Performance

How fast is a PLC that costs a few euros?

No promises: timings measured on real boards, and from there the maths for two concrete plants, a hotel and an irrigated field. Where a number is an estimate, it says so.

measured0.79 µsone ladder network “contact → coil” on ESP32

measured0.92 ms1000 ladder networks, one full scan

measured5.4×faster on the P4 with machine code, on a large program

measured0.3–0.5 msround trip of a µLink I/O module

The virtual machine

How long it takes, function by function

The ladder program becomes bytecode, which the VM runs every scan. The P4 is a more powerful processor and can also turn the simple parts into machine code, which the CPU runs directly. Lower times = faster.

Function	ESP32-WROOM	ESP32-P4
One contact → coil network	0.79 µs	≈ 0.4–0.5 µs
One network with an addition (ADD)	1.4 µs	≈ 0.7–0.9 µs
Simple machine, ~300 instructions	0.25 ms	≈ 0.15 ms
1000 contact → coil networks	0.92 ms	≈ 0.5 ms
Program with 101 functions (14.8 KB)	1.44 ms WiFi off only	≈ 0.7–1 ms
Large plant: 129 blocks, 4.3 MB	doesn't fit	38.9 ms with machine code 209.7 ms with the VM only

Values without “≈”: measured on the board. With “≈”: estimated, the P4 1.5 to 2.2 times the ESP32 with the same VM. Machine code pays off on large programs; on a small program full of loops and jumps it's slower than the VM (168 µs vs 136 µs, measured): that's why on the P4 it's a choice, not a must.

Program size

How much program fits

All memory and space values on this page are those of the PLC board: the program, DBs and data live there. I/O modules don't hold the program and the PC is only used to write it. Work memory is the board's RAM where the bytecode runs; on the ESP32 it shares it with the network, and WiFi alone takes a lot.

ESP32 PLC board

work memory, WiFi + Ethernet11.4 KB
work memory, Ethernet only~50 KB
DB area8 KB
load memory (flash)256 KB

In practice: 11 KB is about 1000 simple ladder networks, already a real plant. 50 KB holds a 100-function program.

ESP32-P4 PLC board

work memory (with PSRAM)~7.6 MB
program code, at most8 MB
DB area2 MB

How big is 4 MB? A simple network takes about ten bytes; one from a real plant, with comparisons and boxes, 30–60 bytes. An automatic machine has a few hundred networks; a complete production line a few thousand. The test's 4.3 MB correspond to tens of thousands of networks: the size of a whole department, not of one machine. For a hand-written program the P4's memory never runs out: the limit becomes inputs and outputs.

Communication

One PLC and three I/O modules

A real test: an ESP32-P4 PLC and three ESP32 modules on µLink, with the editor Online and a Download running to stress it. Every round the PLC sends each module the state of all its outputs and gets back all its inputs.

I/O module 1ESP32

309 µsround trip

I/O module 2ESP32

495 µsround trip

I/O module 3ESP32

359 µsround trip

PLCESP32-P4

Average round trip measured for each module, exchange every 5 ms, with the editor Online and a Download running.

measured0.31–0.50 msaverage round trip, per module

measured200 per secondexchanges per module: 64 outputs and 64 inputs each time

measured22,608exchanges per module during the test, all valid

measured< 0.5 %replies arriving late, under stress

A lost frame does no harm: every frame carries the full state, and the next round puts everything right. The worst point measured, during a heavy Download, was 12–15 ms.

In practice · 1

The hotel: how many rooms?

Each room: 10 buttons, 10 lights, climate with temperature probe, valve and fan. A “Room” function written once and called for every room; one 64-I/O µLink module in the floor cabinet serves 5 rooms.

Rooms: 40 WiFi on the PLC

ESP32

ESP32-P4

A blink lasts 100–400 ms: under 100 ms from button to light no guest notices anything. The P4 in its extended configuration (in development) reaches 1000 I/O boards, 128 kbit of I/O and a 128k-value REAL area: that's the top of the slider. Keeping lights under 100 ms, about 1300 rooms fit; beyond that, split over more buses or PLCs.

In practice · 2

Irrigation: how many hectares before RAM runs out?

Each hectare: 2 sectors with valve, watering time and soil-moisture threshold, plus one probe. Plus a fixed station: pumps, pressure, dosers, sludge treatment, mixers, chemical probes.

Hectares: 50 WiFi on the PLC

ESP32

ESP32-P4

A field is in no hurry: even a half-second scan is fine. Here the real limits are memory and I/O points, not speed. On the P4 probes are not a problem: the 128k-value REAL area is enough for over 100,000 probes, so the ceiling stays the DB area. Analog probes need modules with analog inputs or Modbus sensors.

In practice · 3

Baggage conveyors

Each conveyor: 4 photocells, motor, bag tracking (where every item is), jams and diverters. Here speed matters: at 2 m/s a bag travels 4 cm in 20 ms, so we want every conveyor refreshed at least every 20 ms. One 64-I/O µLink module serves 15 conveyors.

Conveyors: 200 P4 with machine code

ESP32

ESP32-P4

A large airport doesn't use one PLC: the system is split into zones, each with its own PLC. An ESP32 handles a small zone, a P4 a zone of hundreds of conveyors, over 2000 with machine code. Estimates: a real baggage system must be engineered and tested on site.

Summary

One row per board

Board	Hotel today (8 modules)	Hotel at the limit	Irrigation: hectares before RAM runs out	Baggage conveyors, refreshed every 20 ms

“At the limit”: ESP32 with 15 modules (all the program's I/O); P4 in its extended configuration, in development: 1000 I/O boards, 128 kbit of I/O, 128k REAL. The time is how often all rooms are refreshed.

🤓 For the curious: raw numbers and how we did the maths

Measurements (real boards, firmware 0.3.39–0.3.71, November 2026)

What	Value	Where
LD → ST network (3 VM instructions)	788 ns	ESP32-WROOM 240 MHz
Network with ADD DINT on DB	1,364–1,460 ns	ESP32
Empty scan (I/O, system DB, checks)	~0.13 ms	ESP32
carico-100fc: Main + 101 FCs, 14.8 KB	1.44 ms	ESP32
4.3 MB of machine code, 129 blocks: VM / mixed	209.7 / 38.9 ms	ESP32-P4, 60 comparisons with the interpreter identical
Small program with loops: VM / machine code	136 / 168 µs	ESP32-P4
Free work memory, WiFi + Ethernet	11,702 bytes	ESP32
Free work memory	7,989,150 bytes	ESP32-P4 with PSRAM
µLink, average round trip (3 ESP32 modules)	309–495 µs	P4 PLC, 5 ms period
µLink, late replies under stress	0.06–0.42 %	22,608 exchanges per module
µLink, ESP32-C3 + W5100 (SPI) module	4.7–7.3 ms	the SPI chip is the bottleneck
µLink, 8 modules	~1.3 ms of bus every 5 ms	µLink documentation
µLink, 192 modules (test)	16 ms to send everything	ESP32-P4
Download of 127 KB over USB (serial cable, including STOP and flash write)	~28 s	ESP32-P4, MiaConfig automatic measurement
Editor over Ethernet / USB	0.7–1.3 s / 33.5 s	370 KB, 18 files
Status request over Ethernet, PLC in RUN	31–40 ms	ESP32 and P4

Estimates (to be confirmed)

P4 with the interpreter VM: at least 1.5× the ESP32 (RISC-V CPU at 360–400 MHz; the estimate ranges 1.5–2.2×). We use 1.5× in the maths.
FC call: ~0.3 µs, plus ~0.25 µs per parameter.
Room: 23 bit networks, 3 boxes (climate scaling and comparisons), 26 parameters → ~29 µs on ESP32, ~19 µs on P4; 107 bytes of call + 24 bytes of DB; the “Room” FC 600 bytes, once. 12 outputs per room: 5 rooms per module with 63 usable outputs.
Hectare: 2 sectors × (5 networks, 4 boxes including the timer, 12 parameters) → ~26 µs on ESP32; 102 bytes of calls, 56 bytes of DB (two 16-byte TIMERs, times, thresholds, probe). Fixed station: 4 KB of code, 1 KB of DB, 0.4 ms.
Conveyor: 20 bit networks, 10 boxes (tracking, timers, comparisons), 15 parameters → ~34 µs on ESP32, ~23 µs on P4 with the VM; with the P4's machine code 5.4× faster (the value measured on the large program). 63 bytes of call, 64 bytes of DB (bag queue, timers); 4 inputs per conveyor: 15 conveyors per module. Target: everything refreshed within 20 ms.
Bus: period = at least 5 ms, with the bus busy at most 60 %; on the P4 0.083 ms per module (192 modules in 16 ms). Worst-case reaction = 2 periods + 1 scan.
P4 extended configuration (in development): 1000 I/O boards, 128-kbit I/O areas, a REAL area of 131,072 values (128k). One REAL probe per room and per hectare.

The formulas

scan    = 0.13 ms + N × (networks × 0.79 + boxes × 1.4 + 0.3 + params × 0.25) µs ÷ speed
memory  = fixed code + N × call bytes          ≤ work memory
data    = fixed data + N × DB bytes            ≤ DB area
period  = max(5 ms, modules × time per module ÷ 0.6)

The calculator is in the site file assets/js/prestazioni.js: all starting numbers are at the top, with their source.