Nvidia’s DAQIRI now feeds a 500 Gbps electron microscope camera straight into GPUs
Nvidia's latest DAQIRI release adds hardware packet reordering, a faster ibverbs engine and a low-latency polling mode, demonstrated on a Berkeley Lab camera that streams over 500 Gbps.
Nvidia has released an update to DAQIRI, its library for moving sensor data into GPU memory, and showed it on a Lawrence Berkeley National Laboratory electron microscopy camera that produces more than 500 Gbps, according to a developer blog post published November 19.
A camera that outruns the usual EELS data stream
The case study comes from the Molecular Foundry at Berkeley Lab. Electron energy-loss spectroscopy (EELS) measures how much energy electrons lose in a sample, which helps identify elements and how they are bonded. Those signals mix strong and faint components. The conventional way to capture both is two separately tuned acquisitions in quick succession, which needs tight coordination of electron optics, exposure timing and detector readout, and the results must then be aligned and merged.
Image: Nvidia
The new high-dynamic-range camera captures both at once. Nvidia says that simplifies the setup and microscope optics, but the stream is over 500 times faster than conventional EELS data. In the described pipeline, FPGA readout electronics send detector data over Ethernet, DAQIRI places the packets directly into GPU memory, and GPU code assembles and corrects frames before reducing them to energy-loss spectra. Nvidia says the computation also recovers submicron spatial resolution from pixels 5 micrometers across, and that researchers can inspect data and adjust acquisition parameters and regions of interest during the measurement.
Over 500 Gbps of detector data processed while the experiment is still running
Image: Nvidia
Advertisements
Three additions: NIC reordering, ibverbs, direct polling
The release adds three features. First, hardware packet reordering on supported ConnectX-7 or newer network cards. For compatible sequence formats, the card reads the sequence field and writes each payload to its ordered position, so the application gets assembled data without a separate software reorder copy. It is configured per receive queue, including the byte offset, sequence bit width, packets per batch and what to do with missing packets (the example drops partial batches on timeout).
Second, an ibverbs engine, built on libibverbs and NVIDIA DevX, which is now the default for raw Ethernet streams and can be selected with engine: "ibverbs". Nvidia says it delivers about twice the per-core throughput of the previous DPDK backend for packets up to 1024 bytes at rates above 200 Gbps. Those figures come from benchmarks on DGX Spark, IGX Thor and RTX PRO server platforms; the published chart is a single-queue run on IGX Thor and shows similar throughput for 8 KB packets.
Image: Nvidia
Third, a direct polling mode, enabled with poll_mode: "direct", in which the application thread submits and receives packets itself rather than handing off to an I/O worker. In cabled loopback tests, Nvidia reports application send-to-receive latency of “2-4 µs”. It can be applied to control queues while data queues keep batched processing. A benchmark, daqiri_bench_raw_latency, is provided so users can measure their own hardware.
How this compares with earlier pipelines
The Molecular Foundry has handled huge streams before. Its 4D Camera, a different detector, reads out at 87,000 frames per second and about 480 Gbit/s, according to a Microscopy and Microanalysis abstract. In that design the camera head feeds four FPGAs, which pass data to four receiver PCs, with a fifth PC for storage and analysis. The new EELS workflow instead lands packets in GPU memory on one path. That is a contrast in architecture rather than a measured head-to-head, and the Nvidia post does not compare the two.
DAQIRI is also not an entirely new bet. Nvidia describes it as part of Holoscan, and its developer page says it prepares UDP and RoCE v2 streams for CUDA-accessible memory for use with Holoscan, TensorRT and Omniverse. The code is on GitHub under NVIDIA/daqiri. The main alternative remains the DPDK path Nvidia itself used as DAQIRI’s previous default, and the ibverbs claim is measured against that backend, not against rival acquisition stacks. Nvidia’s own earlier post on DAQIRI also covered a CERN project, which suggests the library is aimed at more than microscopy.
Advertisements
What the post does not establish
The post is a developer article, and every performance figure is Nvidia’s own. The “approximately twice” gain applies to small packets at very high rates, not to all traffic. The 2-4 µs latency was measured on a cabled loopback on one host, not across a network or inside a full experiment. The EELS section gives no end-to-end latency for frames-to-spectra, no GPU model used at Berkeley Lab and no data on how many experiments ran this way, and the camera itself is described without a publication or release plan. Hardware reordering also depends on a compatible sequence format and ConnectX-7 or newer cards.
For developers, the practical entry point is the DAQIRI documentation, configuration examples and benchmarks linked in the post. Instrument makers with Ethernet-capable detectors can test throughput and latency on their own platforms before redesigning anything. Nvidia’s argument is that designers could stop shaping detectors around what can be stored or moved, and instead around what needs measuring.
To provide the best experiences, we use technologies like cookies to store and/or access device information. Consenting to these technologies will allow us to process data such as browsing behavior or unique IDs on this site. Not consenting or withdrawing consent, may adversely affect certain features and functions.
Functional Always active
The technical storage or access is strictly necessary for the legitimate purpose of enabling the use of a specific service explicitly requested by the subscriber or user, or for the sole purpose of carrying out the transmission of a communication over an electronic communications network.
Preferences
The technical storage or access is necessary for the legitimate purpose of storing preferences that are not requested by the subscriber or user.
Statistics
The technical storage or access that is used exclusively for statistical purposes.The technical storage or access that is used exclusively for anonymous statistical purposes. Without a subpoena, voluntary compliance on the part of your Internet Service Provider, or additional records from a third party, information stored or retrieved for this purpose alone cannot usually be used to identify you.
Marketing
The technical storage or access is required to create user profiles to send advertising, or to track the user on a website or across several websites for similar marketing purposes.
To provide the best experiences, we use technologies like cookies to store and/or access device information. Consenting to these technologies will allow us to process data such as browsing behavior or unique IDs on this site. Not consenting or withdrawing consent, may adversely affect certain features and functions.
Functional Always active
The technical storage or access is strictly necessary for the legitimate purpose of enabling the use of a specific service explicitly requested by the subscriber or user, or for the sole purpose of carrying out the transmission of a communication over an electronic communications network.
Preferences
The technical storage or access is necessary for the legitimate purpose of storing preferences that are not requested by the subscriber or user.
Statistics
The technical storage or access that is used exclusively for statistical purposes.The technical storage or access that is used exclusively for anonymous statistical purposes. Without a subpoena, voluntary compliance on the part of your Internet Service Provider, or additional records from a third party, information stored or retrieved for this purpose alone cannot usually be used to identify you.
Marketing
The technical storage or access is required to create user profiles to send advertising, or to track the user on a website or across several websites for similar marketing purposes.