Comparison
Thick client or thin client HMI for a control system upgrade
Short answer
A thin-client HMI does not remove processor, memory, disk and network load; it concentrates all of it in one server and one network, so one failure can blank every operator position. Choose thin clients only if the site can staff, back up and recover that server; otherwise refresh thick clients, where a failure costs one station. Virtualizing the Windows XP HMI is a transition, not a modernization.
On this page
- What is the actual difference between a thick client and a thin client HMI?
- Does a thin-client HMI reduce the total computing load?
- How big does the control server need to be for thin client sessions?
- How many operator positions does one failure take out?
- What does the network have to carry for thin clients?
- Do thin clients force the operators to learn new displays?
- Is virtualizing the existing Windows XP HMI a real upgrade?
- Which architecture should I choose for the upgrade?
- Questions asked
The decision is about where load and failure live, not about which label the proposal carries. A thin-client control system HMI does not eliminate the computing work behind displays, alarms, trends and commands; it concentrates that work in one control server and one network, so a single fault there can blank every operator position at once. Refreshed thick clients keep each fault inside one station and keep the network dependency small, at the cost of maintaining a full computer at every desk. This comparison gives the sizing arithmetic, the failure-domain counts, the network planning figures and the operator-training reality for both architectures, then commits to a choice for each situation — including where virtualizing the Windows XP HMI fits: as a transition, not a modernization.
What is the actual difference between a thick client and a thin client HMI? #
A thick client is a full workstation that runs the HMI application locally; a thin client draws its screens from a session running on a control server, so the application work moves to the server and only screen updates and input travel the network.
In the vendor's terms: the thick-client option puts new workstation hardware at each operator position and runs the HMI application locally, as the old machines do. The thin-client option runs the HMI sessions on a control server, and the devices at the desks draw screens and forward keyboard and mouse input over the network.
A virtualized HMI is neither of these. A virtual machine running the old HMI is still a complete operating-system instance — on newer host hardware — while a thin client is a terminal into a session that runs somewhere else. That difference changes licensing, patching, backup and fault containment, so keep the three options separate even when one proposal bundles them.
Does a thin-client HMI reduce the total computing load? #
No — the same display animation, alarm scanning, trend retrieval and controller communications still run somewhere; a thin-client design moves that work to one server and adds the remoting protocol on top.
Every architecture still executes the same work; what changes is placement. Disk writes for trend and alarm history that were spread across six station drives now land on one storage system, which becomes both a shared wear item and a shared failure point.
Heat moves too, on illustrative numbers: six thick stations dissipating 150 W each put 6 × 150 W = 900 W into the control room; a server carrying the same work dissipates a comparable total in one rack. The room gets cooler, the server location gets denser, and its cooling becomes a single point of dependence.
The per-desk hardware does get simpler — that part of the promise is real. What also moves is the skill: keeping the system alive stops being a station swap and becomes server, hypervisor and network administration.
How big does the control server need to be for thin client sessions? #
Size the session host from measured per-session load, not from seat-count rules: with six stations, a measured 0.6 core and 2.5 GB per session, host overhead of 2 cores and 4 GB, and a 70% sustained ceiling, you need 8 cores and 32 GB.
Measure one existing station at the busiest credible plant condition — all displays open, trends and alarms active — using the same disciplined load capture you would use for proving an IPC slowdown is hardware, not software. The inputs below are illustrative; substitute your measurements and rerun the arithmetic.
| Sizing input | Illustrative value | Where it comes from |
|---|---|---|
| Operator positions (sessions) | 6 | Control-room count |
| CPU per session at peak | 0.6 core | Measured on the existing station |
| Memory per session | 2.5 GB | Measured working set |
| Host and remoting overhead | 2 cores, 4 GB | Session host OS and remoting protocol |
| Sustained utilization ceiling | 70% | Design limit for this example |
- Cores: 6 sessions × 0.6 core = 3.6 cores, plus 2 cores of overhead = 5.6 cores; 5.6 cores ÷ 0.70 = 8.0 cores. Order 8 physical cores and do not count hyperthreads as cores.
- Memory: 6 × 2.5 GB = 15 GB, plus 4 GB of overhead = 19 GB; 19 GB ÷ 0.70 = 27.1 GB. Install 32 GB.
- Split option: two hosts with three sessions each need (3 × 0.6 + 2) ÷ 0.70 = 5.4 cores, so 6 cores, and (3 × 2.5 + 4) ÷ 0.70 = 16.4 GB, so 24 GB apiece.
The decision from these inputs: one 8-core, 32 GB server carries all six sessions and exposes all six positions to its failure. Two 6-core, 24 GB hosts cap any single failure at three of six positions — 50% instead of 100% — and that is the configuration to order if you go thin.
How many operator positions does one failure take out? #
With six positions, a failed thick client removes one of six (about 17%); a failed single thin-client server removes six of six (100%), and a network interruption between server and clients can do the same without any hardware failing.
Recovery time follows the same asymmetry, on illustrative targets. A pre-imaged spare thick client swapped onto the desk should have an operator back online inside 60 minutes. A server restored from a 500 GB backup image at 100 MB/s takes 500,000 MB ÷ 100 MB/s = 5,000 s ≈ 83 minutes before service restart and session reconnection, so plan and rehearse about two hours. If two hours of total position loss is unacceptable, split the servers or stay thick.
Staffing decides as much as arithmetic: when a thin-client system faults, someone on shift has to troubleshoot a server, a hypervisor and a network. If that person is a phone call away rather than on site, the fault duration is the vendor's response time, not yours.
What does the network have to carry for thin clients? #
Budget for remoted screen updates on top of controller traffic: a planning figure of 15 Mbit/s per active session with animation-heavy full-screen displays puts six sessions at 90 Mbit/s peak on the server uplink, and latency matters more than bandwidth.
Controller traffic exists in both architectures; the thin-client design adds a remoted screen stream between server and desks. At 15 Mbit/s per active session, six sessions put 6 × 15 Mbit/s = 90 Mbit/s of peak screen traffic on the server uplink. Verify that figure during a pilot using managed-switch port counters before committing cable and switch capacity, and build the path redundant — two switches, two uplinks — because the link is now part of the HMI.
Latency is the number to hold: measure alarm response time before the change, re-measure at acceptance with all sessions active, and fix any regression before cutover, not after.
Do thin clients force the operators to learn new displays? #
In practice yes: the thin-client proposal usually ships with a new display runtime and rebuilt screens, so budget formal operator training and a parallel-running period, while a thick-client refresh can often keep the existing displays and navigation unchanged.
Operators relearn where things live even when the new screens are meant to be simpler, so involve the people who use the busiest displays in acceptance. A thick-client refresh carries existing displays forward only if the new HMI version imports them — verify the import with the real project, not a demonstration.
Either way, start with the discipline of identifying HMI software before replacing an unknown panel PC: list every display, its runtime and version, and every object on it — pushbuttons, setpoint entries, alarm links, trends, permissives — marking each object read-only or state-changing. Objects that change plant state get witnessed tests under approved plant conditions with a rollback agreed beforehand; that is the one place this work can hurt the process.
Is virtualizing the existing Windows XP HMI a real upgrade? #
It is a transition, not a modernization: the virtualized HMI keeps the familiar displays on new host hardware, but the Windows XP guest stays an unsupported operating system that must be isolated, and the host adds a hypervisor to patch and back up.
What virtualization buys is real — the displays operators already know, on new hardware, usually with faster startup. What it does not buy is support: the guest remains an operating system no vendor patches, so treat it as a bridge with a demolition date, not a destination.
Handle the host per ISA/IEC 62443-3-2, which in practical terms asks you to group assets into zones and control every conduit between them: the host carrying the legacy guest sits in its own zone, every network path in is inventoried, and removable media is a controlled conduit rather than an open port. Test host updates against the hypervisor before deployment, keep installation media and license tools recoverable, and set the guest's end date at approval time.
Which architecture should I choose for the upgrade? #
Choose thin clients only when the site can operate, monitor, back up and recover the shared server and network with its own staff; otherwise choose refreshed thick clients, and use virtualization only as a dated interim step for the Windows XP HMI.
The fixed host overhead — 2 cores and 4 GB in the example — spreads over station count. At eight stations it costs 0.25 core and 0.5 GB per seat (2 ÷ 8, 4 ÷ 8); at three stations, 0.67 core and 1.33 GB per seat (2 ÷ 3, 4 ÷ 3). Below about four stations, concentration pays back nothing.
| Refreshed thick clients | Thin clients, control server | Virtualized legacy HMI | |
|---|---|---|---|
| Where the work runs | Each station, locally | One shared server | Guest on new host hardware |
| Worst single failure | 1 of 6 positions (about 17%) | 6 of 6 (100%), or 3 of 6 split | 1 of 6 positions |
| Network dependency | Controller data only | Controller data plus remoted screens | Controller data only |
| Operator training | Minimal if displays carry over | New displays, formal training | None, displays unchanged |
| Administration | Distributed, per station | Server, hypervisor and network skills on site | Hypervisor plus unsupported guest |
| Choose when | Few stations, local fault containment | Many stations, staffed and tested server operation | Interim only, with an end date |
The decision for each case:
- Stay with refreshed thick clients when you have fewer than about four operator positions, no on-site server and network administration, or no completed, timed restore test inside your tolerated outage. Each desk then needs a computer that drives the existing monitors, and a fanless mini PC fits that role.
- Go thin when you have eight or more positions, staff who can operate and recover a server without a vendor phone call, a restore rehearsed inside the tolerated time, and a redundant network path to every desk; order two session hosts, not one. Between four and eight positions, let the staffing and restore-test answers decide, not the hardware count.
- Virtualize only as a dated interim step when the driver is the unsupported Windows XP guest and the displays must not change yet; isolate the host by zone and conduit, and write the removal date into the approval.
Questions asked
Can I run thick and thin clients together during the migration?
Yes, and it is the lower-risk path. Keep the refreshed thick-client stations in service while you commission the control server, then move positions across in pairs. Run both against the same controller data and the same display project where the vendor supports it, and set a firm cutover date so the mixed state does not become permanent.
How many thin-client sessions can one control server host?
With the worked-example inputs of 0.6 core and 2.5 GB per session, 2 cores and 4 GB of host overhead, and a 70 percent sustained ceiling, an 8-core, 32 GB server carries six sessions. Your own per-session measurement decides your number; remeasure at the busiest credible plant condition, because the per-session term dominates the result.
Do thin clients improve control system security?
They move the target, they do not shrink it. The clients expose little to attack, but every session and every operator command now terminates on one session host, which must be patched and hardened to the same standard and schedule as the SCADA server itself. Treat that host as a tier-one asset in your segmentation design.
What should the acceptance test cover before I sign off either architecture?
Write the limits before installation: all sessions concurrent at peak load, sustained CPU at or below the ceiling from the sizing calculation, alarm response no slower than the pre-change baseline, trend retrieval within that baseline, and a timed restore of the server or a spare station with the result recorded. Every display object that can change plant state gets a witnessed test.
What should I demand from the integrator before approving the design?
An architecture drawing that names every component whose failure removes one or more operator positions, the network path and its redundancy, the licensing components per server and per seat, where backups land, and the recovery method with an owner and a time for each failure point. If a vendor cannot fill that in, the design is not ready.