NVIDIA’s open-source Holoscan Sensor Bridge has achieved 1.74 µs round trip latency with Backline, a reduction from 2.47 µs and, according to the authors, the lowest published for quantum controllers between an FPGA and a CPU over a switched network protocol. Researchers directly compared AMD ERNIC and NVIDIA’s Holoscan Sensor Bridge using the same AMD Ryzen Threadripper PRO workstation, network card and cable. Switching to the Holoscan Sensor Bridge reduced latency to an AMD Instinct MI210 GPU from 4.56 µs, shortening the time data spends in transit and maximizing decoding capacity.
Backline Achieves 1.74 µs Latency with NVIDIA’s Holoscan Sensor Bridge
Backline establishes communication between a control FPGA and CPUs or GPUs responsible for decoding quantum information, employing RDMA over Converged Ethernet, or RoCE v2. The system’s RDMA engine, the FPGA logic converting data into network packets, previously relied on AMD’s ERNIC; now, HSB provides an alternative. This addition allows users to select between the two engines without altering their QNodes, compiler passes, or decoders, offering flexibility in system configuration, NVIDIA says.
The Backline hardware handshake and Holoscan Sensor Bridge were tested on an AMD Versal VPK120 FPGA, connected via a single 100 Gbps cable to a ConnectX-7 network card. The team acknowledges that real-world installations often include network switches, each adding latency; THQLink reports 0.13 µs per switch hop.
Comparisons to other systems reveal Backline’s performance relative to existing solutions. THQLink, connecting an FPGA to an Xeon server, achieves a median round trip latency of 2.87 µs, 1.1 µs higher than Backline’s current HSB-enabled result. However, THQLink has since reduced its HSB-based latency to 2.7 µs using Blackwell GPUs, approximately 1 µs faster than Backline’s current measurement, according to NVIDIA.
The difference in performance, the researchers explain, arises from how each system handles replies from the GPU kernel; Backline’s setup allows the kernel to directly post replies to the network card, while the current Backline kernel hands off replies to a CPU. With more of every round now free for decoding, they anticipate further performance gains as users implement their own decoders with Backline.
AMD ERNIC and NVIDIA HSB RDMA Engine Comparison
It’s lightweight, open source under the Apache 2.0 license and a natural fit for a stream. From PennyLane, switching to HSB requires two settings. Users should load the HSB bitstream onto the board, pass backend=”hsb_hwhs” to the controller and add peer=hsb to the coprocessor’s config.
The team notes that HSB uses one-sided RDMA writes to reach the CPU or GPU, a design choice that contributes to its performance characteristics. “Since everything else stays the same whichever RDMA engine sits on the FPGA, you can choose the one that best fits each experiment with a simple configuration change,” the researchers explain, highlighting the flexibility of the Backline system.
The choice between ERNIC and HSB depends on the specific experimental requirements; HSB excels in scenarios demanding the shortest round trip latency over a direct attach cable, while ERNIC offers reliable connections and retransmission capabilities, offering clear guidance for selecting the optimal RDMA engine for specific quantum computing tasks.
The observed reduction in latency translates directly to increased decoding capacity. Shorter round trips free up more processing time for the decoder. The 1.74 µs round trip latency achieved with HSB, therefore, represents a tangible improvement in the efficiency of quantum control systems. This optimization is a key component of Backline’s goal to provide a low-latency environment for quantum computing tasks, enabling faster and more complex calculations.
The researchers detailed the architecture and original measurements in a published paper, offering a technical overview of the system’s capabilities and performance characteristics, the company says. A physicist and software developer leading the performance team at Xanadu notes the potential for further optimization and customization. He and his colleagues encourage users to explore the technology and contribute to the ongoing development of Backline, offering a pathway for integrating custom decoders and RDMA engines.




See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
