Coral NPU: A full-stack platform for Edge AI
research.googleI'm guessing you can still find Coral TPU-based boards somewhere but not sure what the support for these will be now that the focus is shifting. Coral TPU also uses subset of tensorflow and its nice to see that the open standard is targeting jax and torch.
When I went to see if anyone is selling the boards or their "Partners" page regarding manufacturing design I got 404 even after signing in: https://developers.google.com/coral/guides/coral/resource
Also seems the core hardware was open-sourced? Github: https://github.com/google-coral/coralnpu
Will be interested to see what their developer kit looks like when (and if) they release it.
I've been experimenting with the BeagleY-AI to build a little edge AI gizmo with a camera (Texas Insturments SoC + 4 TOPS NPU in RPi 5 form factor)
https://docs.beagleboard.org/boards/beagley/ai/demos/using-e...
That's cool, any success stories, challenges or other feedback you can share?
I've only heard of people using Coral PCIe / USB for edge image AI processing tasks like classifying subjects in a stream. Curious if you have the same use case or something different!
I'm trying to make a DIY security camera that can run local models, and stream video over wifi.
The TI SDK makes it easy to run demos but making any custom apps quickly gets complicated unless you are familiar with embedded Linux dev, Yocto, etc. Certainly much more complex than iOS/Android.
Hopefully over time the tools for embedded can catch up to mobile.
This seems cool:
> Hardware-enforced privacy
> A core principle of Coral NPU is building user trust through hardware-enforced security. Our architecture is being designed to support emerging technologies like CHERI, which provides fine-grained memory-level safety and scalable software compartmentalization. With this approach, we hope to enable sensitive AI models and personal data to be isolated in a hardware-enforced sandbox, mitigating memory-based attacks.
It seems here that Google provides the core IP, and that Synaptic packages this (and probably other related IP blocks) block that can be used to build a SoC. As of now there are no chips announced. So it will be some years before we as software/electronics engineers get to play with it.
The architecture seems to be RISC-V array with standard RVV vector instruction set. That is a quite familiar environment for software developers compared to custom systolic arrays.
I think they announced 5 chips[0] in the SL2610 product line[1] today. It appears they're combining 1-2 ARM Cortex-A55 with a Cortex-M52 and a 1 TOPS NPU. Somewhat more complete data sheet here[2] (which is still a bit anemic, IMHO) There are photos of the devkit hardware that will be offered at [6] if you scroll up a little bit.
The original Google Coral lineup offered a 4TOPS accelerator back in 2018-2020.[3][4]
The original (4 TOPS) Coral used ~1 watt while this new Coral TPU is designed for 10mW/0.5 TOPS.[5] That power budget fits well alongside low power MCU's.
It doesn't appear to include any hardware accelerated video encoding[6] for H264/etc, which was also a massive limitation of the original google coral (improved slightly in the Dev Board Mini). There's a lot of documentation on consuming WebRTC content, and while streaming out is mentioned ... any encoding would have to be performed on one of the A55 cores at dubious performance levels. The RK3588, for example, includes a VPU for hardware accelerated video encoding (H264/HEVC encoding @ 8k30fps).
0: https://cdn.bfldr.com/ZU41R0OK/at/tjm5s8hrmz5mgqrjtsrc5c4/sl...
1: https://www.synaptics.com/products/embedded-processors/sl261...
2: https://cdn.bfldr.com/ZU41R0OK/at/ks4thp8bw9n3bt2ktms3k34s/s...
3: https://abopen.com/news/google-launches-coral-edge-tpu-devel...
4: https://developers.googleblog.com/en/new-coral-products-for-...
5: https://developers.google.com/coral/guides/power
6: https://synaptics-astra.github.io/doc/v/latest/linux/index.h...
Looking at the state of the original Coral TPU (which was basically abandoned, just like regular other Google stuff), would make me very wary to use this is a long term product.
the coral tpu does what it does and it's not bad at it. The documentation is good , and quite a few people use them practically. They're available readily.
What's upsetting about the state? Continued development?
however, to your point : being google affiliated is a huge red-flag for longevity.
They were completely sold out for 2-3 years (2020 onwards), and Google wiped documentation (https://coral.ai/products/accelerator/ redirects to the main page, which has no reference to the original Coral). I can't tell where there's an official place to buy this. I see some on Amazon, but that might be resold.
Huge amount of drivers, all for 6.12, which is already pretty old. https://github.com/synaptics-astra/linux_6_12-drivers-synapt...
Agreed that that no hardware video encode is pretty damned deflating.
Having such a low power device is incredibly enticing. Thanks for the good details. One random thing I'm kind of excited over, there's 3x I2S audio connections, which has some weird fun use cases (ideally a little field recorder?).
One curious thing, Google's post says they are still trying to finalize the matrix extensions for RISC-V. I'm assuming those simply aren't on these Synaptics chips?
Did you check out the STM32N6? It apparently has an h264 encoder
Didn't they abandon the previous Coral accelerators?
Pretty much. They used to be „plug and play“ but running them with modern tensorflow is a huge pain
They aged out of being competitive against SoC NPUs. This is ultimately the successor to that.
Does "it aged out of being competitive" mean the same thing as "they abandoned it," or does it mean something else?
This seems like important work and at first I wondered what this does for Google's bottom line. However reading about the simulator for software dev and the hardware kits, Google is aiming to win the AI glasses, etc. edge wars. All makes sense.
Google's track record of suddenly dropping something they have developed and their stellar record on privacy make me pretty wary of this. I would like something like it, but Google as the champion just doesn't give me confidence.
How much would cost to produce these ?
Price?
is hardware still squirrelled away behind mega paywalls of one or two companies holding all the EDA software?
someone at ycombinator should create a "github for silicon IP" company. that would be awesome.
Like OpenCores?
opencores is cool but its not very New and Shiny... upon further thought, i suppose a github for ip would just be github
Interesting that in 2025 they only provide C compiler support as building tool, who cares about security in AI systems.
This is blatantly false. There are MLIR based backhands for other languages. This is explicitly mentioned on the landing page at https://developers.googleblog.com/en/introducing-coral-npu-a....
Please language troll somewhere else.