A fast Ethernet feed still leaves work for the receiver if its packets arrive out of order. NVIDIA’s DAQIRI can now assign that work to the network adapter.
In its November 19 report, NVIDIA describes a Berkeley Lab microscopy pipeline receiving more than 500 Gb/s of detector data directly into GPU memory. The GPU assembles frames, applies corrections and reduces them to spectra during acquisition.
The networking detail is in DAQIRI’s November 10 release: compatible ConnectX-7 or newer adapters can read repeating packet sequence numbers and write payloads into their ordered buffer positions. That removes a separate software reordering step. Hardware reordering is optional; software remains the default.
Control traffic gets a different option. Direct polling lets the application handle transmit and receive queues without handing work to an I/O worker. NVIDIA reports 2–4 microseconds of application send-to-receive latency in cabled loopback tests. That isn’t a measurement of the complete microscopy feedback loop.
The engineering takeaway: configure the bulk-data and control queues for their different jobs. Before enabling hardware reordering, check sequence-field compatibility and decide how incomplete batches should be handled. A correctly ordered buffer still needs an explicit policy for missing packets.


