I bought their m.2 dual tpu and the drivers are in a really bad state
They got in the linux kernel, staging, then basically no work done. They continued the development off-tree, going out of sync with what the kernel had. Only 4.19 supported if I remember correctly.
Finally removed from the kernel in...5.13?
Driver wise I can't say that this company is trustworthy
Yep, raspberrypi, arduino and in the last few years, esp* are defacto standards just because of the great software (and library) support.
Yes, STM, XYZ, CabbgePi, etc., is better, faster, has sata, has emmc storage, has more pins, less power usage, has this and that, but if I can't program it now, because of software bugs, and won't be able to program in in 5 years, because the toolchain will be abandoned by now, I really don't need it.
source: I have a bunch of microcontrollers and single-board computers, collecting dust in a drawer at work.
Spot on, better a crappy hardware platform with excellent software than the other way around, even if in theory you can fix it all, if you value your time the decision is easy.
This was the bane of my existence for a few months. Tried it on several different systems and the best I could ever do was getting a single TPU to show up.
Eventually I just sold it, and back when you could buy them, picked up a couple more USB Corals. Works well enough, but it’s ugly.
it's a hobby project for a few folks with tenure at google. I've tried to buy a bunch of these over time and the few I was able to get were close to useless.
Coral products in general are made from unobtanium. Especially the m.2 variant that's keyed right for desktop motherboards. The mouser page says it's EOL but support claims it's still manufactured.
Also at first glance, the docs and software support are questionable at best.
Brave of them to release new hardware, when even their old stuff has ridiculous lead times.
I've been trying to find the Coral USB Accelerator[0] for a Frigate installation for a few years now. Mouser Europe has a lead time of ... 81 weeks! Other distributors just say "nope".
Agreed, their existing products are impossible to source. I've been trying to find a USB Accelerator, Mini PCIe Accelerator, or M.2 Accelerator for quite a while. I don't even care which, just any one of them! Deciding to release a "Dev Board Micro" when they already have a "Dev Board" and "Dev Board Mini" and all their other "Accelerator" products are unavailable seems a bit silly.
Crazy, they were handing these things out as conference freebies a while back and now they're $400 on eBay. Pretty sure I still have one in a box somewhere.
At the same power level and general availability with OK support, you're looking at the MyriadX/NCS2. Even then I think the Coral beats it on performance. Next step up is the Jetson Nano (10-15 W) which is a lot more beefy, but power hungry. There are also a bunch of chip companies that make accelerators, like the RK3399Pro from Rockchip and some other OEMs from China, but I have no idea how good the docs and availability are if you're not a big partner. The RK3399 has mainline kernel support which is attractive, but again there aren't many reliable board vendors. Khadas make some nice looking SBCs.
I was also faintly surprised to learn that the Paralella still exists! I always wanted one to play with and I thought they discontinued them.
I was playing with the K210 RISC-V 64 AI Camera from M5 Stack. It is fun, but the training takes place on a server vs. the chip. There is supposedly a way to do it on your own laptop, and then download it to the chip, but I stopped tooling around with it by that point. I was looking to use it for some hydroponics ML work.
You can buy the chips from them for a sufficient price.
The price is probably something unreasonable, with a high minimum order quantity or high lead time - because that's typically what happens when you see "Ask for Price" or "Request a Quote" (folks who do that often want to reduce your ability to comparison shop easily).
I am pretty good at devices like this. Have done a lot of raspberry pi development, arduino, nvidia jetson etc. for real ML production application. I ordered a Coral before and just had no luck. Worked for some contrived applications, but had nothing but trouble with the drivers and anything remotely custom just failed. It's cool tech, but I'll pass on this vanity project that will undoubtedly get cancelled right as people start to rely on it.
I'm on the fence. It's a very nice device if you can get your models working on it - basically untouched at the price/power point. Drivers for me have been OK. I have an M.2 card connected to a Jetson devkit (makes for a nice embedded test bench) and it runs fine, no worse than the NCS for setup anyway^. There were a couple of PCI settings to tweak but I documented the setup here [0]. For common use cases it's a decent option, I think. For custom models you really need to know what you're doing.
The main issue I've had is that the compiler behaviour differs between versions (and it's very difficult to find older releases), so where previously you could run a big model and delegate things to the CPU, now it sometimes won't compile at all. There were also problems where we trained a model in AutoML - using free credits but the real cost would have been over $100 - and the edgetpu compiled model lost a lot of performance for no apparent reason. The developers have been very helpful when I've contacted them, and generally you can get through to real devs (not generic support) who can look at your model for you. Mostly I think you need to take care when training models for these devices, but quantisation-aware training is not trivial to use in Tensorflow and there are only a few off-the-shelf models which are supported in the various toolkits. Model maker looks promising, but it's also finnicky in my experience [1]. I guess that's my main caution: it's often very difficult to figure out why your model performance goes to crap when you compile it, especially if the input TFLite model is fine.
I'm not super worried about hardware availability. They're suffering from the chip shortage like everyone else, so it's not surprising that lead times are long. I was able to buy my device in late 2020 without any trouble.
They also make a nice little environmental sensor + OLED board that's compatible with Pi headers and is the same form factor as the Pi Zero. A bit odd, because on the Coral dev board the heatsink is squarely in the way and you have to buy an extension header for it.
I'm not terribly familiar with the GPU/TPU market landscape. I know you can get tensor cores in Nvidia chips, but are there no other options on the market that provide tensor cores only?
39 comments
[ 2.9 ms ] story [ 96.2 ms ] threadI think it is impressive cramming an ML model into something small like this, wonder if you could go as far as VIO.
Random thought: you attach one to the front of a football, you throw it and it can steer itself through trees (servo tail).
4 TOPS with 2 TOPS/W is also interesting, this thing has 2W TDP!?
That's 400mA at 5V, seems reasonable to me.
I bought their m.2 dual tpu and the drivers are in a really bad state
They got in the linux kernel, staging, then basically no work done. They continued the development off-tree, going out of sync with what the kernel had. Only 4.19 supported if I remember correctly.
Finally removed from the kernel in...5.13?
Driver wise I can't say that this company is trustworthy
Yes, STM, XYZ, CabbgePi, etc., is better, faster, has sata, has emmc storage, has more pins, less power usage, has this and that, but if I can't program it now, because of software bugs, and won't be able to program in in 5 years, because the toolchain will be abandoned by now, I really don't need it.
source: I have a bunch of microcontrollers and single-board computers, collecting dust in a drawer at work.
Eventually I just sold it, and back when you could buy them, picked up a couple more USB Corals. Works well enough, but it’s ugly.
Also at first glance, the docs and software support are questionable at best.
I've been trying to find the Coral USB Accelerator[0] for a Frigate installation for a few years now. Mouser Europe has a lead time of ... 81 weeks! Other distributors just say "nope".
[0] https://coral.ai/products/accelerator
I've been trying to get one for ~1yr but it's not happening any time soon.
It's not a complaint, but I'm surprised they're releasing new hardware soon when people can't even get the hardware they have now.
I don't need it _that_ much. =)
Camera is the same and i wanted to compare TOPS, but it's not listed on the Pico4ML specification page.
https://coral.ai/docs/edgetpu/benchmarks/
I'm wondering about the price though. The ArduCam is only 35$.
Ah yes, something that will be discontinued 3 months from now.
No thanks!
I was also faintly surprised to learn that the Paralella still exists! I always wanted one to play with and I thought they discontinued them.
disclaimer: I sell these :)
The price is probably something unreasonable, with a high minimum order quantity or high lead time - because that's typically what happens when you see "Ask for Price" or "Request a Quote" (folks who do that often want to reduce your ability to comparison shop easily).
There are 3 configurations on the webpage: Basic, Standard, Advanced (https://rapidlab.io/raco-edge-ai-gateway/).
The prices are respectively: 219, 309, 409 USD + shipping
For 1, 2 or 3 units the lead time is 2-3 weeks. For more, I would have to ask my engineers.
The main issue I've had is that the compiler behaviour differs between versions (and it's very difficult to find older releases), so where previously you could run a big model and delegate things to the CPU, now it sometimes won't compile at all. There were also problems where we trained a model in AutoML - using free credits but the real cost would have been over $100 - and the edgetpu compiled model lost a lot of performance for no apparent reason. The developers have been very helpful when I've contacted them, and generally you can get through to real devs (not generic support) who can look at your model for you. Mostly I think you need to take care when training models for these devices, but quantisation-aware training is not trivial to use in Tensorflow and there are only a few off-the-shelf models which are supported in the various toolkits. Model maker looks promising, but it's also finnicky in my experience [1]. I guess that's my main caution: it's often very difficult to figure out why your model performance goes to crap when you compile it, especially if the input TFLite model is fine.
I'm not super worried about hardware availability. They're suffering from the chip shortage like everyone else, so it's not surprising that lead times are long. I was able to buy my device in late 2020 without any trouble.
They also make a nice little environmental sensor + OLED board that's compatible with Pi headers and is the same form factor as the Pi Zero. A bit odd, because on the Coral dev board the heatsink is squarely in the way and you have to buy an extension header for it.
[0] https://github.com/jveitchmichaelis/edgetpu-yolo/blob/main/h...
[1] https://www.tensorflow.org/lite/guide/model_maker
^Remember that time when Intel released an OpenVINO installer that clobbered any existing OpenCV build on your system?
Er... Sure you can use it for robots, but that device seems pretty well suited to to do covert face/voice recognition...
>NXP i.MX RT1176 (Cortex-M7 and Cortex-M4)
An cortex M that has dual GBit Ethernet and is usable with at least FreeRTOS and Zephyr.