Open the text reading version
1. Start with the arithmetic
Suppose you need to update firmware on 500 battery-powered devices already installed in the field. Half of them sit in meter chambers, pipe voids and metal cabinets.
At typical municipal project rates in China, a single site visit — travel, labour, coordination, downtime confirmation — costs somewhere between CNY 300 and CNY 1,000. For 500 devices that is CNY 150,000 to 500,000, spread over a quarter or more. That is the real price of having no remote update capability.
FUOTA offers a different path: split the image into thousands of fragments and broadcast once to an entire device group, with each device reassembling the file locally. Five hundred devices and five devices take almost the same airtime, because a broadcast does not care how many listeners there are.
But this path is not something you switch on. It is assembled from five independent specifications, each with hard constraints, and the budget has to be calculated per airtime, per duty cycle and per device class. This article takes the mechanism apart specification by specification, and ends with a decision checklist you can take into a design review.
2. Three kinds of "update" are not the same thing
The most common confusion in engineering discussions is treating three different updates as one. Let us fix the boundaries first.
| Target | Channel | Executed by | Typical tooling | Covered here |
|---|---|---|---|---|
| Gateway firmware / ThinkLink itself | Ethernet / Wi-Fi / 4G | Operations, via web or API | gwTool, gateway upgrade packages, system upgrade | No (A/B partitioning was covered separately) |
| End-device firmware | LoRaWAN air interface (multicast or unicast) | Application server orchestration | FUOTA server | Yes |
| End-device firmware (local) | NFC / serial | Field staff with phone or fixture | TapLink and similar | Only as a fallback comparison |
The failure consequences differ completely. A bricked gateway can be opened up and recovered over serial. An end device that fails a FUOTA campaign with no rollback space is usually scrap. That is why, in the third category, redundancy and verification are not optional features — they are the passing grade.
3. Five specifications, each owning one segment
FUOTA is not a single specification. It is a set of LoRa Alliance application-layer specifications plus a process summary (TR002, FUOTA Process Summary) that ties them together. They are building blocks:
| Spec | Name | Owns | Default application port |
|---|---|---|---|
| TS003 | Application Layer Clock Synchronization | Gives devices the network's absolute time | 202 |
| TS004 | Fragmented Data Block Transport | Slices the image, transports it, recovers with redundancy | 201 |
| TS005 | Remote Multicast Setup | Distributes a shared group address and group key, defines the session | 200 |
| TS006 | Firmware Management Protocol | Queries versions, delivers a pending image, reboots | 202 (implementation-dependent) |
| TS007 | Multi-Package Access Protocol | Asks which packages the device actually implements | Implementation-dependent |
The critical point: none of these five is mandatory at the MAC layer. "Supports LoRaWAN 1.0.4" does not mean "supports FUOTA". A device may implement TS004 (it can receive fragments) without TS005 (it cannot join a multicast group), or even without TS003 (it does not know when to open its window). That is precisely why TS007 exists — ask which packages are implemented instead of assuming.
4. Six stages of a real campaign
A real FUOTA campaign has a strict order. Skip a step and the failure surfaces later as "the update failed", while the root cause sits much earlier.
Stage 0: capability self-check
Three hard conditions must be confirmed. Miss one and you need a different approach.
- Device class. Only Class B or Class C can join a multicast session. A Class A device opens receive windows after its own uplinks, so it cannot be synchronised to an agreed moment; it can only receive fragments by unicast.
- Flash space. There must be room for a second image (A/B), or at minimum support for page-by-page writing with progress persisted across resets.
- Bootloader. It must verify the image signature, check hardware model and version, and recover if power is lost mid-write.
The second and third are decided in hardware and production, not in software. Verify them on a sample, not on a datasheet.
Stage 1: clock synchronisation (TS003)
Multicast is fundamentally "agree on an absolute moment and everyone opens the window together". Without a common clock there is no multicast.
Low-cost crystals drift by tens of ppm per day, so a device that has not synchronised for a few days will miss the window. The AppTimeReq / AppTimeAns exchange defined in TS003 lets the device compute its offset from network time and keep correcting it.
This step gets skipped because "the device reported in just this morning". Reporting in is not the same as being time-synchronised. Many devices sync once at join, or do not implement the package at all.
Stage 2: set up the multicast group (TS005, port 200)
The server sends McGroupSetupReq with the group parameters: group ID, multicast address, the encrypted group key McKey_encrypted, and minMcFCount — the minimum downlink frame counter accepted for that group.
The specification's hard limits: a device must support at least one multicast context and at most four (McGroupID 0 to 3). If the server assigns a group ID the device cannot host, the device sets the IDerror bit in McGroupSetupAns and refuses explicitly. That is a pragmatic design — it turns "unsupported" from a silent failure into an observable error.
minMcFCount protects against replay: devices must discard multicast frames with a frame counter below that value. Its purpose is adding a device to a group that is already running without letting it replay historic frames.
How the group key arrives is the most overlooked version difference in this stage:
- LoRaWAN 1.0.x requires an additional root key, GenAppKey, provisioned in the device. It derives McKEKey, which decrypts McKey_encrypted into McKey, from which McAppSKey and McNwkSKey are derived.
- LoRaWAN 1.1 and later derives McRootKey from AppKey. No second root key is needed.
The conclusion is blunt: on 1.0.x devices, whether multicast is even possible is decided at production provisioning time. If GenAppKey was never provisioned, no software update will fix that.
Stage 3: set up the fragmentation session (TS004, port 201)
FragSessionSetupReq defines the transfer: session index, total fragment count, fragment size (a multiple of 8), the fragmentation algorithm, redundancy ratio and session expiry.
A single session carries at most 16,383 fragments. That ceiling is fine for most firmware, but it carries a warning: fragment count is not unbounded, and small fragments on a large image push you toward it.
The redundancy ratio is the heart of TS004. The server transmits a proportion of extra fragments so a device can rebuild after collecting any sufficient subset, without requesting retransmissions. This converts reliable delivery from a retransmission problem into a coding problem — at the cost of proportional extra airtime.
Stage 4: multicast transmission (port 202)
A DataFragment frame is minimal: a 3-byte header (session index plus fragment counter) followed by raw image bytes. So the 51-byte application payload at DR0 actually carries 48 bytes of image.
Class C devices keep their receive window open throughout; Class B devices use ping slots. The server does not wait for acknowledgements — it broadcasts to the end.
Speed and risk come from the same property: a broadcast does not care who received it. "Transmitted" and "installed" are two different things, separated by stage 5.
Stage 5: completion and verification
The device reports completeness with AllFragReceivedReq, then runs the local sequence: reassemble, verify the signature against the update server's public key, check that the image header matches its hardware model and current version, mark the image pending, reboot, and let the bootloader install it.
TR002 makes an important point here: implementing fragment transport does not imply signature verification and power-loss recovery. Those belong to device and bootloader implementation quality; the specification recommends but does not mandate them. Your acceptance checklist therefore needs its own "pull the power mid-write" test, rather than trusting the phrase "supports FUOTA".
Stage 6: acceptance
TS006 queries versions. In practice, check three things: the version number actually changed, sensor readings are sane (a new image that boots but loses its driver is still a failure), and the next uplink arrives on the new interval.
5. The budget: three ledgers
The part of a FUOTA project that goes wrong is rarely the protocol; it is the time estimate. Three variables set total duration: image size, data rate and redundancy. Here is the arithmetic with real EU868 parameters.
Ledger one: airtime
Take a 100 kB image (counting 100,000 bytes), DR0 (SF12BW125), 48 bytes of image per frame:
| Item | Value |
|---|---|
| Image fragments | about 2,084 |
| Total frames with 10% redundancy | about 2,300 |
| Airtime per frame at SF12 | about 2.8 s |
| Pure airtime | about 6,400 s, i.e. 1.8 hours |
1.8 hours of pure airtime looks tolerable. But that is only how much of the air the transmission physically occupies. Regulation has not been accounted for yet.
Ledger two: duty cycle
The 869.4–869.65 MHz sub-band in EU868 is limited to a 10% duty cycle under ERC Recommendation 70-03. That gives a gateway 360 seconds of transmit time per hour — roughly 129 frames per hour at 2.8 s each.
2,300 frames ÷ 129 frames per hour ≈ 18 hours.
So the honest answer is not 1.8 hours but close to 18 — and that is the optimistic case, ignoring retransmissions and ignoring every other downlink the gateway must send on that channel. This is why so many campaigns "run all night and are still not done".
Ledger three: data rate is the biggest lever
The same 100 kB image at DR5 (SF7BW125) carries 242 bytes per frame, or 239 bytes of image after the 3-byte header: about 460 frames at 0.39 s each — 3 minutes of pure airtime, roughly 30 minutes after the duty cycle limit.
From DR0 to DR5 the difference is about 36×. The price: every device in the group must receive reliably at SF7. And the worst device in the group is usually the reason the gateway was installed in the first place — underground, inside a metal cabinet, far from the antenna.
This is the real trade-off in FUOTA planning: the data rate is set by the worst link, not the average. Three ways out: run a coverage check before the campaign, temporarily relocate the marginal devices, or split the fleet into batches at different rates.
A fourth comparison: field versus airtime
It is worth putting the three paths side by side for a 500-device fleet:
- Per-visit field updates: CNY 150,000 to 500,000 at the estimates above, a quarter of scheduling, and a downtime negotiation every single time.
- Batch unicast: the server pushes to each device in turn, and each device can only take one fragment per uplink window. At a 15-minute reporting interval and 2,300 frames, a single device needs more than three weeks — while consuming downlink capacity.
- Multicast FUOTA: one broadcast, 30 minutes to 18 hours on the arithmetic above, essentially independent of fleet size.
The cost structures differ completely: the first buys labour, the second buys time, the third buys upfront engineering and device-side capability.
6. The security price of multicast
Multicast saves airtime but changes one thing in the security model: every device in the group shares the same session keys (McAppSKey and McNwkSKey).
That means if one device is opened up and its keys extracted, an attacker can forge valid multicast downlinks and broadcast arbitrary content to the entire group. In unicast, a compromised device affects only itself. In multicast, one compromised device is the whole group.
The specification's answer is to keep keys inside hardware: after McKey_encrypted is decrypted with McKEKey, the ideal implementation never lets McKey, McAppSKey or McNwkSKey leave the secure element. Then even physical possession yields no keys.
Three practical conclusions:
- Multicast members should be models whose key protection you can actually attest, not "everything that fits in the group".
- On 1.0.x, GenAppKey provisioning belongs to production and supply chain, not to operations.
- The larger the group, the wider the blast radius — so split groups by business domain, not by convenience.
7. Six ways campaigns fail in the field
Ordered by how often they show up in real projects:
- Class A devices pushed into a multicast group. They never receive anything, and the log only says "some devices did not respond". Check the class per device, not per model name.
- Group created before clock sync. The symptom is intermittent misses and partial fragments; the cause is device time drifting out of the window. Run TS003 first.
- Multicast timeout set too short. The session ends at 80% and the remaining fragments are never delivered. Derive the timeout from fragment size, image size and data rate. Third-party platforms usually calculate it; default values tend to sit in the 4,096-second range.
- Data rate chosen from average signal quality. The marginal devices fall behind, redundancy is consumed absorbing their losses, and effectively you have none left.
- Redundancy set at the theoretical value. Ten percent redundancy covers ten percent loss, while instantaneous loss behind metal can be far higher. Twenty-five percent is a more realistic engineering value, at proportional airtime cost.
- TS004 without signature verification or power-loss protection. This is the only category that causes irreversible loss: an interrupted write bricks the device, and no field visit can recover it.
One more item that is not an error but is frequently left out of the budget: energy. A Class C device keeps its receiver open for the whole campaign — 30 to 60 minutes of continuous receive. For a battery device, that is the most concentrated energy draw in a multi-year lifecycle. Budget it explicitly using the battery-life method covered in our earlier article, and schedule the campaign when batteries are fresh, for example in the first month after a replacement cycle.
8. Platform reality: batch push is not multicast
This section draws a boundary that sales language tends to blur.
On the ThinkLink side there are four update-related capabilities today: a device firmware library (import an obin built with EB compiler, or save it directly from the cloud compiler), per-device firmware upgrade, batch one-shot RPC dispatch, and local NFC upgrade through the TapLink mobile app. On the gateway side there are two separate channels: gateway upgrade packages and system upgrade.
Batch one-shot RPC is not FUOTA multicast, and the difference is worth spelling out:
- Batch one-shot RPC is unicast, device by device, with a configurable inter-device send interval in milliseconds to avoid instantaneous concurrency spikes. Five hundred devices means five hundred unicasts, and duration grows linearly with fleet size.
- FUOTA multicast is one broadcast, the whole group receiving at once, essentially independent of fleet size and dependent only on image size, data rate and redundancy.
Put differently: batch unicast solves "I do not want to click 500 times". Multicast solves "I cannot wait for 500 uplink windows". The first is an efficiency problem; the second is a physics problem.
For ThinkLink specifically, no multicast FUOTA capability under TS004 / TS005 is publicly claimed; the platform wording here stays conservative and official confirmation governs. For private-network users, the practical advice scales with fleet size:
- Dozens of devices across a few sites: batch unicast plus scheduled windows is entirely sufficient; multicast would not repay the engineering and device-side work.
- Hundreds of devices spread across a region: worth evaluating multicast — and the first question is whether the device implements TS003 / TS004 / TS005, not whether the platform "supports FUOTA".
- Ten devices or fewer, or mechanical maintenance needed anyway: NFC upgrade in the field is simpler, and you get a fastener and seal inspection for free.
9. Decision checklist
Multicast FUOTA is worth the investment when (three or more should hold):
- 100+ devices installed, physically dispersed, with high per-visit cost
- Devices are Class B / Class C and confirmed to implement TS003, TS004 and TS005
- Image size above 30 kB and more than two updates per year expected
- A defined maintenance window exists, and hours of continuous receive are acceptable
- A rollback path exists: A/B partitions, or field recovery over NFC / serial
Batch unicast is enough when:
- Dozens of devices concentrated on a few sites
- Low update frequency (once a year or less) and small images
- Devices are Class A and cannot be changed
Go to the field when:
- Ten devices or fewer
- No multicast capability and no A/B partition
- The visit also needs mechanical maintenance, seal inspection or a battery swap
Five questions to ask in review:
- Which packages does the device implement? Demand TS numbers, not "supports FUOTA".
- How large is the image, what data rate is targeted, and can the worst link in the group sustain it?
- What redundancy and what multicast timeout, and how were both derived?
- At which layer are signature verification and power-loss recovery implemented, and has a power-cut test been run?
- What is the rollback path after a failed campaign, and does it require a site visit?
10. Closing
FUOTA is not a button. It is an operations project that needs a budget, a time window, a rollback plan and acceptance criteria. Its technical core is not protocol complexity but three pieces of arithmetic: how many fragments the image becomes, how much airtime each fragment costs, and how long the duty cycle lets you spread them.
Get those three right and updating 500 devices is one overnight broadcast. Get them wrong and it is 500 site visits.
References: LoRa Alliance TS003 Application Layer Clock Synchronization; TS004 Fragmented Data Block Transport; TS005 Remote Multicast Setup Specification v1.0.0; TS006 Firmware Management Protocol; TS007 Multi-Package Access Protocol; TR002 FUOTA Process Summary; ERC Recommendation 70-03 (10% duty cycle for 869.4–869.65 MHz); LoRaWAN Regional Parameters RP002 payload limits; public FUOTA implementation documentation from third-party platforms (ChirpStack, AWS IoT Core for LoRaWAN) and reported field figures. Airtime and duty-cycle figures assume EU868 without repeaters; real projects must recalculate with their regional parameters and gateway downlink load. Field visit costs are rough domestic estimates used only for order-of-magnitude comparison.
Buildings, HVAC & campuses · Global partners · Send an enquiry ↗
Explore more technical notes and field practices.
Back to articles ↗