Checklist
Running a compact industrial PC as an always-on server
Short answer
Yes, a compact fanless industrial PC can host small IoT, VPN and SCADA services for years, but its rugged build guarantees survival, not capacity. Qualify the exact model against measured peak load: CPU below 70%, memory below 75%, SSD write endurance with margin, thermal margin at the site, and a restore test that proves recovery.
On this page
- Can a compact industrial PC run as an always-on server?
- How much workload will the server actually carry?
- What CPU and memory does the exact model have?
- How long will the SSD last under continuous writes?
- Will the fanless case stay cool enough at the site?
- Can IoT, VPN and SCADA share one host safely?
- Will it really run non-stop for years?
- What must pass before a customer uses it?
- Questions asked
Yes, a compact fanless industrial PC can host small IoT, VPN and SCADA services around the clock for years: the sealed heatsink build, wide temperature rating and absence of wearing parts are exactly what a 24/7 cabinet box needs. What it does not give you is capacity or a runtime promise; you earn those by qualifying the exact model against measured peak load and by proving recovery. This checklist runs the qualification in the order you meet it: define the workload, confirm the CPU and memory you actually get, check storage endurance and thermal margin, decide what shares the host, and finish with the acceptance test that earns the word server.
Can a compact industrial PC run as an always-on server? #
Yes for small workloads: the build is made for 24/7 duty (sealed, fanless, silent), but the enclosure proves nothing about capacity. Treat the unit as a candidate server and qualify CPU, memory, storage, thermal margin and recovery against your measured load before a customer depends on it.
Server here means locally hosted services — a message broker, a VPN endpoint, a small SCADA host with its historian — not a cloud architecture, and the word carries no availability of its own. Fix the ground rules before touching hardware:
- Write down every service, its consumers and its dependencies: devices, clients, database, upstream links.
- Set recovery targets now: tolerable downtime and tolerable data loss per service.
- Adopt the pass rule used throughout: measured peak CPU at or below 70 percent, peak memory at or below 75 percent.
- Proceed to hardware checks only once those targets exist; every later check compares against them.
How much workload will the server actually carry? #
Measure it: run the planned combined stack under representative load — every device publishing, VPN in use, SCADA clients open — and record peak CPU, peak memory, disk writes and network throughput. Idle numbers tell you nothing about a service that fails only when everything runs together.
A service that starts cleanly can still fall over under load through slow response, queue buildup or memory exhaustion, so your measurement, not the datasheet, feeds the 70/75 percent rule. Illustrative sum: ingestion at 18 percent, VPN at 4 percent, the SCADA host at 27 percent and the database at 16 percent total 65 percent combined, a pass; the same roles at 78 percent fail though nothing has crashed yet.
- Reproduce the worst realistic hour: all devices publishing, remote clients connected, historian writing.
- Record peak CPU, peak memory, daily write volume in GB and peak network throughput.
- Pass at 70 percent CPU and 75 percent memory; a miss means dropping a role to another box or upgrading before going further.
For the ingestion side, the guide on bridging OPC UA PLC data to MQTT with an edge PC shows the traffic pattern this test should reproduce.
What CPU and memory does the exact model have? #
Read it from the nameplate and the datasheet of your exact variant, not from the family page: compact industrial PCs commonly ship low-power two- to four-core CPUs and a single memory slot, which means the module you fit at day one is all the memory you will ever have.
Record the full model and order identifier, then pull the matching datasheet for four facts: installed CPU, maximum memory and slot count, storage type, network interfaces. With one slot, buy the end-state module on day one, because there is no second chance. And the measured 65 or 78 percent from your load run is the honest verdict on that low-core processor, not the headline on the box.
- Confirm the planned operating system release is supported on that CPU generation; the article on choosing an operating system image for an industrial PC covers that decision.
- Run the load test from the previous section on this exact unit, not on a bigger sibling.
- If you are specifying hardware rather than qualifying a unit you already own, start from the industrial mini PC collection and pull each candidate's datasheet before shortlisting.
- Pass: the fitted configuration carries the combined stack inside the 70/75 percent rule.
How long will the SSD last under continuous writes? #
Endurance is arithmetic, not a feeling: divide the drive's rated total writes by your measured daily writes. A 480 GB industrial SSD rated 480 TBW carrying 30 GB per day wears out around 44 years out, so the drive is rarely the limit — logging discipline is.
Worked example, with example inputs; take the endurance rating from your drive's actual datasheet. A 480 GB drive whose manufacturer publishes 1,000 full-drive program/erase cycles carries 480 GB × 1,000 = 480,000 GB = 480 TBW. Assume the measured write rate of this historian stack is 30 GB per day: 480,000 GB ÷ 30 GB/day = 16,000 days, and 16,000 ÷ 365 ≈ 44 years, beyond any service window. The same host with debug logging left on writes 500 GB per day: 480,000 GB ÷ 500 GB/day = 960 days ≈ 2.6 years, inside a four-year window. That second number is why log levels, not drive shopping, usually decide this check.
- Take the TBW rating from the drive manufacturer's datasheet.
- Measure GB per day written during the worst realistic week of the load test.
- Pass when rated TBW ÷ measured GB per day ÷ 365 gives at least twice your planned service window; a miss means capping log levels, rotating aggressively or fitting a higher-TBW drive.
- Enable SMART wear monitoring and alert on it; the guide on sizing a self-hosted Node-RED InfluxDB Grafana historian shows how retention settings set both capacity and write rate.
Will the fanless case stay cool enough at the site? #
It depends on a number you must measure: the datasheet states a maximum ambient temperature at rated load, and your cabinet has to stay under it with margin. This checklist passes the site only when the measured worst-case internal temperature sits at least 5 °C below that rated maximum.
A fanless PC dumps processor heat through its enclosure into the surrounding air; run it above the rated ambient and the CPU throttles, and a server that quietly slows through every heatwave afternoon is worse than one that fails loudly. The rating is a documented limit of the exact model; the site temperature is yours to measure: hottest day, doors closed, all services loaded, clearance kept around the fins, never mounted above a drive or a power supply.
- Read the maximum ambient temperature at rated load from the exact model's datasheet.
- Measure worst-case internal temperature at peak load on the hottest day; if go-live comes sooner, measure now and repeat at the first hot spell.
- Pass when the measured temperature is at or below the rated maximum minus 5 °C; a miss means filtered ventilation, moving or shading the cabinet, or moving the role to actively cooled hardware.
- Re-check after any cabinet change; a reworked layout can add ten degrees.
Can IoT, VPN and SCADA share one host safely? #
Yes on one host, provided each role stays in its own lane: consolidate only when losing the whole box is an acceptable outage for every role on it, keep the VPN as the single remote door, and never expose SCADA or database ports to the wider plant network.
Consolidation couples failure modes — a database burst can starve SCADA visibility, one reboot interrupts everything — so the decision is about consequences, not hardware count. Hold one boundary: if the SCADA role is the operator's live window onto a running process, treat the host as critical, give it UPS power and a second viewing path, and never let the VPN be the only way in.
- One inbound path only, the VPN; everything else listens on the plant-side interface or loopback.
- Least-privilege accounts per role, with firewall rules exposing exactly the ports each service needs.
- If the host serves PLC tags to clients, apply the hardening steps for a Node-RED OPC UA server exposing PLC tags before any client connects.
- Test each role alone, then rerun the combined peak-load test with all roles active.
- Pass: every service reachable only through its intended path, response targets still met with everything running.
Will it really run non-stop for years? #
No datasheet promises you years of uptime, so stop looking for one: MTBF is a population mean, not a commitment for your unit — a quoted 50,000 hours works out to 50,000 ÷ 8,760 ≈ 5.7 years on average, and some units fail years earlier. Your real guarantee is recovery.
Answering the two-to-four-year question directly: continuous service for years is a realistic outcome for a correctly qualified small workload, and you earn it through failure design rather than the heatsink. Build so any single failure costs minutes: automatic restart after power loss and after a crash, backups on a different device, and remote access while the site is empty. A short successful run proves nothing about year three; a proven restore proves everything.
- Enable automatic restart after power loss in the firmware's power settings, plus a service watchdog per role.
- Keep backups off the host; a backup that lives only on the drive it protects is a decoration.
- Baseline the healthy host: CPU, memory, temperature and disk usage at peak, so drift is visible later.
- Deploy when recovery is proven and timed, not when a trial run simply has not failed yet.
What must pass before a customer uses it? #
Four tests, all timed and recorded: the combined peak-load run, a controlled reboot with automatic recovery, a network outage, and a full restore to a spare unit. Nothing counts as passed until the restore finishes inside your recovery target with data verified.
Each test exists because a different failure kills small servers: saturation, power events, network flaps, drive loss. Run them in this order on the finished build:
- Combined peak-load run: all roles active at the worst realistic hour, inside the 70 percent CPU and 75 percent memory rule; capture the baseline readings.
- Power test: pull power mid-write, restore it; every service returns automatically in the correct order, with no manual starts.
- Network test: disconnect the uplink for several minutes; the ingestion path buffers through the gap and resumes with no data loss you cannot explain.
- Restore test: restore the latest backup onto a spare unit or blank drive; data, configuration and time sync come back correct; record elapsed minutes against the recovery target.
- Close-out: review logs and alarms after all four tests; clear every unexplained error before go-live.
Go live when all four pass and the numbers are on file. If the restore misses the recovery target, fix the procedure or re-provision the hardware; do not shorten the target to match the result.
Questions asked
How much RAM is a workable floor for IoT, VPN and SCADA on one box?
For a Linux host running a message broker, a small historian and a SCADA server together, this library treats 8 GB as the practical floor and 4 GB as workable only for a single lightweight role; Windows hosts should start at 16 GB. Whatever you fit, size from the measured peak in the load test and keep about a quarter of it free, because memory exhaustion takes services down before CPU saturation does.
Should the server run Linux or Windows?
Either, on qualified hardware. Linux keeps the footprint light on small memory and reboots only when you choose the moment; Windows IoT or LTSC editions fit when a server role requires Windows, because consumer editions force update reboots you do not schedule. Decide from the support matrices of the applications you plan to host: the operating system each vendor actually supports for your version is the one to install.
Can one of these compact PCs run a few virtual machines too?
Where the CPU includes virtualization extensions, technically yes, but on a low-power two- to four-core processor with a single memory module, one host running its roles directly usually serves better than a hypervisor juggling several small VMs. Reserve virtualization for real needs such as tenant isolation or snapshots, and decide from the headroom left after the combined-load test rather than from the appeal of a tidier setup.
What upkeep does a fanless server need over a multi-year run?
Almost none mechanical, because there are no fans or filters on the computer itself; the upkeep is environmental and logical. Keep heatsink fins dust-free at annual cabinet inspections, watch SMART wear and disk headroom alerts, confirm time synchronization stays healthy, and rerun the peak-load test after any operating system or application upgrade so drift from the qualified baseline is caught while it is still small.