Throughput and latency get treated as two words for one property. One asks how much can be moved per minute; the other asks how long the slowest packet waited. When device counts climb, the second one usually breaks first.
Throughput is an average: gigabytes per hour, and a couple of stalls in between do not change the figure. Latency is an extreme value: the time a single packet spends between ingress and egress, and the longest queue wait is the figure that matters.
The two do not scale together. Split bandwidth across ten devices and each gets a tenth, a linear loss. Queue time is not linear: at seventy percent utilization it looks fine, at ninety percent it starts to jump. A speed test reading 30 Mbps while a screen spins is consistent rather than contradictory, because throughput is still sufficient while the queue is already long.
Device count acts on the two metrics differently. Another device takes a share of the bandwidth, a straight subtraction. Every device also needs a transmit opportunity inside the same time window, and near saturation those opportunities stop arriving evenly. Jitter is layered on top of the subtraction.
Cellular and Wi-Fi are both shared media. Other subscribers in the same cell and other clients on the same access point change when a given device gets its turn. Add devices at a remote site and the first symptom is usually a click that waits, not a transfer that stalls.
Insufficient bandwidth moves less data, and latency can stay normal. Queuing has enough bandwidth and still long waits, because packets sit behind a queue. Poor link quality, from weak signal or interference, brings retransmissions and degrades both numbers at once. One further case sits outside the device: a plan that grants a high-speed allowance and then reduces the rate until the billing cycle resets. The link never drops, the speed test falls, and every service slows together.
Triage is straightforward. Measure in-site latency and internet-bound latency at the same moment; normal in-site with slow outbound points to the egress segment, and slow in both means continuous heavy traffic inside the site. Record idle latency and loaded latency as well, and subtract one from the other: the difference is queueing. A single ping reporting the minimum round trip cannot show this.
The cellular side deserves a second pass: measure with the cabinet door open, with it closed, and with the antenna feeder routed outside, then compare the three readings.
Cellular links are asymmetric. LTE Cat 4 is rated at 150 Mbps down and 50 Mbps up, while 5G SA puts theoretical uplink in the 900 Mbps range. Video backhaul, telemetry reporting and bulk delivery all sit on the uplink, so a headline downlink peak says nothing about an uplink-bound load.
A common field case: during commissioning, a phone's cellular data is tethered to a portable router and several devices share that single uplink. Nothing feels wrong with two devices, and the slowdown appears once they all start working. The asymmetry explains it. Before adding a camera, run the arithmetic: a 1080p main stream at 2 Mbps sends about 21 GB per day when running continuously. Add a second live view and one historical playback and the uplink is full; PLC polling packets fall to the back of the queue, and latency rises from there.
The USR-G816 keeps more headroom on that side, with a theoretical SA uplink of 900 Mbps (a rated figure rather than a field measurement), four gigabit ports one of which can be reconfigured as LAN, dual SIM with automatic failover and failback across cellular, wired and Wi-Fi, and email or SMS alerts for offline status, weak signal and traffic overrun.
"Devices supported" means associated wireless clients, not devices busy at the same time and not sessions. One camera holds two or three streams and one host computer holds a dozen long connections, so twenty devices can mean forty sessions or four hundred.
The USR-G809s is rated for 256 clients across 2.4 GHz and 5.8 GHz, with eight gigabit RJ45 ports plus two SFP slots, which keeps in-site lateral traffic on the LAN side instead of the uplink. The USR-G806w offers 100 Mbps Ethernet and 2.4 GHz Wi-Fi with up to 20 wireless clients, adequate for three to five devices in one cabinet, though adding local video recording turns its own 100 Mbps ports into the new bottleneck. Which column applies depends on how many devices work at once. Session counts need not be guessed: the real-time session page shows current connections, and a screenshot taken during the busiest hour gives a better estimate than a device count.
The order is to move bulk traffic off the uplink first, then prioritize what remains. Storing video locally, and scheduling batch upgrades and historical playback into the night window, often removes the need to replace hardware at all.
If it still lags, prioritize at the egress device: critical traffic first, everything else capped, with caps applied by IP range or MAC address rather than guessed service names. Set the cap loose, watch critical traffic latency for two days, then tighten, because rate limiting itself introduces retransmissions. Whether a decision should run on site is the same question in another form: local configuration puts thresholds and interlocks on the device, so they still trigger when the network is down instead of waiting for a cloud poll, and the USR-EG628 class of edge computer ships with local configuration and Node-RED.
Acceptance takes two actions: run critical traffic during the busiest hour and record latency and packet loss, then pull the primary link and time the recovery. Put both figures into the handover document.
Throughput answers whether the data can be moved, latency answers whether the wait is acceptable, and device count decides which of the two reaches its limit first. For an industrial wireless router, the order is to count devices working at once, then check uplink headroom, and compare ports and rates last.
1. What separates an industrial wireless router from a consumer router?
Not peak rate, but sustained operation and controllability. Consumer buying advice centers on several clients streaming video and on coverage; the industrial side needs wide-voltage input, wide temperature range, DIN rail mounting, a watchdog, and traffic that can be governed by rules and priorities. When the network misbehaves, one is fixed by a reboot, the other has to be diagnosable down to the segment.
2. How far can real throughput fall below the rated figure?
Half or more is not unusual, since rated values describe an air interface or port ceiling and the real path pays for protocol overhead, retransmission and scheduling waits. The useful test is not comparing two numbers but loading the site's own traffic during its busiest hour and watching how much is lost when one more device joins.
3. How can insufficient bandwidth be told apart from excessive latency?
Measure in-site and internet-bound latency at the same moment. Normal in-site with slow outbound points to the egress segment; slow in both means continuous heavy traffic inside the site. Then compare idle and loaded latency: a wide gap points to queueing and buffering, where more bandwidth rarely helps.
4. Does a 256-client rating mean 256 devices can be connected?
That figure counts associated wireless clients, not concurrent throughput and not sessions. Twenty devices may hold forty sessions or four hundred, depending on how many connections each opens. Estimating from devices working at once, plus the session count at the busiest hour, is more reliable than reading the capacity column.
5. Is 5G always lower latency than 4G?
The air interface usually is, but end-to-end latency is set by the slowest segment. Insufficient uplink, long routing queues and slow tunnel re-establishment can cancel the advantage. For a remote site carrying only a few dozen telemetry points, 4G headroom is often easier to estimate.
6. How much uplink does one camera at a remote site need?
Convert bitrate to volume: bitrate times seconds, divided by 8, divided by 1024, gives daily traffic. A 1080p main stream at 2 Mbps sends roughly 21 GB per day and about 600 GB per month. When the uplink covers only one stream, keep recording local and pull the sub stream or event clips remotely.
7. Is data lost during the seconds a link fails over?
It depends on which side holds the buffer. A device-side buffer backfills after the link returns, while a platform-side backfill depends on the platform's reconnect logic, so both need confirming before deployment. Count failover time by service recovery rather than by link switching, since tunnel re-establishment is usually slower.
8. What should be confirmed before choosing a model?
Four items: devices working at once and the session count at peak, real uplink demand, port count and whether port roles can be reassigned, and what restores service after a link drops. Wide voltage, wide temperature and a watchdog are baseline requirements; they do not change the capacity calculation, but they decide how long the unit survives in a cabinet.
