A digital twin without real-time data is only a 3D model

Daniel Litwin opened with a blunt question for Annie Yang at the Premio booth at Embedded World North America 2026: what separates a digital twin from a random 3D rendering? Yang, a product marketing specialist at Premio, gave an answer that puts a hardware decision in front of every plant team planning one. A model is a digital twin only when it is wired to live data.

Speaking in a recent interview recorded at the booth, Yang described a digital twin as a virtual version of something real: a machine, a production line or an HMI system, the human machine interface screen an operator uses to run equipment. The difference between that and a picture is a connection to real-time information that lets a team monitor, visualize and understand what the physical asset is doing. The wiring that makes the connection, in her telling, is the edge computer.

The timing explains why the booth conversation centered on hardware. Nvidia, whose Jetson modules Premio builds into its JCO series, says in its own digital twin glossary that a new class of AI-enabled twins now connects to enterprise data and production systems at the edge, and its RTX Pro server material tied to its GTC 2026 announcements describes the RTX Pro 4500 Blackwell Server Edition as carrying GPU acceleration from the data center to edge computing platforms. Yang told Litwin that the newest GPU Premio supports is the NVIDIA RTX Pro Blackwell. For an operations manager, that means the acceleration layer of a twin can now sit in a box on the plant floor rather than in a server room.

Two kinds of twin, and only one can answer how the line is running now

Premio splits digital twins into two uses: simulation before a build, and operation on live data. Yang's example for the first is a factory that does not exist yet, where a team builds the twin first, tests scenarios and catches problems before pouring concrete. The second is the one she says edge computing exists for.

The mechanism she lays out is a loop with five stations. Cameras and sensors on the physical asset collect data; the data goes to an edge computer, a hardened PC installed near the machines rather than in a data center, which processes it close to where it was generated; the processed data goes to the digital twin; the twin surfaces a pattern or a problem; and a person or a system acts on it. New data then starts the cycle again.

Nvidia's glossary fills in what the twin is built from. On its own site the company says twins combine 1D data, meaning tabular records from IT and OT systems, with 2D and 3D data such as CAD files, reality-capture scans and BIM, the building information model an architect or contractor keeps for a facility, and that IoT sensors are what keep the twin accurate over time. For the operator, the split matters: the edge computer in Yang's loop handles the 1D stream of sensor values, while the geometry comes from the engineering side and does not change minute to minute.

Connect: four port types, a slot for more, and cables that exit downward

Premio's edge computers carry Ethernet, COM, USB and DIO, with an Edge Boost design that adds connectivity modules. DIO is digital input and output, the simple on-off signal lines that read a limit switch or trip a relay, which is how most legacy machines report state. Yang says the Edge Boost design is there for a customer who runs out of built-in ports.

She also singles out the RCL and BCL series for ground-facing I/O, meaning the connectors sit on the underside of the unit so cables drop straight down. Her reason is practical: it makes installation and maintenance easier in a cabinet or on a mount. A technician swapping a unit in a panel does not have to fish cables around the sides.

One thing a buyer will not get from the interview is the messaging layer. Yang names the physical ports, and Premio's interview does not name the protocol, such as OPC UA or MQTT, that would carry sensor values from the box to a twin platform. OPC UA is the industrial automation standard for exchanging machine data between vendors' systems, and MQTT is the lightweight publish and subscribe protocol most IoT platforms accept, so the twin software, not the hardware, usually decides which one runs.

Process: Nvidia's own numbers put the workload range at 157 to 2,070 TOPS

Nvidia's Jetson page lists the Orin NX at up to 157 TOPS and the Thor series at up to 2,070 TOPS. TOPS stands for trillions of operations per second, the standard yardstick for how much AI inference a chip can run. The same page puts the power draw at 10 to 25 watts for Orin NX and 40 to 130 watts for Thor, which is a 13-fold spread in compute inside one module family that Premio ships in its JCO series.

That spread is the reason Yang gives for offering three processor architectures rather than one.

Her map of the catalog runs Intel on the RCO and VCO series, NVIDIA Jetson on the JCO series, and AMD as a third option. When Litwin asked whether a single Premio box carries enough compute to handle every data source, her answer was no: it depends on the workload, and a 3D visualization job or an AI-heavy job calls for a higher class of processor. A plant that only wants a vibration twin on a pump and a plant that wants a rendered line walking in sync with the real one should not expect to buy the same unit.

Nvidia's own material shows how far that range has moved in one generation. On the Jetson page the company says Thor delivers up to 7.5 times the AI compute of the Jetson AGX Orin, with 3.5 times the energy efficiency, inside a 130 watt envelope, and that its Orin Nano 2 nearly doubles the inference of its predecessor while drawing 40 percent less power. Yang's pitch that workloads differ is, in effect, a pitch that the same twin can be specified at several price points.

Accelerate: two full-length GPUs in one rugged chassis

Premio's VCL, RCL and JCL series hold up to two full-height, full-length GPUs. Full-height, full-length is the largest standard PCIe card form factor, the size of a workstation graphics card rather than a half-height server card, so a chassis that takes two of them has been designed around their cooling and power from the start. Yang said the newest GPU in that lineup is the NVIDIA RTX Pro Blackwell, and that the RCO series supports GPUs up to the RTX Pro as well.

What the card adds comes from Nvidia's side. On its RTX Pro server page the company describes Blackwell as its current architecture and lists industrial AI and digital twins alongside agentic AI and large language model inference as the workloads it targets, and it says the RTX Pro 4500 Blackwell Server Edition brings that acceleration to the enterprise data center and edge platforms it calls the most widely adopted. The nearest performance figure Nvidia publishes for Blackwell at the edge is the Jetson Thor comparison above, since Thor is also a Blackwell GPU.

For the operator, the GPU is the layer that makes the twin visual and predictive rather than tabular. Yang ties GPU support to 3D visualization and to what she calls very demanding AI workloads, which in the feedback loop means the step where patterns get found and problems get flagged before a person sees them.

Operate: four rugged traits, a UL mark, and no truck roll

Yang names four rugged traits: wide operating temperature, wide power input, shock and vibration resistance, and UL listing. A UL listing means UL, the safety certification body, has tested samples of the product against its applicable safety standard and allows the mark on the unit, which is what an electrical inspector or a corporate safety officer looks for before a box goes into a panel.

The interview gives the categories, not the figures. Premio does not attach a temperature range, a shock rating or a voltage range to any series in the conversation, so a buyer has to pull those from the datasheet for the specific RCL, VCL, BCL, RCO, VCO, JCL or JCO unit under consideration. For a sense of what the category publishes, Nexcom, another maker of rugged Jetson-based edge computers, says on its own product page that its units run at minus 20 or minus 30 to 70 degrees Celsius with an IP67 rating on selected models, meaning sealed against dust and against temporary immersion in water.

The trait Yang spends the most time on is the one that is not physical. Selected Premio products carry remote management, and her example is a twin deployed somewhere a technician cannot easily reach.

Scale Computing, in a white paper on its own site dated June 30, 2025, frames the same problem from the IT side: it describes manufacturers trying to run edge workloads at potentially thousands of sites with little or no IT staff. A remote console is Premio's answer to that staffing gap for the hardware it sells.

The twin itself is built in software, and the edge box is one layer of it

Nvidia's Omniverse material makes clear that the twin layer lives in software, not in the edge computer. On its own site the company describes Omniverse as a set of libraries for OpenUSD data interoperability, RTX rendering, PhysX physics simulation and asset validation, and its digital twin glossary shows a ship twin visualized in Siemens Teamcenter X using Omniverse APIs. OpenUSD, the open Universal Scene Description format, is the file structure those tools share so that a scene built in one application can be read in another.

That is the honest limit on the interview's thesis. An edge computer can connect sensors, process the stream, accelerate inference and survive the plant, and it still does not decide what the data means; the physics, the rendering and the scene structure sit in a platform like the ones Nvidia and Siemens name. Yang's own description agrees with that boundary: in her loop the edge computer sends processed data on to the twin, and it is the twin that finds the pattern, and she does not tie Premio to any one platform.

The Nexcom page also shows that the four-stage pitch is a category pitch. Nexcom describes its Jetson line as spanning 20 to 275 TOPS, with industrial interfaces built in, a wide temperature range, and an in-house Linux build that exposes CAN bus, I/O and power management to AI applications. A buyer comparing rugged edge vendors should expect the connect, process and operate stages on every datasheet and should compare on the specifics: port counts, GPU slots, temperature figures and the management console.

Edge or cloud: the June 2025 case for keeping the loop local

Scale Computing's white paper argues the latency case without a stopwatch figure. It says the endpoints and IoT devices a factory adds generate data sets large enough to cause high latency, bandwidth congestion and bottlenecks in centralized networks, that localizing compute cuts the time to process and act on data for robotics, automation and predictive maintenance, and that in manufacturing any downtime becomes extremely costly. Yang's version is shorter: the edge computer processes operational data closer to where it is generated, and that proximity is what makes the loop real time.

The same white paper complicates the hardware story. Scale Computing says plants that already run edge gear end up with hardware from many technology providers, that interoperability then depends on resources from multiple OT vendors, and that the result is complexity that drives up the cost of managing and maintaining the edge. The company's remedy is standardizing on a platform; Premio's remedy, in Yang's telling, is a consistent I/O design across series and remote management so each box can be watched from one control room.

Neither source puts a dollar figure on edge versus cloud for a twin, and neither claims the cloud goes away; Scale Computing's point is that smart factory teams decide what runs in the cloud, in the data center and at the edge, and Nvidia's own servers sit in the data center. The edge box, by both accounts, is where the loop closes fast enough to adjust a machine.

Litwin closed by telling viewers that anyone launching an operational twin to capture data at the source of a factory operation will need edge computing in the mix. The test Yang gives is the one to carry into a vendor meeting: ask whether the twin can answer how the machine is working right now, and if the answer depends on a round trip to a data center, the loop is not yet operational. Premio points buyers to premioinc.com for the series-level specifications the interview leaves off.