When a machine leaves the factory, the networked things are easy to count: one PLC, one HMI panel, one host computer. Open the same cabinet two years later and the list changes. The line picked up barcode scanners and handheld terminals, an AGV joined the dispatch system, two cameras at a critical station, an energy monitoring module, and a maintenance channel requested by the customer's IT group. Three devices became twenty-something and the network starts misbehaving: dropouts, latency spikes.
The instinct is to find one number on the datasheet: how many devices it can carry. The capacity of an industrial cellular router is not set by any single number. Four lines run side by side, and which tops out first depends on what the site is doing.
Three different counts get mixed together. The online device count is the number of entries in the address table. The session count is the number of connections at the same moment. The wireless client count is a separate figure on the Wi-Fi side. A camera holds one or two streams, a host computer may keep a dozen open at once, and a PLC poll never really ends. Twenty devices can mean forty sessions, or four hundred.
The wireless side usually reaches its limit first. Handheld terminals, inspection tablets and AGVs sit on Wi-Fi, and client capacity is independent of how many Ethernet ports the box has. The USR-G806w carries up to 20 Wi-Fi users; the USR-G809s is rated for 256 clients in theory. Count the wireless terminals first, then see which tier they land in.
More accurate is looking at what is connected right now. The real-time session view shows terminal connection state and abnormal connections at a glance; static MAC and IP binding matches every unit to a name. Only then does the question of how many are actually running have a firm answer.
Only the outbound half consumes the cellular plan. Traffic between devices on site presses on the LAN-side ports and the backplane, and the two cannot be added. With cameras and the recorder on the LAN side, recording stays on local switching and routine upstream is close to zero; remote viewing and cloud backup consume it. If every device reports to the platform on its own, twenty units of concurrent upstream is a different calculation.
Measurement beats the datasheet. Run an upstream test in the busiest period, morning, noon and evening, and take the steady value rather than the peak. Repeat with the antenna outside the cabinet, since a metal enclosure eats mostly upstream. The daily and monthly traffic bar charts give a forecast of monthly usage and deserve a look before the SIM is active.
Running out of ports is a physical limit: add a switch, and it has nothing to do with forwarding capacity. Two other things are overlooked more often: whether port roles can be reassigned, a WAN port switched to LAN takes one more device with no need to change model for a port; and distance. A single copper run tops out around 100 meters, so a point several hundred meters away comes back more reliably over an SFP fiber port, which also breaks the ground potential difference.
One more class of device does not use Ethernet: instruments and older machine tools have only serial ports, and with RS232 and RS485 on the router the cabinet holds one less converter and one less power feed.
Someone once considered building a router out of a small computer with two Ethernet ports, on the grounds that a real IP-layer firewall was wanted rather than filtering at the DNS layer, which collapses the moment a device has a resolver hard-coded into it. The same boundary shows up on industrial sites: filtering by domain name does not hold back acquisition modules and cameras that only work with IP addresses, and rules only bite at the layer of IP, MAC and port.
Rule count grows with device count: allow rules are added one at a time, one catch-all deny goes last, and twenty devices mean twenty-odd rules. Forwarding has to carry bandwidth and rules at the same time. That is what performance means on the industrial side: not a benchmark score, but whether things hold at full load.
A few places cause rework: matching by source MAC requires the source IP left empty, and the same in reverse; rule names should carry the device name, so a year later the answer to who a rule was for still exists; changes take effect only after a restart in the system. Acceptance is one line: devices on the list reach their destination, devices off it do not.
When bandwidth runs short, split traffic before enlarging the plan. QoS flow control and per-terminal flow control give critical traffic the right of way, and policy routing sends public-network and private-network data out separate exits, which costs less than expanding everything.
The USR-G806w fits small sites where cabinet space is tight and the device count is low: 104 x 102 x 28 mm, Wi-Fi 802.11b/g/n at 2.4 GHz up to 300 Mbps with up to 20 Wi-Fi users, one WAN plus two LAN or all three configured as LAN, 10/100 Mbps auto-negotiating, DC 9-36 V wide input with reverse polarity protection, -20 to 70 degrees Celsius.
The USR-G809s fits sites where the device population has grown: two ports configurable as WAN or LAN plus six LAN, plus two SFP slots, one of which shares with a copper port, all RJ45 at Gigabit, Wi-Fi 6 dual-band 2 x 2 MIMO rated for 256 clients in theory, RS232 and RS485 terminals for instruments, DC 9-60 V wide input with reverse polarity protection, -25 to 75 degrees Celsius.
The USR-G816 fits sites with no broadband in reach, where everything leaves over cellular: a Qualcomm quad-core processor with the X62 modem, 5G SA and NSA, one Gigabit WAN that can be reconfigured as LAN plus three Gigabit LAN, dual SIM with the second slot optionally an embedded eSIM, automatic switching to a backup network when the primary link drops, three alarm classes including traffic overrun, DC 9-36 V wide input.
The first is the measured upstream value that holds steady, plus antenna placement. The second is session count and packet loss at the maximum online device count: load the site to its design limit and test again, rather than accepting it with three devices. The third is recovery time with a stopwatch after the primary link is pulled, plus a check of whether on-site traffic retransmits or stops during that wait.
Once the count climbs, choosing an industrial cellular router is no longer about one client capacity figure. It is about which of the four lines, device count, traffic, interfaces and rules, tops out first. Measure first, choose second; reversing the order overspends on bandwidth and keeps rework going on the rules.
The line is the SIM card slot, and whether the unit can reach the internet on its own at a site with no broadband. Also, 5G is not the same as 5 GHz Wi-Fi: the first is the fifth generation of cellular network, the second is a frequency band of Wi-Fi, and the two numbers mean entirely different things.
Decide on upstream, not downstream. For a site that only reports status values and minute-level telemetry, 4G is enough. Where several video streams go back at once, or programs are pushed out in batches, only 5G upstream holds up. Where 5G coverage does not reach, the link falls back to 4G on its own, and that fallback belongs in the plan.
Confirm that the data card is permitted for use in a router, since some carriers require the IMEI to be registered. The APN decides how the connection is made, and private-network or VPDN cards have to use the APN the carrier supplies. A wrong APN shows up as good signal with no connection.
Cellular networks usually hand out carrier-grade private addresses, so connections initiated from the public internet cannot get in and port forwarding is not an option. The arrangement is reversed: the router builds an encrypted tunnel outward on its own, or the remote management platform opens the built-in web page from a distant location, and nothing inbound has to be opened on site.
Security comes from tunnels and rules, not from the medium. Three things belong on the device side: a VPN tunnel carrying the remote channel, outbound traffic allowed destination by destination with everything else denied, and the default password on the management page changed at first login.
One, changing plan or carrier without changing hardware. Two, automatic switch to the second card when the base station serving the first is congested or down. Two cards from different carriers also avoid local dead spots. Whether data is buffered on the device or backfilled by the platform during the switch is worth confirming once during commissioning.
It depends on which side holds the buffer. Storing collected data on the machine or the gateway and backfilling after recovery is common practice, while a link that relies on live upload with no local buffer loses whatever happened during the outage. The acceptance action is to pull the primary link once, stopwatch the recovery time and check the backfill.
Device count, session count and wireless client count are three separate figures. Start with how many units need Wi-Fi and check that against client capacity, then look at how many connections exist at one moment using the real-time session view, and finish with concurrent upstream volume. If n
