Last verified: 2026-11-09 (cold-start runs on the GH200 server, image aerial-26-2-gh200).