Hacker News new | past | comments | ask | show | jobs | submit login
Google Coral Dev Board mini SBC is now available for $100 (cnx-software.com)
149 points by todsacerdoti 5 days ago | hide | past | favorite | 104 comments

The Nvidia Jetson Nano costs the same and is less likely to be killed-off after you’ve invested in the platform.

Jetson is actually an important product for Nvidia and Google tends to kill off this type of pet project.

Google/alphabet might have more success with their side-bets if they spun them out as separate companies like Xiaomi and Haier (both Chinese) seem to do.

Coral is powered by an Edge TPU (Tensor Processing Unit), which wipes the floor with GPUs like the Jetson Nano when it comes to running Tensorflow:


...and Google is pretty invested in TPUs, since it uses lots of them in house.


They might be great for inference with tensorflow - but from what I can tell from Google's documentation, Coral doesn't support training at all.

I'm sure an ML accelerator that doesn't support training will be great for applications like mass-produced self-driving cars. But for hobbyists - the kind of people who care about the difference between a $170 dev board and a $100 dev board - being unable to train is a pretty glaring omission.

You wouldn't want to use it for training: This chip can do 4 INT8 TOPs with 2 watts. A Tesla T4 can do 130 INT8 TOPs with 70 watts, and 8.1 FP32 TFLOPs.

Assuming that ratio holds, you'd maybe get 231 GFLOPs for training. The Nvidia GTX 9800 that I bought in 2008 gets 432 GFLOPs according to a quick Google search.

Hobbyists don't care about power efficiency for training, so buy any GPU made in the last 12 years instead, train on your desktop, and transfer the trained model to the board.

On the other hand, it would be useful for people experimenting with low-compute online learning. Also, those types of projects tend to have novel architectures that benefit from the generality of a GPU.

Last I’ve heard covid was making GPUs about as difficult to find as the other things it’s jacked the prices up on, too.

You can get pretty much any GPU at pre-COVID prices right now, except for the newest generation NVIDIA GPUs that just came out to higher-than-expected demand.

As a hobbyist in a state with relatively high electricity prices, I do care about the power efficiency of training.

Training is what the cloud is for.

That makes a $170 board that can also do training look dirt cheap in comparison

Good luck training anything in any reasonable time on it.

Useful for adapting existing models. Not everything needs millions of hours of input.

If you want to train yet-another-convnet sure, but there could be applications where you want to train directly on a robot with live data, as in interactive learning.

See this paper for an example of interactive RL: https://arxiv.org/abs/1807.00412

or a highly rigged machine, this looks more for fast real time ML inference on the edge

You can adapt the final layer of weights on edge tpu.

Training on a dev board should be a last resort.

Even hobbyists can afford to rent gpus for training on vast.ai or emrys

Google is pretty invested in TPUs for their own workloads but I fail to see any durable encouragement of them as an external product. At best they're there to encourage standalone development of applications/frameworks to be deployed on Google Cloud (IMHO of course).

AFAIK, apart from toy dev boards like this, you can't buy a TPU, you can only rent access to them in the cloud. I wouldn't want my company to rely on that. What if Google decides to lock you out? If you've adapted your workload to rely on TPUs, you'd be fucked.

What's the difference between Coral's production line of Edge TPU modules and chips [1] and Google's cloud TPU offering?

Note: I haven't tried sourcing these in production (100k+) quantities so I have no idea what guarantees that product line gives customers.

[1] https://coral.ai/products/#production-products

Edge tpu is 2 tflops at half precision, cloud tpu starts at 140 tflops single precision and scales further.

Also edge tpu is 2-5Watts. Supposedly cloud tpus are more power efficient than GPUs, and for eg the 14 tflops 2080 ran at 300 W regularly.

They're nothing alike at all. Similar to how a low end laptop GPU differs from a top of the line NVIDIA datacenter offering. Google's cloud TPU offering is the strongest ML training hardware that exists, the edge devices simply support the same API.

Coral can only run inference, and is optimized for models using 8-bit integers (via quantization).

A full TPU v2/v3 can train models and use 16/32 bit floats. They also have a Google-specific (?) 16-bit floating point type with reduced precision.

And don't forget, TPUs are horrible at floating point math! The errors!

Yea I've been wondering about charts I've seen comparing tpu model quality perf to gpu model quality like here [1], whether that could be due to error correction. At the same time training on gaming gpus like 1080 ti or 2080 ti is widely popular, though they lack the ECC memory of the "professional" quadro cards or V100. I did think conventional DL wisdom said "precision doesn't matter" and "small errors don't matter" though.

I've noticed this difference in quality perf in my own experiments tpu vs gaming gpu, but don't know for sure what the cause is. I never did notice a difference between gaming gpu trained models and quadro trained modela. Have more info/links?

1: https://github.com/tensorflow/gan/tree/master/tensorflow_gan...

Until you want to use Pytorch or another non tensor flow framework the support goes down dramatically. Jetson Nano supports more frameworks out of the box quite well, and it ends up being same cuda code you run on your big Nvidia cloud servers

Not only that, nvidia cares deeply about pytorch. Visit pytorch forums and look at most upvoted answers. All by nvidia field engineers.

That benchmark appears to compare full precision fp32 inference on the nano with uint8 inference on the coral, that floor wiping comes with a lot of caveats

There seems to be more than one jetson board.

The Jetson nano has 2GB of RAM and is $59.


what a nice price point

This is rpi4 territory.

But I would not recommend the 2GB version. The 4GB versions is barely useful without a swap file on a SSD.

Why not just use ZRam and go headless? There are plenty of good ncurses apps out there.

Note that Zram and ML are not best friends, for a number of reasons.

Maybe 2GB is barely usable (I'll take your word), but the google coral dev board this thread is about is $100 and has 1GB.

The base board for the 2GB version removed the mini PCIE slot, meaning you can’t swap in a cheap drive either.

Google Brillo was renamed to Android Things four years but apart from the change of names, the boards are still supported: https://developer.android.com/things/get-started/kits

While they may be supported, the store links for the kits either 404 or the item is discontinued.

It's google, what did you expect?

This is why I would never recommend this SCB to anyone. Beside being locked to tensorflow, you are also at merci of some random manager at Google.

jetson's x1 core (note: not arms new x1 architecture!) is already 5 years old. once upon a time that would scare me, but now, it seems almost comedically safe to say "I guess it's not going anywhere!"

And it's still faster than the Coral Dev Board mini... (Cortex-A35 is a CPU tier _below_ the A53, there's no contest).

The fastest SBC at CPU tasks priced below $100 is the Raspberry Pi.

"The fastest SBC at CPU tasks priced below $100 is the Raspberry Pi."

The Odroid N2+ costs $79 and is over twice as fast as the Pi4. The Khadas Vim3 costs $100 and is about 30-40% faster than the Pi4.

The number of SBC boards out there is becoming huge; although the PI price has dropped significantly wrt performance and features (especially RAM), there's a lot of comeptition, and it's growing.

https://hackerboards.com/spec-summaries/ https://all3dp.com/1/single-board-computer-raspberry-pi-alte...

> Cortex-A73 at 2.4GHz

That's indeed much faster than the Pi4. Do you know the state of kernel support for that board?

It uses an Amlogic S922X aka the G12B. Support is generally pretty good, there's a dedicated community that has been very active pushing upstream[1].

Except the ARM G51 Bifrost gpu, which has only recently started to see viability[2] thanks to one hacker's reverse engineering. If you want to read a lot of words, there's a status report from the libreeelec Kodi-based media player distribution distribution that's a year old, that lays out a lot of what needs be done, from a very video-intense perspective[3]; this is before recent reverse engineering efforts, & largely discusses uses closed proprietary blobs, but still interesting. Most recently & very interestingly, there are signs that ARM itself may be willing to start helping out the reverse engineered development[4], which would be a new potentially interesting state of affairs.

[1] http://linux-meson.com/

[2] https://www.phoronix.com/scan.php?page=news_item&px=Bifrost-...

[1] https://forum.libreelec.tv/thread/21134-what-aspects-of-hard...

[4] https://www.phoronix.com/scan.php?page=news_item&px=Arm-Panf...

According to the Armbian (one distro to support them all:^) page, mainline kernel support is complete, although they say there still could be some network problems. From what I read on their forum, the Hardkernel Ubuntu-based image is currently more stable than the Armbian one.

https://www.armbian.com/odroid-n2/ https://forum.armbian.com/search/?q=odroid%20n2%2B&fromCSE=1


Dietpi also supports the N2, which is very similar to the N2+. https://dietpi.com/

Aren't there better odroid options if you mainly care about compute?

The $63 N2+ has the latest "C" rev of the S922X, which is a dual 2.4GHz A73 + 4x A53 and a "MP6" variety of bifrost GPU, a G51. The C4 has the newer S9005X3, which has 4x 2GHz A55 cores and a smaller G31 bifrost gpu. Those A55's, while improved over the A53's, are going to be sigificantly outmatched by the A73 cores on the N2+.

The H2+ has an Intel J4115 Atom celeron running 2.3GHz all-core, which I expect would trounce these ARM chips. It's also $120.

Alas there hasn't been any update to the excellent Exynos5422 that started HardKernel's/Odroid's ascent as the XU4. Lovely 2GHz Cortex 4x A15 4x A7 with (2x! wow! thanks!) USB3 root hosts and on-package RAM: really an amazing chip way ahead of it's time. These days it's way outgunned but this chip really lead the way for SBCs with it's bigger cores for the time, USB3, and on-package RAM (which we really need to see a comeback on).

Worth noting that the A73 on the N2/N2+ and RPi4 are from ARM Artemis, which hails from 2016. Maybe some year SBC won't all be running half decade old architectures, but at least we're at the point where half a decade ago we were doing something right. ;) Still, one can't help but imagine what a wonder it would be if an chip & SBC were to launch with an ARM X1 chip available.

it's an a57 on the x1 (an architecture from 2012, but a big core), so this coral mini's a35 (new but quite small) very significantly below.

the attraction of coral is supposed to be the inference engine. 4 TOps/s at 2 watts is... impressive. Jetson takes 10 or 15 watts & tops out a little under 0.5 TOp/s. those are much more flexible gpu cores but that's 60x efficiency gain & centered around a chip that is much easier to integrate into consumer products.

Compared to the Nano yeah, which is just the same SoC since a long time.

Xavier NX is 21TOPs at 15W for the whole SoC... but the pricing at $399 puts it in a different category...

Google should just start selling the USB sticks at the same price as the M.2 Corals, with them being used on RPis I think...

> Jetson is actually an important product for Nvidia and Google tends to kill off this type of pet project.

Jetson has obsolete distros though? Linux support is probably better with Google if anything.

Which is odd because they created Alphabet to do that very thing.

And here I thought they created Alphabet because of Wall Street pressure.

If you're looking for an even easier & cheaper way to start experimenting with an Edge TPU, the $59 Coral USB Accelerator has been out for a while now: https://coral.ai/products/accelerator

Check out the Level Up series for getting started: https://www.youtube.com/watch?v=-RpNI4ZrfIM

And here's a fun real-life application: https://www.youtube.com/watch?v=jIyM_qT9RZw

I’m not versed in this field so curious, in that real life application, why not just use the computer you’re broadcasting from to run the camera switching client and ML inference. Does the USB accelerator do something that a standard desktop can’t?

Real-world use cases wouldn’t typically have a powerful computer broadcasting video. Rather, this board would be used at the edge to drive a camera and offer on-device, low-power inference.

Useful for robotics/drones, surveillance cameras, vehicles, consumer devices etc.

I put it on my Raspberry Pi 4 with some mobile-resnet version (88 objects) to do real time object detection from a Pi camera. It's quite a small package.

The accelerator is much slower than just running on the host PC... an RTX 3070 has 81TOPs at FP16 and 162TOPs at INT8 for example.

Doesn't the 3070 cost much more than a basic PC and the accelerator? Isn't this targets at those wanting to get their feet wet without vast outlay?

> SoC – MediaTek MT8167S quad-core Arm Cortex-A35 processor @ 1.3 GHz with Imagination PowerVR GE8300 GPU

Is my knowledge outdated or are MediaTek SoC generally a pain to work with in GNU/Linux ?

Any information whether this SoC is supported by mainline Linux or if support will be merged in mainline Linux?

That is very important imho.

Last time I checked SBC opensource GPU driver development pretty much everyone gave up on PowerVR.

This is why my Beagleboard is gathering dust. The proprietary GPU drivers only worked with particular kernels that were usually out of date. Steer clear of PowerVR would be my advice.

I feel like this would be a dramatically better device if they worked with the Raspberry Pi Foundation instead of heaping it into a board with a PowerVR GPU, one of the few remaining major mobile/embedded GPU lineups with no meaningful upstream support, nor any acceleration through Mesa.

You can get that better device by attaching the $60 Coral USB Accelerator to one of the Raspberry Pi 4's USB3 ports. You can even install more than one USB Accelerator per host machine if you want. (But if you go far down that line maybe you should get a nVidia Jetson board instead.)

I don't know why the Coral folks bother with the dev boards; I think their USB accelerators and M.2 cards are better buys. The Pi4 + USB accelerator + a SD card costs roughly the same as this dev board and is more powerful with better community support.

I'm hoping they release a newer, faster version of the Edge TPU itself. There are some newer, faster AI accelerators from other vendors (the Gyrfalcon Lightspeeur 2803S, the Hailo-8) but the Coral USB Accelerator is reasonably priced, actually available to hobbyists, (as of recently) has open-source driver for TensorFlow, etc.

Note: I work for Google, but I don't have any inside knowledge about Coral.

Or use an Allwinner SoC.

Building software for new Google platforms is like farming on the river delta. Yeah, soil is good ... but any day the river will change its path and wash all your work into the fucking ocean.

I'll never do it.

You aren't wrong.

Google has screwed their biggest fans too many times and we aren't Apple drones that can be pushed around.

The MIPI-DSI display is nice for prototyping actual devices I suppose, but I don't think this is a great fit for hobbyists. The only thing hobbyists might use this for is the machine learning chip and with a $70 price difference between this and a Raspberry Pi (or similar) with much more normal processing power, I don't think many people will take this deal.

I guess this board is for specialized applications.

>> $70 price difference between this and a Raspberry Pi (or similar) with much more normal processing power

> AI/ML accelerator – Google Edge TPU coprocessor with up to 4 TOPS as part of Coral Accelerator Module

Maybe this will make a difference in the specialized AI/ML workloads ?

Not for anything I'd wanna use it for. USB 2 makes it hard to get data in and out. Video is 720p single stream. I'm grasping at straws here.

Seems like a total loser compared to the NVidia Jetson line. It does have a few more TOPs than the Nano, I guess? And it costs a bit less than the Xavier series?

Jetson Nano doesn't have dedicated AI hardware... so 472 GFLOPs that you can use there pretty much. Same applies to TX2.

NV hardware with dedicated AI starts with Xavier, with the NX having 21 TOPs at $399 and 32 TOPs for the AGX at $699... that might leave a space for Google Edge TPUs.

I don’t think the Jetson nano even has a DL accelerator, so I could see the Google SBC achieving much better INT8 inference performance. The Jetson has a CUDA capable GPU, so the comparison is kind of apples and oranges.

The Google unit will definitely achieve better DL performance than the Nano. But the set of applications where:

* The Nano's 0.5 TFLOP of CUDA is insufficient; AND

* Google's 4 TOP of DL performance IS sufficient; BUT

* You're not bottlenecked by things like USB2; AND

* You can't afford a higher-end Jetson with dedicated ML

Well, let me just say I can't think of one on things I've actually done. I could come up with ones in abstract, but it seems like Google is aiming for 1% of the market at best. With 1% of the market, NVidia will win on community, tooling, R&D budget for v2, etc., so in the end, that means 0.1% of the market.

And that's if the two businesses were starting on even footing. They're not. Google has a horrible reputation for leaving customers high-and-dry, while NVidia's reputation is pretty good for B2B. Even if the boards were equivalent, most would pick NVidia based on that alone. Plus, NVidia is coming in with established users; they have a first-comer advantaged.

Getting late to the game with a generally inferior product with backing from a less reliable company? It seems like a loser to me.

Not related, but I found amazingly rich x86 SBC with 2x M.2 slots, Seeed Odyssey, $188 or $218 with 64GB eMMC. https://www.seeedstudio.com/ODYSSEY-X86J4105864-p-4447.html

Paired with a $25 Coral M.2 Accelerator [1] this is actually an interesting alternative. Twice the price of the Coral Dev Board Mini, but much better hardware.

1: https://coral.ai/products/m2-accelerator-bm

Neat. Looks pretty similar to the ODROID H2+. Some differences I see: the Odyssey has soldered-on RAM and eMMC, the extra M.2 slot, fewer USB3 ports, and 1 GbE rather than 2.5 GbE. I think they're similar in price (once you add the missing RAM and storage to the ODROID H2+).

Is 110x110mm a standard form factor? Are there cases available that could fit one or two 3.5" HDDs?

I wanted it look pretty, so i went with cheap slim brushed aluminium miniITX box. Fixing MB, notebook PSU and disks will be custom job.

Note that one m.2 is key M, and the other is key B.

Almost all NVMe SSDs are key M.

J4105 have total of 6 PCIe lanes, thats probably why. It'll still take SATA SSD in second M.2.

Is this going to be one of the killed by Google projects or can we only hope that it'll be alive

Absolutely do not expect long term support.

I'm more curious if the community will be able to keep the project alive or if Google is going to keep the project theirs.

right now google is at the "makes cool projects and you should get what value you can from it in the first year then dip" stage.

so yeah, it's a cool product that came out. give it a whirl but don't rely on it.

the problem with these boards is that there are not enough general purpose cores, and the TPU is only really usable for a very small niche set of computing problems. only 0.00001% of developers is really interested in doing anything with AI, which is mostly a fad hyped up by the media anyway. what we need is a 16 core A17 or A72, without peripherals that are only useful to 0.01% of developers out there. these cores can then be applied to a variety of real world use cases where 4 or 8 cores is simply not enough.


You did not even mention Google. Your comment reminds me of people who said GPU is not an important development a decade or more back. I was in high-school back then and can see why TPUs can be super relevant 10 years from now. Someone has to start...

the problem with GPUs is the same - really only applicable to graphics and now since a few years AI, shoehorned into something which was never meant to deal with AI workloads. meanwhile nothing has been done to increase the number of cores, which is really important to be able to run a wider variety of computationally intensive applications on embedded/mobile devices. what most SoC companies are doing is simply slapping on a variety of irrelevant peripherals instead of focusing on what's important - more cores.

I find RAM to be the limiting factor on my AArch64 systems, not the number of cores.

I can imagine that is a problem for some applications. So again, cores/RAM/... basically essential stuff, more important than TPUs or whatever.

I'd rather think your comment has been downvoted because of the baseless claims therein. If you don't like that device, then, well, that's like your opinion man.

it's not about me not liking the device, read what I wrote - and I can add that I've designed and shipped hardware products with SoCs in them, so I know what I'm talking about.

Google is criticized all the time here, far more than it's praised, and your comment is not downvoted.

What's a good use case for a board like this?

Low-power and very fast inference for a relatively constrained set of model architectures. Think robots, drones, surveillance cameras, consumer devices.

So the Edge TPU has been out for a while now, can anyone point out a large enterprise customer that is using this for real for something new that we couldn’t do before? Or is the “edge” just made up by salespeople. Like there needs to be super low latency requirements before someone decides they can’t use an internet api right?

Not an actual existing product, but imagine something like a fish trap that only closes when a certain species of fish goes in. Not everything can be readily hooked up to the internet. And the cell bandwidth of constant upload video stream isn't readily affordable even in things like cars.

Visual inspection for assembly lines is a pretty classic example.


my problem with the coral usb that I got was that Google's own AutoML service did not fully support it. Which sucks since they own it...

I created a raspberry pi cluster with the usb coral sticks and some ssds. Coral sticks do get pretty hot. I wonder how this will do.

I like the that my pi’s all have usb 3 ports and 8gb of RAM.

It was also pretty easy to setup K3s and work through any issues thanks to the Pi community activity.

I have a desktop with an excellent processor, RAM, and decent gaming GPU in it. Is this thing remotely worth it for me, or should I stick with what I've got? Who is the target market for these? People who only have laptops with shitty GPUs?

Does anyone make Beowulf clusters of these? I could see that being more cost-effective if you're resource-limited even by what higher end GPUs can do.

The target market seem to be people who want something to integrate ML in their robot or RC car. It features decent inference performance with (comparatively) low power demand. But they can't even do training.

Unless you're building something with space and/or power constraints you are much better off with a laptop or desktop.

I have to say, this sounds like quite a niche market they're targeting here.

IoT and embedded computing isn't a niche market. Most households have more embedded and mobile devices than they have desktop/laptop computers.

It's a variant of their development board for their AI accelerator ship for embedded setups (think quality control, licence plate detector, analysing visitors via webcam etc). This variant is probably just to make it more attractive for hobbyists; increasing mindshare and community size. It doesn't need a huge market, and maybe people come up with cool ideas nobody has thought of yet.

I’m going to use this or something like it to control a hose to blast water at the deer eating my plants.

It's much slower than your desktop, or higher-powered embedded options from NVIDIA for example (4Tops of AI inference perf... and no training support by design)

Note that the Cortex-A35 is a CPU tier _below_ the A53 (80% of the perf at 32% lower power).

I hate the use of SBC for Single Board Computer. First, the acronym is already in wide use for Session Border Controllers. Second, most PCs are single board computers (just a motherboard), so the name is meaningless.

The "single board computer" meaning of SBC dates back to the mid-1970s when microprocessors first started coming out. Whatever a session border controller is, it's unlikely to pre-date that.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact