Part of: NVIDIA Aerial on COSMOS · 4G / 5G & O-RAN
A live 5G standalone cell with the physical layer on a GPU: the Duranta layer 2 (MAC, scheduler, RLC, PDCP, RRC) runs above NVIDIA Aerial 26-2 on the GH200 server, and a commercial O-RAN 7.2 radio unit carries the cell over the air to a commercial 5G modem. This page records what works today and the measured performance. The step-by-step tutorial (aerial-ota) will follow. The same L1 with the OCUDU gNB is on OCUDU on NVIDIA Aerial.
Status (verified 2026-11-04). All three COSMOS O-RUs work one at a time: Foxconn RPQN-7801, Benetel RAN650 and LiteOn FlexFi. In each case the modem attaches, gets an IP address and carries traffic. Three cells on air at once also work: one Aerial L1 on one GPU serves three cells, one per O-RU, each with its own Duranta layer 2, and a modem attaches and carries traffic on each cell in turn. Traffic on two or more cells at the same time does not work yet with Duranta; see Three O-RUs at once.
| Part | What runs | Where |
|---|---|---|
| Core | OAI 5G core (AMF, SMF, UPF), containers | the GH200 server |
| Layer 2 (DU high + CU) | Duranta, OAI 2026.w44 tree, built for arm64 with the Aerial FAPI interface | container on the GH200 server, Arm cores |
| Layer 1 (high PHY) | NVIDIA Aerial 26-2 cuphycontroller (cuPHY on the Hopper GPU) |
cuBB container on the GH200 server |
| L1-L2 interface | SCF FAPI 10.02 over Aerial nvIPC shared memory | same host |
| Fronthaul | O-RAN 7.2x split, eCPRI over VLAN, BFP-9 compression, PTP (G.8275.1) timing | BlueField-3 NIC (NIC mode) → O-RU switch |
| Radio | one O-RU from the O-RU pool, n78, 100 MHz, 30 kHz SCS, 4T4R | O-RU room |
| UE | Quectel RM520N-GL on USB 3, SA, test PLMN 001/01 | UE node in the same RF space |
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.
| O-RU | Centre frequency | PCI | TDD (5 ms period) | Fronthaul notes |
|---|---|---|---|---|
| Foxconn RPQN-7801 | 3750 MHz | 3 | 6 DL / S / 3 UL | the combination OAI and NVIDIA test against |
| Benetel RAN650 | 3350.01 MHz | 2 | 7 DL / S (10 DL symbols) / 2 UL alone; 6 DL / S / 3 UL in the three-cell setup | static compression; TDD must match the RU's two-part pattern (see below) |
| LiteOn FlexFi | 3549.84 MHz | 1 | 6 DL / S / 3 UL | UL arrives with VLAN priority 7; centre must be on the OAI 360 kHz raster |
All cells: band n78, 100 MHz (273 PRB), 30 kHz subcarrier spacing, 4 DL layers.
TCP throughput with iperf3 (receiver line), Quectel RM520N-GL on a USB 3 port, test PLMN 001/01. All three O-RUs measured on the same GH200 server, image aerial-26-2-gh200, Duranta 2026.w44 layer 2, Aerial 26-2 layer 1.
| CPU / GPU platform | O-RU | DL (Mbit/s) | UL (Mbit/s) | Ping | Measured |
|---|---|---|---|---|---|
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | Foxconn RPQN-7801 | 528 | 105 | 18.5 ms | 2026-11-04 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | Benetel RAN650 | 340–352 | 107 | 17.6 ms | 2026-11-04 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | LiteOn FlexFi | 472–487 | 56.8 | 18.5 ms | 2026-11-04 |
Notes:
The modem and its host are not the limit: UDP DL reaches 690 Mbit/s with the host 89 % idle. The limits are:
One Aerial L1 (cuphycontroller) runs three cells on the GH200's GPU, one per O-RU, all on the same BlueField-3 fronthaul port. Each cell has its own Duranta layer 2 container connected to the L1 through its own nvIPC shared-memory instance. All three gNBs register with the same 5G core.
The modem moves from cell to cell with a cell lock while all three cells stay on air.
| CPU / GPU platform | O-RU (cell) | DL (Mbit/s) | UL (Mbit/s) | Ping | Measured |
|---|---|---|---|---|---|
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | Foxconn RPQN-7801 | 502 | 84.8 | 18.5 ms | 2026-11-04 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | Benetel RAN650 | 450 | 191 | 18.5 ms | 2026-11-04 |
| NVIDIA GH200 (Grace, 72 Arm Neoverse V2 cores / 72 threads + Hopper GPU) | LiteOn FlexFi | 430 | 60.2 | 18.5 ms | 2026-11-04 |
Same modem, USB 3 port and core as the single-O-RU table above. One run per cell; these numbers are not yet tuned.
Traffic on two or more cells at the same time does not work with Duranta yet. When modems load two or more cells at once, the L1 stops within about 15 seconds (an uplink-processing timeout under multi-cell load). Use one modem at a time: with all three cells on air, that is stable. The OCUDU gNB runs all three cells loaded at once on the same L1.
Three cells on one L1 need:
DDDD + DDSUUU gives the 6 DL / S / 3 UL pattern the Foxconn and LiteOn use natively. The Benetel's 7 / 2 pattern cannot be used for the others: those radios keep their own pattern and Msg3 lands in a downlink slot.The non-obvious settings each radio needs with Aerial. They apply to anyone pairing Aerial with these O-RUs.
allmulticast is enabled on the port.DDDDDDSUUU is invalid there, so the radio falls back to DDDDDDDSUU and Msg3 lands in a downlink slot. The gNB therefore uses 7 DL slots, an S slot with 10 DL symbols, and 2 UL slots.fs_offset_dl) is per radio: the Foxconn value leaves the Benetel 25 dB low, and 0 overdrives it.Ta4_min) is 0.The step-by-step tutorial is in preparation. Until then:
rrail domain and the O-RU you want (each O-RU is booked separately).aerial-26-2-gh200.ndz on the server with omf load … -r 0 -o 3600. See OMF commands. After boot, wait until systemctl is-active aerial-images says active: the Aerial L1 and Duranta containers come from the COSMOS network share and take about 2 minutes to load (see NVIDIA Aerial on COSMOS).For all three O-RUs at once, book all three and power them on before step 4. The patched layer 2 container comes from the share with the L1, the three-cell L1 configuration is in the image, and the O-RU configurations ship in the cosmos-oru-tools package (version 1.16 or later).
Ask the COSMOS team for the configuration files for steps 3–4.
Last verified: 2026-11-04, image aerial-26-2-gh200, NVIDIA Aerial 26-2, Duranta (OAI 2026.w44).