Google Coral Dev Board

schematics

published (baseboard schematic + Altium source + Allegro layout, Apache-2.0

toolchain

partial (open runtime (libedgetpu) + TFLite, closed Edge TPU compiler)

datasheets

public

blobs

required (proprietary Edge TPU ASIC + closed compiler)

retail

open_market (broadly purchasable)

Google publishes a public datasheet and the baseboard design files, and there is an open runtime with a TensorFlow Lite path, but the Edge TPU silicon and the model compiler are proprietary closed binaries. The datasheet page links design files throughout: a baseboard schematic PDF, the Altium schematic source, an Allegro .brd layout and a STEP model of baseboard and SoM together, all served from github.com/google-coral/electricals, which carries an Apache-2.0 LICENSE. That is more than radxa-rock-5b releases, and it stops short of fully open because only half the product is covered - the SoM that carries the i.MX 8M and the Edge TPU has no design files. Published board files reach the next tier up only alongside a stack that builds from source, and the Edge TPU Compiler is a closed binary, so the board lands at 3, documented, either way; what the published files change is that the 3 is derived from what Google actually publishes. The public documentation is that same v1.7 datasheet. The toolchain is partly open: libedgetpu ships its runtime source under Apache-2.0 against an Edge TPU Compiler distributed only as an apt binary from packages.cloud.google.com with no source anywhere, and a repository being archived does not make it closed. Firmware blobs are required because that same closed compiler stands between any model and the proprietary Edge TPU. Retail is the awkward component: the board reads as withdrawn rather than on open sale. Seeed, its named reseller, marks both the 1GB at $129.99 and the 4GB at $169.99 discontinued and out of stock; coral.ai/products/dev-board now redirects to developers.google.com/coral, a page that names no hardware and links no purchase route; and no reachable distributor confirms stock, with Mouser serving a bot-denial page and Farnell and Digi-Key refusing automated requests. Open market, distributor-only and restricted all describe a part somebody can still buy under some terms, so a withdrawn part is recorded instead as discontinued. The wording is score-neutral either way, since no rung on the hardware ladder reads the retail component.

• https://coral.ai/products/dev-board recorded 2026-08-14

  the product page for this board 302-redirects to developers.google.com/coral, which names no hardware and links no purchase route - establishes that the vendor no longer offers the board for sale

• https://coral.ai/docs/dev-board/datasheet/ recorded 2026-08-14

  Datasheet v1.7 (December 2022) - 4 TOPS, 2 TOPS/W, SoM form factor, TFLite, Google-designed Edge TPU; its "Schematic and layout files" section links Coral-Dev-Board-baseboard-schematic.pdf, Coral-Dev-Board-baseboard-schematic-Altium.zip, Coral-Dev-Board-baseboard-layout-Allegro.brd and Dev Board.STEP, and points anyone wanting the SoM on custom hardware at the SoM datasheet rather than at design files

• https://github.com/google-coral/electricals recorded 2026-08-14

  "Electrical designs for coral.ai projects", Apache-2.0, public; dev_board/ holds Coral-Dev-Board-baseboard-schematic.pdf, -schematic-Altium.zip and -layout-Allegro.brd, which is where the datasheet download links resolve to

• https://coral.ai/docs/edgetpu/compiler/ recorded 2026-08-14

  The Edge TPU Compiler is distributed as an apt package from packages.cloud.google.com and nothing else - installation is an apt-key add plus apt-get install, the page publishes a version history to 16.0 and points at no source repository, no build instructions and no license. It "compiles a TensorFlow Lite model (.tflite file)" and nothing else. The closed half of the toolchain, and the blob that stands between any model and the Edge TPU

• https://github.com/google-coral/libedgetpu recorded 2026-08-14

  "The userspace level runtime driver for Coral devices", Apache-2.0, source published - the open half of the toolchain. The repo carries "This repository was archived by the owner on Oct 14, 2025. It is now read-only", last pushed 2024-08-11. The source is still published under the same licence, so the openness value is unaffected; what the archive bears on is adoption

• https://www.seeedstudio.com/Coral-4GB-Dev-Board-Version-p-4683.html recorded 2026-08-14

  Coral Dev Board 4GB, SKU 102110458, $169.99, marked Discontinued and Out of stock at the board's named reseller. The 1GB listing at Coral-Dev-Board-p-2900 reads the same way, $129.99, Discontinued and out of stock. Read for `retail` and it does not support `open_market`, which is why the openness axis above is left undated

The entire Edge TPU software stack is archived. libedgetpu is read-only, archived on 2025-10-14, and the Coral GitHub org shows 34 of its 40 repositories archived - libedgetpu, libcoral, pycoral, tflite, examples-camera and tutorials among them, all last pushed 2024-08-11. The six live repositories are a different product line: coralnpu, coralnpu-compiler, coralnpu-mpact, coralmicro and two project demos. Combined with retail, this is a wound-down platform rather than a niche one - Seeed marks both variants discontinued and out of stock, and coral.ai/products/dev-board now redirects to a page that names no hardware and links no purchase route. The category already places nvidia-jetson-agx-orin and memryx-mx3 at level 3, niche, on stronger evidence than this board can now show. It is held at 3 rather than lower because the hardware remains in use and the archived stack still functions.

• https://api.github.com/orgs/google-coral/repos?per_page=100 recorded 2026-08-14

  34 of 40 repositories archived, including the whole Edge TPU stack (libedgetpu archived 2025-10-14, libcoral, pycoral, tflite, examples-camera, tutorials), all last pushed 2024-08-11; the six live repos are coralnpu/coralnpu-compiler/coralnpu-mpact/coralmicro plus two demos, a different product line

Modest 4 TOPS limited to int8 TFLite models; suited to small CNN inference rather than large or generative models. The datasheet (version 1.7) reads "4 trillion operations per second (TOPS)", "2 TOPS per watt" and an Edge TPU that "performs 4 trillion operations per second" with TensorFlow Lite models. That is the lowest AI throughput in this category apart from sipeed-maixcam, and the TFLite-only path is what keeps it below the 6 TOPS RK3588 boards.

• https://coral.ai/docs/dev-board/datasheet/ recorded 2026-08-13

  "High-speed and low-power ML inferencing (4 TOPS @ 2 W)", "4 trillion operations per second (TOPS)", "2 TOPS per watt", and an Edge TPU coprocessor used "with TensorFlow Lite models"