Release Notes

This section describes the new features and limitations for this release of Aerial cuPHY.

What’s New in 26-2

• L1

  • Support MGX ARC Pro (Alpha).

  • Support multi-port and multi-NIC by single L1 instance.

  • Distributing new LDPC algorithm implementation for FP8 and performance-optimized implementations for large lifting sizes.

• cuMAC

  • New deep reinforcement learning–based framework for training MU-MIMO link adaptation models, consisting of supervised teacher cloning followed by proximal policy optimization (PPO) fine-tuning.

  • Integrated the MU-MIMO UE-grouping module into cuMAC-CP and TestMac.

  • Enabled SRS channel-estimation GPU-buffer sharing between cuphycontroller and cuMAC-CP.

  • Automated test-vector generation for cuMAC and cuBB integration test. Support cuMAC-CP + cuBB F08 4C 66c with default cuMAC configuration: 4 cells, 256 SRS UEs per cell, 16 scheduled UEs per cell/TTI.

• E2E (DGX Spark)

  • 4 Peak cells (in Daisy Chain configuration)

  • eCPRI connection to test equipment

  • 4 DL layers and 2 UL layers per UE per Cell

  • 1.46 Gbps DL and 210 Mbps UL TPUT per cell

• E2E (MGX ARC Pro)

  • MGX ARC Pro - 11 Peak 4T4R cells

    • eCPRI connection to test equipment

    • 400G FH optical connection

    • 4 DL layers and 2 UL layers per UE per Cell

    • 1.46 Gbps DL and 210 Mbps UL TPUT per cell

  • MGX ARC Pro – 1 Peak 4T4R cell

    • RF Cabled to O-RUs

    • 400G optical connection (FH & BH on single port)

    • Support simultaneous PTP & SyncE with DOCA FireFly

    • Up to 8 connected UEs

    • 4 DL layers and 2 UL layers per UE per Cell

    • 1.46 Gbps DL and 210 Mbps UL TPUT per cell

  • MGX ARC Pro – 1 Peak 64T64R cell

    • eCPRI connection to test equipment

    • 400G optical connection (FH & BH on single port)

    • 16 UEs with 1Layer each (16 Total Layers)

    • 8 UEs with 2Layers each (16 Total Layers)

    • Note: This was also performed on GH200.

• dApp

  • Multi-cell, multi-UE E3 indications: The E3 Agent now emits one indication per slot for all active cells, each with its own shared-memory references and per-UE metrics.

  • Additional PUSCH data streams: The E3 Agent exposes and Data Lake stores 15 additional PUSCH streams for link quality, timing, HARQ and H-estimate locators, plus a TAI slot timestamp.

  • SRS telemetry streams: A new SRS path adds 34 streams to the E3 Agent, taking the subscribable set from 28 to 78, and to the new srs, srs_iq and srs_hest Data Lake tables.

  • Shared memory layout v1.1.0: H-estimate and SRS blobs move to per-cell rows located by byte offsets in the indication, with all UE groups now covered. New pyAerial notebooks reconstruct them offline.

  • E3 Agent Standalone: e3agent-standalone runs the production E3 Agent without cuPHY, CUDA, or ClickHouse, from synthetic or replayed traffic, allowing for offline dApp development.

• pyAerial

  • pyAerial works in its own virtual environment within the Aerial container. A separate pyAerial container is no longer needed.

• RANPerf

  • RANPerf testbench is a comprehensive framework that provides the methodology, documentation, and tools to Aerial developers for benchmarking their GPU-based RAN implementations.