Jin Daily Tech Trivia: Google launch CoralBoard. AGAIN. 4th Time
Google Coral is basically what happen when Google see the community doing IoT + edge AI and go:
“Wah, got hype ah? We also want enter.”
But instead of working properly with Pi and the community, Google choose the very Google route:
“Let’s make our own board, own OS, own compiler, own stack, own suffering.”
First gen was the NXP-based Coral Dev Board. For that time, the Edge TPU was actually more powerful than the Pi 3. 4 TOPS of AI, WOW.
But then developers ask the normal questions:
“Where to buy this?” “Why so expensive?” “Why my model cannot compile?”
And since it use the NXP i.MX 8M, which is subject to US export regulations, pretty much no developer outside the US can get it easily. And let’s not even talk about the weird runtime versions, model compatibility, and all the other Google rituals.
Google: “Why no one buying our Coral?” Community: “Maybe because expensive, supply limited, and the stack hard to use?”
So came the MediaTek Coral Dev Board Mini.
NXP too expensive? Taiwan chip then. Chromecast also use MediaTek, should be cheaper right? RIGHT???
Still expensive. Still Google stack. Still fragmentation. Different SoC, same problem. And worse, because the MediaTek documentation is not even there for community to read.
So Gen 2 also become like this: hardware exist, but nobody know how to use it.
Google: “The previous boards power hungry. Maybe people need battery life?”
So came the Coral Dev Board Micro, based on the NXP RT1176 microcontroller + Edge TPU.
This one is not a normal Linux SBC anymore. It’s more like TinyML teaching-kit energy. Very low power, yes. Can do embedded examples, yes. But if you ask:
“Can run a server or a bigger project?”
The board will look at you and resign.
It’s microcontroller world. Not “computer with AI” world.
Then the AI hype train arrive. Gemini everywhere. Gemma everywhere. Local AI everywhere. Some Google manager probably look at the Coral Micro and say:
“We need to get Gemini running on it!” Engineer: “Boss, cannot lah. Gemini too big, the board will cry. Best I can do is Gemma 3 270M. Technically still Google AI, right?” Manager: “…fine. Ship it.”
So the newer Coral direction changed again. Instead of only selling one proprietary Edge TPU board, Google now push a RISC-V-based Coral NPU architecture and let silicon partners implement it.
First major partner: Synaptics, with the Coralboard / Astra SL2619.
Still called a Coral dev board. But new chip, new stack, new toolchain, new world. Old Coral work cannot just carry over at all. And worse performance somehow, now only 1 TOPS of AI.
And that’s not even a joke - the new Coralboard literally ships running Gemma 3 270M out of the box. The smallest Gemma they got. On a 1 TOPS board.
Meanwhile, the Raspberry Pi side already got AI HATs with Hailo accelerators (13 TOPS), a stronger community, better tutorials, and the usual Pi advantage: when something break, at least someone online already suffered before you.
So the moral of the story: don’t buy or use experimental Google products. Whatever they launch will just hit end-of-life suddenly, and nobody gonna care about you after that.
