cuPHY System Overview

Aerial CUDA-Accelerated RAN 24-1 (Archive)

Aerial cuPHY is a software-defined workload hosted on NVIDIA-certified EGX servers and a stack that uses the CUDA OS platform and GPU/NIC/CPU firmware and toolkits. This section highlights the Aerial cuPHY workload configuration interdependencies as part of the NVIDIA platform stack.

  • Grace Hopper MGX system supports 20 average loaded (50%) BFP9 cells

  • Supports Massive MIMO: 64T64R (16DL | 8UL) @ 100MHz w/ SRS-based Beamforming

  • AX800 with Dell R750 supports 8 4T4R peak and 16 4T4R average BFP9 cells

  • A100X with Dell R750 supports 5 4T4R peak and 10 4T4R average BFP9 cells

  • A100 with Aerial Devkit supports up to 4 4T4R peak and 8 4T4R average BFP9 cells

Feature Configuration Supported
Grace Hopper MGX Platform
  • 72-core NVIDIA Grace CPU
  • NVIDIA H100 Tensor Core GPU
  • 480GB of LPDDR5X memory with ECC
  • Supports 96GB of HBM3
  • BF3 NIC x2
Y
Platform for Discrete GPU, NIC
  • Servers x2 Model Gigabyte E251-U70
  • CPU Intel Xeon Gold 6240R, 2.4GHz, 24C48T
  • Memory 96GB DDR4
  • Storage 480GB LiteOn SSD x1
  • GPU GA100 x2
  • NIC x2 MLX CX6-DX (MCX623106AE-CDAT)
Y
Platform for Converged Accelerator Dell R750
  • Server Skew 10-AYCG
  • Intel Xeon Gold 6336Y 2.4G, 24C/48T
  • PCIe Gen4
  • Memory 512GB DDR4
  • Storage 2TB
  • GPU+NIC AX800 or A100X
  • NIC CX6-DX
Y
Feature Supported
Maintain integration, qualification of discreet GPU A100 Y
Support Aerial CUDA-Accelerated RAN offload Y
Preserve existing interfaces / API / shared memory buffers integrated with ISVs Y
L2+ integrated with x86 host Y
Previous cuPHY Features Overview
Next Operations, Administration, and Management (OAM) Guide
© Copyright 2024, NVIDIA. Last updated on Apr 26, 2024.