Hi NVIDIA Aerial team,
We have successfully brought up an Aerial setup based on DGX Spark and a WNC O-RU, following the NVIDIA Aerial documentation. The system can run end-to-end and UEs can attach successfully.
However, the measured throughput is significantly lower than the performance shown in the Aerial CUDA-Accelerated RAN 26-1 documentation for DGX Spark, where the documented E2E result reaches approximately:
- DL: 1.5 Gbps
- UL: 210 Mbps
- 1 peak cell
- Validated CX7 performance and timing capability
In our current setup, the best throughput we can achieve is approximately:
- DL: ~500 Mbps
- UL: ~3 Mbps
The UL throughput is especially low.
Our setup:
- Platform: DGX Spark
- Aerial version: ACAR 26-1 / Aerial Testbed 1.0
- O-RU: WNC O-RU, 4T4R
- Band: n78
- Bandwidth: 100 MHz
- SCS: 30 kHz
- O-RAN split: 7.2x
- C/U-plane VLAN: 2
- M-plane VLAN: 100
- PCP: 7
- Compression: BFP, 9 bits
- PTP: synchronized
- Fronthaul NIC: CX7
- UE count tested: 2 UEs
- Traffic test: iperf / throughput test
We would like to understand whether this performance gap is expected with the WNC O-RU, or whether we should adjust some configuration parameters to approach the documented DGX Spark throughput.
Specific questions:
-
Is the documented 1.5 Gbps DL / 210 Mbps UL result on DGX Spark tied to a specific O-RU model, UE model, number of layers, TDD pattern, MCS, or test configuration?
-
For WNC 4T4R O-RU, what DL/UL throughput should we reasonably expect with 100 MHz n78 and 30 kHz SCS?
-
Are there recommended reference configurations for DGX Spark + WNC O-RU, especially for:
- cuphycontroller.yaml
- OAI gNB config
- TDD pattern
- PRACH configuration index
- PUSCH / PUCCH / SRS configuration
- eAxC ID mapping
- compression settings
- phase compensation settings
- MTU / jumbo frame settings
- PTP / SyncE / G.8275.1 settings
-
Since DL can reach around 500 Mbps but UL stays around 3 Mbps, which areas should we check first?
- UL scheduling?
- PUSCH MCS?
- UE capability?
- TDD UL slot allocation?
- SRS configuration?
- PRACH / timing advance?
- O-RU RX gain?
- eAxC mapping for UL?
- PTP timing stability?
- fronthaul packet loss or dropped packets?
-
Are there any known limitations or required tuning steps for UL throughput on DGX Spark with WNC O-RU?
-
Could you share the exact reference setup used for the documented DGX Spark result, including:
- O-RU model
- UE model
- number of UEs
- number of MIMO layers
- TDD pattern
- bandwidth / SCS
- gNB configuration
- cuphycontroller.yaml
- iperf or traffic generation method
We are trying to determine whether the current bottleneck is caused by the O-RU, by the UE capability, or by a configuration mismatch in the Aerial/OAI/O-RU setup.
Any guidance on the recommended debug steps would be very helpful.
Thanks.