The original Google Coral Edge TPU products (USB Accelerator, M.2 and Mini PCIe cards, Dev Board) have not been given a public end-of-life notice, and some distributors still stock them in October 2026. What changed is Google’s focus: coral.ai now redirects to Coral on Google for Developers, which is about Coral NPU, an open-source chip design that other companies build into their own processors. The first hardware with it is the Synaptics Coralboard, sold through a DigiKey wait list. Facts below were checked on Google’s and Synaptics’ pages on October 19, 2026.
Where the old Edge TPU went
If you open an old coral.ai link, such as the USB Accelerator product page, you land on the new Coral NPU homepage. The first-generation documentation still exists. Google links to it as “First-gen Coral Edge TPU” in the Coral NPU site menu, and it opens an archived copy of the old Coral product site with datasheets, setup guides, compiled models and the Edge TPU compiler.
That archived page still lists the original lineup at its old list prices:
- USB Accelerator ($59.99): plugs into Linux, macOS or Windows 10 computers and Raspberry Pi boards.
- Mini PCIe and M.2 Accelerators ($24.99 each; $39.99 for the M.2 card with two Edge TPUs): for Debian-based Linux systems with a matching slot.
- Dev Board ($129.99), Dev Board Mini ($99.99) and Dev Board Micro ($79.99): single-board computers with the Edge TPU built in.
- System-on-Module ($99.99) and Accelerator Module ($19.99): for production hardware.
Those prices are from a page that hasn’t been updated since the 2022 ASUS IoT manufacturing announcement, so treat them as reference only. Real prices and stock vary by distributor. For example, RS UK lists the USB Accelerator (part G950-06809-01) with stock, while Australian retailer Core Electronics marks it retired. Search by part number rather than product name.
What the Edge TPU can still do
The Edge TPU is rated at 4 TOPS using 0.5 watts per TOPS, according to the archived Coral technology page. It is good at small vision models: image classification, object detection, pose estimation and segmentation at low power.
Its limits are the reason Google changed direction:
- TensorFlow Lite only. Models must be converted to TensorFlow Lite, fully quantized to 8-bit integers, then passed through the Edge TPU Compiler. PyTorch models need conversion first.
- Small on-chip memory. Larger models spill to the host or must be split across several Edge TPUs, which slows them down.
- No path to language models. The hardware wasn’t designed for transformer models or LLMs.
- Aging software. The PyCoral and Edge TPU runtime packages target older Python and Debian versions, so new operating systems often need workarounds.
If you already have one and your quantized TensorFlow Lite model runs on it, nothing forces you to stop.
What Coral NPU is
Google Research released Coral NPU on October 15, 2025, according to the Coral updates page. It isn’t a USB stick or a board. It is a design for a neural processing unit, based on RISC-V, that chip companies can put inside their own system-on-chips.
Google’s Coral NPU FAQ spells out the differences:
- Coral NPU is “an evolution” of Google’s edge AI strategy, not a direct replacement for the Edge TPU product line.
- The architecture and its MLIR-based compiler toolchain are open source under Apache 2.0. Companies pay Google no license fee.
- It accepts models from more frameworks than TensorFlow Lite. Google’s homepage names PyTorch, JAX and LiteRT.
- Google says existing Coral models should run on Coral NPU hardware as-is or with minimal conversion, once partners ship that hardware.
- First implementations target hundreds of GOPS. Partners can scale the design toward multiple TOPS.
You can’t buy “a Coral NPU” from Google. You buy a chip or board from a partner that includes it.
The Synaptics Coralboard, the first board you can order
Synaptics and Google Research announced the board on March 10, 2026. The official Coralboard datasheet lists:
- Synaptics Astra SL2619 system-on-chip with the 1 TOPS Synaptics Torq inference engine, the first implementation of Coral NPU
- 2GB DDR4 (a 1GB option exists), microSD and NOR flash
- MIPI CSI camera and DSI display connectors, I2S audio, USB-A host, USB-C power (5V only)
- M.2 slot for optional Wi-Fi and Bluetooth, plus mikroBUS, Qwiic and a 20-pin GPIO header
- Yocto Linux, with models compiled by Synaptics’ open-source Torq compiler
The Google I/O edition adds a mikroBUS hat board and an Arducam 1080p camera module. Google’s Coralboard page sends buyers to a DigiKey wait list and shows no price, so check DigiKey for current cost and shipping dates. Synaptics, not Google, handles hardware support.
Note the raw numbers: 1 TOPS on the Coralboard is lower than the Edge TPU’s 4 TOPS. The new board’s case rests on broader framework support, a full Linux SoC around the NPU and an open toolchain, not on peak throughput.
Which one should you pick?
- You have an Edge TPU project that works: keep it. Buy spare units by part number while distributors still have them, and pin your OS and Python versions.
- You’re starting a small vision project on a Raspberry Pi: the Raspberry Pi AI HAT+ (Hailo-8L or Hailo-8) has current software support and more TOPS. See our AI HAT+ comparison.
- You need more compute for robotics, or want to run language models locally: look at NVIDIA Jetson; our Jetson Orin Nano vs Orin NX guide covers the entry options.
- You want to try Coral NPU or build your own chip around it: join the Coralboard wait list, or start with the free simulators and RTL in the coralnpu GitHub repository.