• expired

Google Coral USB Accelerator $99.43 Delivered @ RS Components Australia

560

Edit: Price update from $100.36 to $99.43

Similar to previous deals around the Coral USB, was keeping an eye on these for use in a Frigate setup. Cheapest I've seen anywhere.

Stock report shows over 800 in stock for 10 working day delivery. Price includes GST.

Page says for use with Raspberry Pi, but these are compatible with Windows, Mac and Linux, just have to be running appropriate software.

Unfortunately the Coral is excluded from the 15% off Your First Purchase:
https://www.ozbargain.com.au/node/653408

Product details as per webpage:

The Coral USB accelerator by Google, allows for an edge TPU coprocessor to be added to any system, enabling high-speed machine learning inferencing. It includes a USB-c socket to connect to a host computer to perform accelerated ML inferencing. The on-board Edge TPU is a small ASIC designed by Google that accelerates TensorFlow Lite models in a power efficient manner: it's capable of performing 4 trillion operations per second (4 TOPS), using 2 watts of power—that's 2 TOPS per watt.

Related Stores

RS Components
RS Components

closed Comments

  • +3

    I've been reading lately that the Intel iGPU's (7th Gen+?) are now as good if not better than the Coral for Frigate detection.

    I'll continue to use mine as I already use the iGPU for Plex transcoding but it seems they are not the only option now.

      • +41

        silence, jv

        • +9

          If only

          • +14

            @iMagoo: This site improves measurably once you block them.

      • +1

        Hence the worldwide shortage on these for years. Dual use civilian/military. I have no doubt autonomous drones are running these or similar for object detection.

      • +1

        With a properly trained model, I don't see why not!

      • Funniest jv comment

    • My understanding of it is that GPU is for processing the image itself, eg encoding/decoding, but for actual detection of objects a TPU (as within the Coral) is significantly more efficient and accurate.

      I haven't looked too much into iGPUs in comparison as my setup is a 5000 series Ryzen, so I don't even have the option.

      • +2

        Frigate has had the option of GPU for detection (OpenVINO for AMD and Intel, TensorRT for Nvidia) for some time but in terms of efficiency the coral TPU is the winner.

        • +1

          OpenVINO works pretty well, on an N100/N105 with its tiny 15W TDP and the UHD 605 iGPU, it may be a close call.

          The coral probably has lower latency and power draw overall, but on the flipside, the models for openVINO will continue to get better and better, and not be constrained to what the original Coral could detect.

          • @Commodore64: N100 is what I am soon to try.. and hoping to manage sans coral

      • GPUs, particularly NVidia, support the CUDA library which is used for a wide range of applications including machine learning. The NVidia GPU is typically going to be more efficient (and much faster) at training. For inference, efficiency will depend on which GPU the comparison is being made against. For light inference workloads, a Coral is going to be much more energy efficient.

    • OpenVino is pretty good, certainly a LOT better than the CPU options a few years back, but I still find it misses a lot compared to the Coral. I've tried a few different models and things with OpenVino, but it still wasn't as good as with the dedicated accelerator.

      YMMV of course.

      • I may switch SSDLite MobileNet V2 over to the YOLO-NAS-S model when Frigate 0.14 goes to release. Will likely still require you to export the model from a collab notebook though, due to license issues with the weights, but at that's a couple minute process at best. In theory the mAP is pretty decent while keeping inference time reasonable.

        • I found YOLO to be noticeably better than MobileNet, though as you say it's a little messing around to do the switch.

          Even YOLO wasn't quite as good as the Coral though.

    • +2

      Nah, Coral is still better. More importantly though it takes some of the load off the CPU/GPU

    • Yes I was all geared up to get one from having read up on requirements for Frigate, when Iearned this is the latest situation. Also these are surprisingly old, I'd have thought some newer tech would have come along by now

    • If you use the iGPU the CPU usage will be much higher and more importantly it can't match the inference time of a Coral (obviously the USB version is the worst of the bunch but can still muster a respectable 8ms or so).

  • +14

    Bought. Now can someone tell me how to use it to enhance my ozbargain experience via my Windows laptop

    • +35

      format c: /fs:NTFS

      • +44

        Settle down Satan, let the man use /quick at least

      • Nooo….

      • +7

        I was thinking more like C:\Users\Ozbargain buynowthinklater /buyeverything

        Or

        C:\Users\jv comments /blockall

      • Tried this now my pc wont boot, tried turning off and on again but still no go, what do I do now?

        I heard running magnets over the hard drives might reset them to their previous state?

        • Only in the good old days of magnetic media with spindles, heads, and RPMs. When wear levelling was not a thing, but defragmentation was an important part of our lives and the consistent fear of the sudden death of a spindle kept us up at night.

        • +1

          sorry grandpa you'll have to go buy some walmart gift cards now

    • +4

      If you have one connected, JV's comments show up in a level of bold that machine learning AI has developed but computer scientists are still baffled by.

    • +2

      First of all, get off Windows. Install debian

      then open terminal and enter sudo rm -rf /*

  • +7

    Hoping I could drop this in the reefs and it would help accelerate coral growth.

    😢

    • Too late, crushed by tourist foot

      • The corals or the usb? 😋

  • Is it still a bargain when it is 2 x RRP?

    • What's RRP?

    • +2

      The only RRP from coral.ai is $59.99USD.

      Current conversion brings it ~$89.93AUD +GST = $98.93AUD.

      That's before any other considerations like shipping.

      • They used to be USD$30. Dual TPU m2. was USD$60

        • +1

          I could be looking at the wrong sku that you're referring to, but this one was released in 2019 at a price of $75USD.

          • +1

            @AlphaDeal: Yup. He must be thinking of the M.2 models.

        • Yep, the USB version was never cheap

        • USB $ always >> M2

  • +2

    Sweet I have frigate/linux running on a refurb micro pc (also found on ozb), keen to try this out.

  • See a M.2 version there has dual Edge TPUs, …and cheaper? 🤷 But absent the nice USB-C enclosure to act as heatsink, if they need it?

    • Oh, that makes more sense to the costs mentioned by Bruceflix.

      Didn't know they had an M.2 chip TPU. Obviously that would suit better if you have the I/O for it.

      Too bad you can't put PCIe chips into a USB enclosure…

      • +1

        Should bear in mind the dual tpu will almost definitely not work with both tpu's in any pc as the PCIE lanes generally don't support bifurcation and you will need to purchase a 3rd party pcie or nvme adapter

        PCIE adapter
        https://www.makerfabs.com/dual-edge-tpu-adapter.html

        Details on adapters
        https://github.com/magic-blue-smoke/Dual-Edge-TPU-Adapter

        And then there is a possibility that it still won't work on both TPUs

        This is my situation after getting the pcie adapter and a warranty request to get a second, either my coral is faulty or it just doesn't work for my pc
        I have a support request in with Google but as others have said its abandon ware and I'm yet to even get an acknowledgement

        the one lane is enough for 8 cameras atm though

        • +1

          I have the makerfabs adapter and can confirm that both TPUs work. Had to add custom passive cooling with heatsink otherwise it will overheat and shutdown quickly.

        • I can likewise confirm the makerfab adapter enables both TPU.

        • Any Dual Edge TPU adaptor to USB/USB-C?

    • +1

      Yeah that'll work, but they only come in E-Key 2230. (Well the dual one only comes in E-key, I think the single comes in other keyouts like A/M if you pick the right SKU)

      Depending on your hardware, you may require an adapter board, or replace your wifi card with the Coral.

      As for heat sinks, the datasheet gives you exact dimensions for both the power and TPU chips, so you can easily source the correct size heatsinks. Although I suspect just a single generic 2230 copper heatsink may be easier than 4 individual smaller ones.

      I think you'd need quite a few cameras to need two though.

      Here's a bunch of people giving their input on various adapters as well. Note that some will have issues having both TPUs detect on the system. https://github.com/google-coral/edgetpu/issues/256

    • +1

      For memory, that one has one TPU on each of 2x lanes in the m.2 E key. Since that slot is commonly used for wifi cards which only typically use 1 lane, most motherboards only have one slot wired for 1 lane. If your motherboard only has one lane on that slot, then you'll only be able to use one of the two TPUs

  • can I use this to accelerate OpenVINO?

  • +9

    "WARNING
    Coral EdgeTPU devices are not recommended. While these TPUs were fast on initial release in 2020, they are now outclassed by both modern Intel and Apple Silicon chips. Furthermore, the aging chips have not seen a single hardware or software refresh. The software (Tensorflow) is no longer in use at Google. For all intents and purposes, the Coral EdgeTPU is now Google abandonware."

    (https://docs.scrypted.app/buyers-guide/servers.html)

    Doesn't mean you won't get value out if it, but I guess consider this.

    • +1

      The only issue I have with that assessment is those other options aren't drop-in replacements for the USB.

      I don't disagree those other options are better overall as most of the 'value' of the Coral is the fact you can plug it into a existing systems easily.

      • I don't disagree those other options are better overall as most of the 'value' of the Coral is the fact you can plug it into a existing systems easily.

        Yeah, was giving the Hailo 8L that comes in the new Raspberry Pi AI Kit a good look, but will be waiting to see if people start picking it up software-wise since it requires models to be converted to their format. (Although the vendor hosts a bunch of different pre-trained models)

        (That and the units are somewhat annoying to acquire themselves, as you either will have to salvage a 8L (Lite) from a Raspberry Pi AI Kit, or reach out to Hailo themselves to purchase the Hailo 8 starter kit.)

    • +3

      Your reference is scrypted, not Frigate. They use different methods of detection

    • +10

      Everything Google is abandonware

  • Thanks OP, good find.

    • RS very slow to despatch.

      • RS dispatched my two m.2 [Dual Edge TPU] on the same day as order. They arrived unexpectedly within two days to Perth, WA - been sent from an eastern state. (Toll Priority)

        As for the Makerfabs adapters, were dispatched next day, and only as of today: "PER / Your parcel has been picked & packed at the sender's warehouse. we'll update you when your parcel is on its way to the delivery destination." (4PX Tracking)

  • Pity it can't be used to mine crypto

    • Get a pi for that.

  • +1

    Why do I want one of these?

    • +9

      Cos it's a bargain.

      You can hand in your badge on way out.

  • just get the pcie one for a lot less

  • +1

    What is this? Why do I need it? Will it make my life better? Thx.

  • Can get a dual TPU M2 card and adaptor for about that price, but yeah USB is definitely more convenient.

  • +1

    Got 1

    What do i use it for

  • Keep in mind these can only perform int8 inference. If you are after a particular model make sure you can in fact quantise the model into int8 precision without sacrificing accuracy. I suppose the most common usage is for cctv hence it makes sense ..

  • -1

    Used this with frigate in ha
    Then found scrypted very very inpressed
    Coral works fine
    Deleted frigate

    • Ahh man, don't tell me that now. Just finished setting up frigate in docker and integrated it with ha
      Took a bloody long time to figure out all the options etc

      You really like scrypted that much more than frigate?
      Can it do the ai stuff like frigate?
      Would a coral still be useful?

      Cheers mate

      • Frigate is free.
        Scripted is subscription based.
        Payed influencers are going to preach. :)

    • Just had a read through of the scrypted docs. Very interested. Will have a play around with this. Frigate, although works fine for me, seems a lot less polished than this.

      • +2

        The upcoming next major revision (0.14) will have a much improved UI.

    • Scrypted gets very expensive to use if you want the NVR and have a lot of cameras. Kind of prefer the FOSS Frigate.

Login or Join to leave a comment