Lucid Demos 200 Gb/s RDMA Camera Link and Triton All-in-One

Twenty cameras, dual 100 Gb/s uplinks, and 200 Gb/s of sustained image throughput into an NVIDIA DGX Spark — that is the headline number from Lucid Vision Labs' RDMA demonstration recorded at VISION 2026 in Stuttgart. Alexis Tiessie, director of product management at Lucid, used the session to map two trends now shaping machine-vision system design: moving ever-larger image streams across multi-camera networks, and collapsing camera, optics and lighting into single deployment-ready units.

What does the 20-camera demonstration actually move?

The Giga 3.0 RDMA architecture connects 20 cameras through a network switch and links the switch to the DGX Spark system via two 100 Gb/s uplinks. Aggregate throughput reaches 200 Gb/s. The critical detail is how the data lands: RDMA (remote direct memory access) places image data directly into system memory without per-packet CPU intervention.

That distinction matters at this scale. Conventional packet processing forces the host CPU to handle every frame arriving from every sensor, and at 200 Gb/s the processor becomes the bottleneck long before the network does. Bypassing per-packet CPU work frees the compute budget for the actual inference or inspection task running on the DGX Spark.

How is synchronization handled across the network?

Tiessie's second theme was timing. As vision systems grow past a single sensor, architects need a shared time base across the whole network, not just adequate bandwidth per camera.

Lucid's approach layers two mechanisms:

• Hardware-based PTP timestamping (the IEEE 1588 Precision Time Protocol, which distributes a common clock over Ethernet) combined with camera synchronization to establish a network-wide time base.
• E-Trigger technology, which takes a trigger received by one camera and distributes it to other cameras on the network.

According to the company, actions can be coordinated within microseconds depending on network configuration. The practical payoff is cabling: distributed triggering removes the need to run extensive dedicated trigger lines to every camera position on a line or in a cell.

What is the Triton All-in-One?

The presentation also introduced the Triton All-in-One, an integrated imaging system launched in Stuttgart that combines a camera, a focus-tunable liquid lens, and illumination in one factory-ready package.

The featured model builds on Lucid's Triton platform and pairs:

• A 2.5 GigE camera (2.5 Gigabit Ethernet interface)
• Sony's 3.2-megapixel IMX900 sensor
• A software-controlled liquid lens for focus adjustment at varying object distances
• Configurable, independently controllable left and right illumination channels

The split illumination control is the detail inspectors will care about. Directing light from one side at a time can reveal surface features that flat, symmetric lighting hides — embossing and other subtle topographic details are the examples Lucid cites.

Which applications is it built for?

Lucid positions the unit at barcode reading, logistics automation, and general machine-vision inspection. Integration is the selling point: camera, optics and lighting ship as one system, which the company says simplifies deployment where separate component selection and mechanical alignment would otherwise slow a project down.

The platform also supports remote focus adjustment and focus stacking (combining multiple frames at different focal distances into one fully sharp image). System designers can scale the attached compute resources to match the application rather than the camera's fixed capability.

The open question

The throughput figures come from Lucid's own demonstration configuration — a controlled showcase at a trade show, not an independent lab result — and buyers will want to verify sustained rates in their own network topologies. The broader question the presentation raises is one of architecture adoption: as multi-camera systems push past 100 Gb/s, does RDMA-to-GPU-memory become the default machine-vision data path, and does integrated hardware like the Triton All-in-One push component-level selection further out of the integrator's workflow?