Skip to content
Emberwire OTA firmware delivery

Firmware delivery infrastructure

Ship firmware to a fleet you cannot reboot by hand.

Emberwire stores your firmware images, signs them, and hands them to devices in controlled waves — smallest cohort first, full fleet last. Every image is verified on the device before it is written. When an update goes wrong, the wave stops and the device goes back to the build it was running an hour ago.

9.6 Mdevices enrolled across customer fleets
712 TBimage data delivered in the last 30 days
14delivery nodes in 12 countries
99.7 %updates confirmed on first attempt

The shape of the system

One origin holds the signed images. Fourteen nodes hold the bytes close to the devices. The devices pull — nothing is pushed at them.

ORIGIN DELIVERY NODES DEVICE FLEETS Image store signed · deduplicated 214 TB · 38,900 builds Frankfurt · fra-a 104 TB / 30 d Ashburn · iad-a 77 TB / 30 d Singapore · sin-a 41 TB / 30 d + 11 more nodes · 800 TB of edge cache in total Smart meters — 4.90 M devices 3.2 MB image · 180 KB delta · check-in 15 min basements, meter cabinets, mobile links Routers — 1.26 M devices 24 MB image · 2.6 MB delta · check-in 5 min home and small-office broadband Payment terminals — 38,000 devices 1.4 GB image · 96 MB delta · session held open one device in canary
Solid lines: a device pulling from its nearest node. Dashed lines: the same device falling back to a second node when the first one is draining. The orange square is the single terminal a wave starts with.

What the service actually does

Stores and signs images

You upload a build; Emberwire computes its digest, stores it deduplicated against previous builds of the same product line, and signs a manifest with your release key. Images are kept for 24 months so that a device that has been in a warehouse for two years still has a path forward.

Rolls out in waves

A release does not reach the fleet at once. It reaches one device, then a hundred, then a percentage, with a soak window between each step. If confirmations stop arriving, the next step never starts. Default ramp: 0.1 % → 1 % → 5 % → 25 % → 100 %.

Sends differences, not images

Where the device can apply one, Emberwire ships a binary delta against the exact build the device is running. Median delta is 6.7 % of the full image. For the five device classes below that is 7.9 TB per full-fleet wave instead of 102 TB.

Verifies on the device

The agent checks the signature chain and the image digest before a single byte is committed to the inactive slot. A manifest that does not chain back to the trust anchor burned into the device is discarded without being written. See Security.

Rolls back without a truck

A device that boots a new image and does not confirm within the watchdog window reverts to the previous slot on its own. No command has to reach it, which matters because the failure mode that needs a rollback is usually the one that broke the network stack.

Keeps the receipts

Every transition — offered, downloaded, verified, staged, booted, confirmed, rolled back — is recorded per device with a timestamp and the build identifier. Exportable as signed JSONL for audits and regulatory filings.

What a wave costs in bytes

Representative device classes across the customer base. These five cover 87 % of enrolled devices. Volumes are for one full-fleet wave, not per month.

Image size and wave volume by device class — measured, 30-day window
Device class Full image Typical delta Delta ratio Devices Wave, full Wave, delta
Environmental sensor 812 KB41 KB5.0 % 2,140,0001.62 TB84 GB
Smart meter 3.2 MB180 KB5.5 % 4,900,00014.9 TB841 GB
Home / SMB router 24 MB2.6 MB10.8 % 1,260,00028.8 TB3.12 TB
CNC machine controller 512 MB41 MB8.0 % 9,4004.59 TB376 GB
Payment terminal 1.4 GB96 MB6.7 % 38,00052.0 TB3.48 TB
Total——6.7 % 8,347,400102 TB7.9 TB

Delta ratio in the total row is the median across classes, not the ratio of the totals. A first-time device, a device that skipped several releases, or a device whose flash content does not match the expected base always receives the full image.

Why the traffic looks the way it does

Devices stay connected because reconnecting is expensive

A fleet is not a set of browsers. A water meter in a basement checks in every fifteen minutes for ten years and asks for nothing else. A payment terminal in a corner shop keeps one session open overnight because re-establishing it across a congested mobile link costs more battery, more time and more failed retries than keeping it alive.

The result is a traffic profile that looks unusual if you expect web traffic: a very large number of long-lived, mostly idle sessions that originate from consumer and mobile networks — home broadband, business DSL, cellular carriers — because that is physically where the devices are. Then, when a release reaches their cohort, those same sessions pull somewhere between 812 KB and 1.4 GB in a burst.

Aggregate: 712 TB delivered in the last thirty days, with roughly 78 % of connections idle at any given moment. See the delivery network for the per-node breakdown.

Nodes sit in twelve countries for one reason

Nobody wants to pay to move gigabytes across an ocean more than once. A 1.4 GB terminal image replicated once to the São Paulo node and then served locally to 4,000 terminals moves 1.4 GB across the Atlantic instead of 5.5 TB.

The second reason is that devices on poor links fail differently at distance. A meter on a marginal cellular connection will finish a 3.2 MB transfer from a node two hops away and give up on the same transfer from a node on another continent, after burning the data allowance three times over in retries.

Nodes are placed where customer fleets are, not where it looks impressive on a map. If a customer has no devices in a region, they pay nothing for it.

What happens when it goes wrong

The interesting part of an OTA system is the failure path

Anybody can copy a file to a device. The question that decides whether you ship on a Friday is what the fleet does when the file is bad, when the power dies mid-write, or when the new image takes the radio down.

Bad image reaches a cohort

Devices that boot the new build and fail to confirm within the watchdog window roll back to the previous slot without being told to. The wave halts automatically when the confirm-failure rate in the current cohort exceeds 0.5 % or when 25 consecutive devices fail to confirm — whichever comes first. In the default ramp that caps exposure at the 1 % cohort.

Power fails mid-write

Nothing is written to the active slot, ever. The download lands in the inactive slot or on the staging partition; the slot pointer flips only after the digest matches. A device that loses power at any point during the transfer boots the build it was already running and resumes the download from the byte it reached.

A node goes down

The agent holds a list of candidate nodes from its last successful check-in and retries the next one with a resumable range request. Transfers survive a node loss; a device does not restart a 1.4 GB download because a node was drained for maintenance.

A device disappears for a year

Devices that reappear after a long absence are placed in a quarantine cohort rather than being handed the newest build directly. Emberwire walks them through any release marked as a required waypoint, so a device on a two-year-old build is not asked to make a jump nobody ever tested.

Who runs on Emberwire

Customers are manufacturers. We do not name them, and we do not publish logos or testimonials — a device vendor's update cadence is competitive information.

Customer profile Fleet What they run into
A European metering vendor 480,000 Regulated market: every firmware change needs an exportable, per-device audit trail with signature evidence.
A North American POS manufacturer 38,000 1.4 GB images, devices that must not be offline during trading hours, strict maintenance windows per merchant.
An industrial controls group 9,400 Machines that cannot reboot on request. Updates are staged for weeks and applied at the next planned shutdown.
A consumer networking brand 1,260,000 Volume: a routine release is 3.1 TB of delta traffic in a weekend, mostly to residential connections.
An environmental sensing startup 2,140,000 Battery-powered devices on metered cellular plans. Every kilobyte is a line in their operating cost.

Start here

If you are evaluating

Read How it works for the rollout model, then Device agent for the memory and flash budget on your target.

If you are integrating

The API documentation covers upload, manifests, targeting, wave control and webhooks. The agent is a single static binary or a source drop for RTOS targets.

If you are buying

Pricing is per enrolled device per month plus egress. Two worked examples on that page show where the money actually goes.