10 KiB
YamBMS - Understanding ESPHome Modbus
Work in progress. The ESPHome Modbus stack is being reworked by @exciton through a long series of pull requests. The first of them landed in ESPHome 2026.7, and more are on the way — including a full overhaul of
modbus_controller(#11781) that will change how commands are queued and how offline devices are handled. This page tracks that work and will be updated progressively as the PRs are merged.
Starting with ESPHome 2026.7, the Modbus client timing defaults have changed
(send_wait_time 250 ms → 2000 ms, turnaround_time 100 ms → 600 ms).
These new defaults are safe for slow third-party hardware, but they are far too
conservative for a YamBMS setup where many fast ESP32-based servers share one bus.
YamBMS therefore ships its own defaults:
substitutions:
# UART
modbus_baud_rate: '19200' # 9600 / 19200 / 115200
# Modbus
modbus_id: '1'
modbus_send_wait_time: '200ms' # Timeout: max time to wait for the FIRST byte of a response
modbus_turnaround_time: '0ms' # Pause after a completed response before the next command
# Modbus Controller
modbus_update_interval: '5s' # Frequency with which BMS/Balancer/Shunt (modbus servers) are queried
modbus_controller:
max_cmd_retries: 1 # Retransmissions after a timeout (default is 4)
offline_skip_updates: 5 # Server re-polled every 30s with an update_interval of 5s
The one idea to remember: what costs time on the bus is the number of Modbus commands per cycle, not the number of servers. A BMS whose register map has gaps generates several commands instead of one. Reducing the command count is usually worth more than raising the baud rate.
What each parameter does
send_wait_time — response timeout
Maximum time the client waits for the first byte of a response. If nothing has started arriving within this time, the command is treated as failed.
- It is a latency budget, not a transmission budget: it does not depend on the baud rate or on the response length. Once the first byte arrives, the client waits for the whole response however long it takes.
- It only costs time when a server does not answer. Lowering it has no effect on a healthy bus — it just bounds the price of a failure.
An ESPHome server on ESP32 answers within one pass of its main loop (16 ms by
default). 200ms leaves ample margin for an occasional slow loop (WiFi burst,
API reconnect, flash write) without ever mistaking a busy server for a dead one.
turnaround_time — inter-command pause
Pause inserted after a completed response, before the next command. It exists to give slow devices time to get ready for the next frame.
It is charged per command, not per server — including between two commands
sent to the same BMS. With 24 BMS generating 4 commands each, a
turnaround_time of 100 ms would add 96 × 100 ms = 9.6 s to every cycle.
0ms is the right value for an all-ESPHome bus. It never produces truly
back-to-back frames: the Modbus specification silence (3.5 characters) is always
enforced, and the ESPHome main loop adds its own scheduling delay on both sides.
update_interval — polling frequency
How often each modbus_controller reads all its register ranges. All
controllers fire at roughly the same time and the resulting commands are
serialized on the bus, one conversation at a time.
max_cmd_retries — retransmissions
Retransmissions of a timed-out command before giving up (not counting the initial
transmission). ESPHome defaults to 4; YamBMS uses 1. One retry absorbs a
transient glitch, and a genuinely dead server is handed to offline_skip_updates
quickly instead of blocking the bus.
offline_skip_updates — offline back-off
Once a server has exhausted its retries it is marked offline, and this many
update cycles are skipped before it is probed again. With 5, a silent server is
re-polled every (5 + 1) × update_interval = 30 s. As soon as it answers a
probe it is marked online and normal polling resumes — no reboot or manual action
needed.
Reducing the number of commands
ESPHome merges consecutive registers into a single range, and reads each range with one Modbus command. A single unmapped register in the middle splits the range in two — and every extra command costs a request frame, a response frame, and a fixed scheduling overhead.
Fill the gaps with register_count
register_count can span a gap. Set it on the sensor before the gap, sized
to reach the next register you actually want. The registers in between are read
and discarded.
Example — a PACE BMS map with gaps at register 8 and registers 13–14:
| Sensor to modify | Value | Effect |
|---|---|---|
| last sensor before the gap (address 7) | register_count: 2 |
absorbs register 8 |
| last sensor before the gap (address 12) | register_count: 3 |
absorbs registers 13–14 |
Ranges 0–7, 9–12 and 15–36 then merge into a single command covering
0–36, taking that BMS from 4 commands per cycle down to 2 — roughly a third off
the total cycle time, for the cost of 3 useless registers read (6 bytes).
Check first that the skipped registers are actually readable on your device. An invalid register anywhere inside a range makes the whole command fail with a Modbus exception.
force_new_range — use sparingly
force_new_range: true (default false, available on every modbus_controller
platform) prevents a sensor from joining the previous range, forcing a separate
command. It works in the opposite direction to the optimisation above, so use
it only to isolate genuinely static registers that can then carry a high
skip_updates.
Two things to know:
skip_updatesapplies to the whole range, and ESPHome keeps the lowest non-zero value among its sensors. One fast sensor inside a large block forces the entire block to the fast rate.- Sensors marked
force_new_rangeare sorted before all others, ahead of the address sort. Marking one sensor in the middle of a block does not split it in two — it pulls that sensor to the front and lets the rest merge. To split a block, mark the first sensor of each intended segment.
It is also set automatically for any sensor using custom_command.
Sensors can share a range for free
Several sensors may read the same registers without any extra bus traffic — they
all decode the same response buffer. A delta cell voltage sensor declared at the
same address as cell voltage 1, with a lambda walking all 16 cells, costs
zero additional commands. Two declarations at the same address produce one
transaction, not two.
YAML order: when it matters
Sensors are sorted internally (by register type, force_new_range, address, then
offset), so their order in the file is normally irrelevant.
Except when two sensors share the same address and offset. The final
tie-break is allocation order — that is, YAML order — and the sensor that comes
first defines the range's register_count.
With cell voltage 1 (1 register) and delta cell voltage
(register_count: 16) both at address 15:
cell voltage 1first → the range opens with count 1, cells 2–16 extend it normally → one command. ✅deltafirst → the range opens with count 16, cell 2 no longer fits the extension test → a second range starts, and registers 16–30 are read twice. ❌
Leave a comment in your YAML so nobody reorders these by accident.
Tuning guide
send_wait_time: set it to the slowest observed "time to first byte" of any server on the bus, with generous margin.200mssuits all-ESPHome/ESP32 buses. Raise it only if a device that is genuinely alive produces spurious timeouts. Do not raise it "to be safe": it is the unit price of every dead-server event.turnaround_time: keep0mson an all-ESPHome bus. For a slow third-party device try10–30msfirst;100ms(the old ESPHome default) is the safe fallback for unknown hardware, and200–600msonly for VFDs and gateways. Remember the cost is per command.- Mixed buses:
turnaround_timebelongs to the hub, so one slow device imposes its value on all your BMS. Put slow third-party hardware on a second UART with its ownmodbus:hub rather than raising the global value. modbus_baud_rate:19200is a good default. Gains flatten out above it, because per-command scheduling overhead starts to dominate transmission time — merge ranges before reaching for a higher baud rate.update_interval: leave enough headroom that the full cycle still fits when a server is dead. If some entities stop refreshing while others keep updating, the cycle is overrunning the interval — reduce the command count, raise the interval, or split the bus.offline_skip_updates:(value + 1) × update_intervalis the re-probe period.5→ every 30 s at a 5 s interval. Lower it to3for faster recovery detection.
Checking your configuration
Set the logger to DEBUG (or VERBOSE for the range details) and watch a few
cycles:
logger:
level: DEBUG
- Ranges are logged as they are built — the quickest way to confirm that a
register_countgap fill actually merged what you expected. Modbus command timed outon a device that is alive meanssend_wait_timeis too low, or the bus is saturated.Modbus device=N set offline/back onlinetraces the offline back-off.