CUDA-Q Realtime

CUDA-Q Realtime is a library for tightly coupling GPU-accelerated compute to the control system of a quantum processor via a networking layer. It requires a host system with NVIDIA GPU and ConnectX-7/BlueField NIC, and an FPGA connected to the NIC.

It fulfills two primary responsibilities:

  1. It provides the low-level basis of realtime co-processing between FPGA and CPU-GPU systems.

  2. It provides the low latency networking layer of the NVQLink architecture, enabling system integrators to achieve few-microsecond data round trips between FPGA and GPU.

The provided networking layer leverages the Holoscan Sensor Bridge to handle high-bandwidth data transfer over Ethernet using the RoCE protocol. This layer can optionally be replaced with a custom implementation.