Editorial Technical Reference

Edge AI gateways

Edge AI gateways for video analytics, industrial IoT and remote sites · NPU TOPS and supported data types, CPU, RAM and storage, 5G/4G, Wi-Fi, LoRa and Ethernet, RS232/485 and I/O, DC input and power draw, operating temperature, IP rating, Docker and algorithm libraries

An edge AI gateway is a rugged computer that connects cameras, sensors and machines on site, runs AI inference — counting, detection, recognition — locally, and sends results to the cloud over cellular, Wi-Fi or Ethernet. Ask three things first: how many video streams or sensors it must analyse and with which models, which sets the NPU performance and memory; which networks and field interfaces it must connect; and the power supply, temperature and enclosure at the installation.

Edge AI gateways and AI boxes — industrial computers with NPU or GPU acceleration, cellular and wired networking and field I/O — for video analytics and industrial IoT. Plain IoT gateways and routers without AI acceleration, AI modules and chips sold alone, cameras with built-in AI, and data-centre servers are outside this page except where compared.

CNFX editorial research · Sources checked October 12, 2026 · 4 sources

Size AI performance Match interfaces and site

Edge AI Gateway in a manufacturing environment — Representative product image. Confirm appearance and specifications with the manufacturer.

Start with your task

Size AI performanceTOPS, memory. Match interfaces and siteNetwork, I/O, power.

Size AI performance, memory and software

Hoditek, an IoT equipment maker, lists the Smart3100 5G edge AI gateway with an octa-core CPU, a 6 TOPS NPU, 8 GB RAM and 32 GB eMMC running Linux, Docker container support and secondary development, and an algorithm library covering face, fall, flame, people-flow and counting and licence-plate applications. [1] For comparison, Rockchip's RK3588 processor combines quad Cortex-A76 and quad Cortex-A55 cores, a Mali-G610 GPU and a triple-core 6 TOPS NPU supporting INT4, INT8, INT16, FP16, BF16 and TF32, with 8K video decoding and multiple camera inputs. [2] At the high end, NVIDIA's Jetson Orin family has seven modules with the same architecture offering up to 275 TOPS. [3]

Editorial note: ask the supplier at which precision its TOPS figure is stated, and to run your model on the box and report frames per second, accuracy and power.

Give the supplier the AI workload (models, number of camera streams or sensors, frame rate), required TOPS and precision, CPU, RAM and storage, operating system and container support, networks (5G/LTE bands, Wi-Fi, LoRa, Ethernet ports), serial and digital/analogue I/O, power input range and budget, operating temperature, IP rating and mounting, certifications for target markets, remote management, and quantity.

Match connectivity, interfaces, power and environment

The Smart3100 offers 5G NR SA/NSA or LTE Cat 4, Wi-Fi, LoRa and Bluetooth through an eight-antenna design with link backup across 4G/5G, LoRa, Wi-Fi and Ethernet; two Gigabit LAN plus two Gigabit WAN/LAN ports; RS232/485, USB, 4 analogue and 8 digital I/O channels; DC 9–36 V input with ignition (ACC) sensing, drawing 330 mA idle and 750 mA during data transfer at 12 V DC; IP43 protection; and operation from −30 to +70 °C with storage from −40 to +80 °C. [1]

Editorial calculation: 330 mA and 750 mA at 12 V are about 4.0 W idle and 9.0 W while transferring data — about 0.2 kWh per day if running at the higher figure around the clock, which sets the size of any solar or battery supply at a remote site. [1]

Compare a low-power DIN-rail AI gateway

Robustel's EG5120 specification lists a 5G/4G industrial edge AI gateway built on a quad-core Cortex-A53 at 1.6 GHz with a 2.3 TOPS NPU, 2 GB or 4 GB LPDDR4 and 64 GB eMMC, running RobustOS Pro (Debian 11) with Docker container support. It has dual mini-SIM cellular with link backup to Ethernet, 1 WAN + 1 LAN or 2 LAN ports, two software-configurable RS-232/RS-485 ports, 2 digital inputs and 2 wet-contact digital outputs, and optional Wi-Fi, Bluetooth and GNSS. It takes 9–60 V DC with reverse-polarity and surge protection and draws 5 W idle and 15 W maximum at 12 V, in an IP30 metal housing of 49 × 105 × 128 mm and 567 g for DIN rail, operating from −40 to 70 °C. [4]

Set this low-power gateway against the 6 TOPS gateway, the RK3588 processor and the Jetson Orin modules described above; the gap in NPU performance is matched by a gap in power and thermal design. [1] [2] [3]

Editorial calculation: 2.3 TOPS is about 38% of the 6 TOPS NPU above; at its 15 W maximum the EG5120 delivers about 0.15 TOPS per watt of total gateway power, and 15 W at 12 V draws 1.25 A.

EG5120 parameter (Robustel)	Published value
NPU	2.3 TOPS
CPU	Quad-core Cortex-A53, 1.6 GHz
Memory	2 or 4 GB LPDDR4 + 64 GB eMMC
Power input	9–60 V DC
Power consumption at 12 V	5 W idle; 15 W max
Operating temperature	−40 to 70 °C
Housing	Metal, IP30, 49 × 105 × 128 mm, 567 g

Ask the supplier for the NPU rating your models actually achieve, the maximum power at your input voltage, and the operating temperature with the cellular module transmitting.

Compare what suppliers offer

Suppliers linked to this page list edge AI gateways; each card links to the page reviewed.

Listings include 5G edge AI gateways and AI boxes, edge computing gateways and edge routers; ask each for benchmark results on your models, interface lists and certifications.

Trace the gateway

This is a functional reference built from the sources cited; it is not one maker's design.

Element	Role	Fix in the order
SoC / NPU	6 TOPS to 275 TOPS examples. [2] [3]	TOPS, precision.
Memory and storage	8 GB RAM, 32 GB eMMC example. [1]	RAM, storage.
Networks	5G/LTE, Wi-Fi, LoRa, Ethernet. [1]	Bands, ports.
Field I/O	RS232/485, AI/DI. [1]	Channels.
Power and enclosure	9–36 V DC; IP43; −30 to +70 °C. [1]	Input, IP, temperature.

Fill the specification fields with values for your project

An edge AI gateway quotation needs values for your project. The table explains each field, gives published values where a reviewed source states them, and says what to obtain from the supplier.

Field	What it means and published values	Ask the supplier for
AI performance (TOPS)	6 TOPS example; up to 275 TOPS modules. [1] [3]	TOPS, precision, FPS.
Processor type	Octa-core; A76 + A55 example. [1] [2]	SoC.
Memory	8 GB RAM example. [1]	RAM.
Storage	32 GB eMMC example. [1]	Storage, expansion.
Power consumption	330–750 mA at 12 V. [1]	W by workload.
Operating temperature (°C)	−30 to +70 example. [1]	Range.
Network interfaces	5G/LTE, Wi-Fi, LoRa, 4× GbE. [1]	Bands, ports.
AI frameworks	Docker, algorithm library. [1]	Supported toolchains.
Input voltage (V)	DC 9–36 V. [1]	Range.
Ingress protection	IP43 example. [1]	IP code.
Certifications	Not stated in reviewed sources.	Market approvals.
Power consumption	Example: 5 W idle, 15 W max at 12 V (Robustel EG5120, 2.3 TOPS). [4]	Max power at your input voltage.

Send the workload, interfaces and site together

Paste it into your enquiry below or into an email to a maker, then replace every bracket.

Workload: [models; streams; FPS]
Compute: [TOPS; RAM; storage]
Network: [5G/LTE bands; Wi-Fi; LoRa; Ethernet]
I/O: [RS232/485; DI/AI; USB]
Power: [input range; budget]
Site: [temperature; IP; mounting]
Compliance: [certifications]
Quantity: [units]

See manufacturer records Send an enquiry through CNFX

Buyer questions

How much AI performance do edge gateways offer?

Examples range from 6 TOPS (Rockchip RK3588 NPU) to up to 275 TOPS in NVIDIA Jetson Orin modules. [2] [3]

What power does an edge AI gateway draw?

One 5G gateway draws 330 mA idle and 750 mA transferring data at 12 V DC, about 4–9 W. [1]

What precisions do NPUs support?

Rockchip's RK3588 NPU supports INT4, INT8, INT16, FP16, BF16 and TF32. [2]

How much power does an edge AI gateway use?

Robustel lists its 2.3 TOPS EG5120 at 5 W idle and 15 W maximum at 12 V, on a 9–60 V DC input, rated −40 to 70 °C. [4]

Sources and what still needs a direct check

Sources checked 2026-10-12. Hoditek, Rockchip and NVIDIA describe their own products; Hoditek also appears as a supplier on this page.

Hoditek — Smart3100 5G edge AI gateway ↗ Maker product page Return to text ↑
Rockchip — RK3588 processor ↗ Chip maker product page Return to text ↑
NVIDIA — Jetson Orin ↗ Module maker product page Return to text ↑
Robustel — EG5120 dual-SIM 5G/4G industrial edge AI gateway, specifications ↗ Maker product page Return to text ↑

Still to check yourself

Benchmark your models on the box.
Check networks and bands.
Confirm I/O channels.
Check power and temperature.
Verify the legal supplier and factory.