Khadas VIM3 board based on Amlogic A311D processor with a 5TOPS Neural-network Processing Unit (NPU) launched last June. We’ve reviewed VIM3 with Android 9 shortly after launch, but until recently it was not possible to leverage the NPU since the software was not quite ready yet.
The goods news is that Khadas has now released the NPU toolkit for both VIM3, and the cheaper VIM3L boards.
The NPU toolkit contains the following directory:
- docs – Model conversion documentation
- acuity-toolkit – Model conversion tools
- linux_sdk – Linux SDK
- android_sdk – Android SDK
The toolkit works in host PCs running Ubuntu 16.04 or 18.04 with Tensorflow framework, and inference can run on both Linux and Android OS in Khadas VIM3/3L board. It includes an Inception v3 sample with 299×299 sample photos, among other demos. You’ll find documentation to get started with model conversion and inference in Linux on Khadas Wiki.
You can also watch the video demo below running Yolo v3 object detection model.
The less good news is that the NPU toolkit has licensing requirements from Amlogic and VeriSilicon (NPU Intellectual Property) which requires you to fill-up a form to apply for VIM3/3L NPU Toolkit with your contact details, project description, and order number.
The toolkit release was announced about two weeks ago in the forum, and it seems easy enough to get access to the toolkit.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress
Well, then maybe it’s better to ignore this NPU toolkit entirely and use another device which doesn’t require filling such forms. It’s sad to see there are still IP vendors who want to make money both on the silicon AND the software needed to exploit their silicon. Let them die and encourage their more open competitors instead. The VIM3/3L remain great boards even without using this (I didn’t even know that my VIM3L did have an NPU).
Hi Willy:
Thanks for ordering the VIM3L.
We will sent out the NPU toolkit(DDK model convert tool & documents) to all the users who buy the VIM3/L.
Thanks again!
Hi Gouwa,
you’re welcome and I have no doubt about this. The issue is more related to this amlogic+verisilicon stuff not being open enough if they require some specifial license. It simply means that their SDK will never progress as much as it should, and users constrained to use this solution will start in a dead end while competitors progress 10 times faster. It’s a pity that in 2019 they haven’t understood this yet.
Most of the fools think they’re going to, “take the market,” and own it. No, you need to get the SDK’s in people’s hands, even if you’re shipping only pre-built binaries. The notion that is behind having, “special agreements” in place is to keep track of whom has the stuff and trying to gatekeeper people getting competitive analysis access to their products.
Well, this doesn’t seem to be the case right now, unfortunately. Most of the vendors need to “get” that this stuff is going to be worked on by DIY enthusiasts as well as major corps like my day job. I do this sort and class of stuff after hours with blessing from my employer as long as it’s not in their space. Since this is the case, I’m not going to cheat and leverage a Fortune 500 just to get access, etc. You need to make it generically available, period. You’re not keeping __ANYONE__ from reverse engineering things. Competitors get… Read more »
In light of the fact one can get rather powerful ARM based stuff that’s fully capable of this without a proprietary NPU that works with stock TensorFlow Lite and similar inference models (Jetson Nano…) for the same prices or less, the “Pro” models of the 3399 with the integrated or the outside NPUs that you have to sign all sorts of crap…yeah, no.
So I’ve just tried it on Khadas VIM3L board.
Good news is that there’s no manual check to download the toolkit, you’ll receive an email immediately with Baidu and Mediafire links.
The bad news is that it does not work on VIM3L board just yet. They’ll release support for it later this month. So It’s only working on VIM3 for now.
does it work on VIM3L now?