NVIDIA DAQIRI 2026.11: Direct GPU Processing of 500Gb/s Electron Microscope Data

·2026.11.21 17:05

Key point

NVIDIA has released DAQIRI 2026.11, a GPU-based data acquisition library capable of processing 500Gb of data per second.

1 / 4

Details

  • NVIDIA released the open-source library DAQIRI v2026.11.00, which writes sensor data directly to GPU memory.
  • In a case study at Lawrence Berkeley National Laboratory, over 500Gb of electron microscope data per second was processed in real time without disk storage.
  • The introduction of the ibverbs engine increased 256B packet throughput by approximately 2x compared to DPDK (99Gbps) and reduced control latency by 1.8~2.7µs through direct polling.
  • Hardware packet reordering is supported on ConnectX-7 or higher NICs, eliminating software copy costs.

Technical Features

  • Utilizes GPUDirect RDMA to allow the NIC to write data directly to GPU memory without passing through the CPU.
  • Adds a hardware reordering feature on ConnectX-7 or higher NICs that reads sequence fields to directly DMA write payloads to their final locations.
  • Reduces device control feedback latency with a direct polling mode where application threads directly handle packet transmission and reception.

Compatibility and Requirements

  • Requires Linux kernel 5.15 or higher and a GPUDirect-capable GPU; ConnectX-6 Dx or higher NICs are recommended.
  • ABI version changed to 3, requiring existing users to rebuild; Apache License 2.0 applies.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.