The Arrival of a Google Coral Ecosystem?
Google's Edge TPU just became a lot more interesting.
Announced just over a year and a half ago now, Google’s Edge TPU based hardware, intended to accelerate machine learning inferencing at the edge, is the fastest accelerator hardware on the market. But, at least until yesterday, you could only get your hands on the Edge TPU silicon if you bought Google’s own Coral Dev Board, or their USB Accelerator.
The first sign this was about to change was the announcement of the ASUS Tinker Edge T board which started shipping at the very tail end of last year.
However, since the initial release of the Coral hardware last year, Google has always said, or at least strongly implied, that the only way you would be able to buy the Edge TPU on its own in volume was to purchase the 40×48mm system-on-module (SoM) used by the Coral development board. While they did at least announce the potential availability of a PCI-e accelerator board “…towards the end of the year,” their partnership with ASUS — whose Tinker board also uses the Edge TPU SoM — seemed to confirm that this was the plan.
Yesterday’s announcement from Google included not just the original Mini PCIe Accelerator board which they’d announced with the release of the original round of Coral hardware, but additionally two more PCI-e boards: an M.2 Accelerator A+E key and a M.2 Accelerator B+M key.
Interestingly, yesterday’s hardware release also included the announcement of another new board, the Coral Dev Board Mini. Built around the MediaTek 8167s SoC, the new smaller developer board is a lower powered version of the original development board.
However perhaps the most significant announcement amongst yesterday’s bag of hardware is the new Coral Accelerator Module.
“…the Coral Accelerator Module, an easy to integrate multi-chip package that encapsulates the Edge TPU ASIC. The module exposes both PCIe and USB interfaces and can easily integrate into custom PCB designs.”
The accelerator module is a solderable multi-chip surface-mounted module that includes the Edge TPU. This is something has been lacking so far in Google's line up. Because while you could pick up Intel’s Movidius chip, and integrate it into your product, you couldn’t do that until yesterday with Google’s Edge TPU — at least not without the overhead of using their SoM.
As a result of the availability, and Intel's first mover advantage, most of the machine learning ‘accelerator’ products you’ll find on the market are built around Intel’s Movidius chip. So the release of the Google Edge TPU as a solderable module is actually sort of a bigger deal than you might think, especially when you consider the performance when compared to the older Movidius chip. Because Intel is really paying for that first mover advantage, and the assumptions they made at the time.
It turned out that quantisation is a big and important thing when it comes to doing machine learning inferencing at the edge, and right now at least, Google are the only ones out there that are taking advantage of that fact.
The Coral SoM is now available to pre-order Seeed Studio for $114.99, while the Coral Mini PCI-e Accelerator module, the M.2 Accelerator A+E Key, and the B+M Key are all available for $34.99. If you’re at CES next week, you can see the new Google hardware demo’ed at the Murata booth at the Las Vegas Convention Center, Tech East.
While it’s not yet available, the Coral Accelerator Module — the Edge TPU on a solderable module — should be available some time during the first half of this year. Although pricing isn’t yet known, the module pricing is presumably going to be below the $35 price point of the PCI-e keys, and that is going to seriously disrupt things for the people building third party machine learning hardware. Because, at least right at the moment, Google is far in the lead when it comes to inferencing hardware.