WEBVTT 00:00:00.320 --> 00:00:07.240 this might look like a GPU smell like a 00:00:03.600 --> 00:00:10.480 GPU stick in a PCI Express 16x slot like 00:00:07.240 --> 00:00:12.679 a GPU and even talk like a GPU look at 00:00:10.480 --> 00:00:18.400 me I'm a little GPU but I assure you its intended purpose is very different and 00:00:15.280 --> 00:00:22.359 to some potentially terrifying it's 00:00:18.400 --> 00:00:26.920 dubbed the ASUS iot AI accelerator PCIe 00:00:22.359 --> 00:00:28.920 card or crl dg16 new- P3 DF for short 00:00:26.920 --> 00:00:33.240 and it's designed for artificial intelligence comput ation 00:00:39.960 --> 00:00:44.840 ooh historically we've mostly shied away 00:00:42.960 --> 00:00:49.879 from covering AI stuff and there's a couple of good reasons for it one is 00:00:47.760 --> 00:00:55.440 we're not AI or machine learning developers so finding practical use 00:00:53.000 --> 00:01:01.640 cases that we are able to set up and demo for you guys can be a little tricky 00:00:58.199 --> 00:01:04.439 and second a lot of those use cases are 00:01:01.640 --> 00:01:09.640 extremely technical behind the scenes kinds of things at least for now so 00:01:07.119 --> 00:01:15.119 Hardware like this isn't too applicable to end users like me and you that is 00:01:12.200 --> 00:01:18.920 until today while doing research for and testing the home automation setup at my 00:01:17.200 --> 00:01:22.240 new house get subscribed by the way you don't want to miss any future content 00:01:20.520 --> 00:01:26.360 around revamping that place we stumbled into a bit of a roadblock regarding 00:01:24.119 --> 00:01:30.119 presence detection or the ability for the systems in the house to be aware of 00:01:28.280 --> 00:01:34.360 whether or not anyone is actually home presence detection provides us with a 00:01:32.000 --> 00:01:38.520 number of benefits for example we can turn off lights and set the HVAC to be 00:01:36.840 --> 00:01:42.759 more energy efficient when there's nobody at home and this is relatively 00:01:41.000 --> 00:01:48.360 simple to implement on a whole house level but therein lies the challenge we 00:01:45.960 --> 00:01:52.119 put a lot of effort into making my house a lot more granularly controllable 00:01:50.520 --> 00:01:56.640 practically every room has its own independent HVAC and lighting which is 00:01:54.439 --> 00:02:00.640 awesome for personal preferences like the kids might want their rooms cooler 00:01:58.680 --> 00:02:05.479 than my wife and I do for example or vice versa but it can also be used to 00:02:03.039 --> 00:02:09.640 improve efficiency and save cost if no one's in the giant wreck room downstairs 00:02:08.039 --> 00:02:14.760 what's the point of heating and cooling it I mean surely that can't be that hard 00:02:11.680 --> 00:02:16.920 to automate well as we found out it can 00:02:14.760 --> 00:02:20.959 be really freaking hard one idea we had was to install Bluetooth low energy 00:02:18.640 --> 00:02:24.120 beacons in every room and read people's presence through their phones the only 00:02:22.720 --> 00:02:28.239 problem with that is that people have this tendency when they're at home to 00:02:26.280 --> 00:02:32.319 put their phone on a charger and oh I don't know walk around their house so 00:02:30.280 --> 00:02:37.239 that won't work another idea was to use motion sensors but those kind of crap 00:02:35.640 --> 00:02:41.560 the bed as soon as you decide to take a nap on the couch and then you wake up in 00:02:39.480 --> 00:02:47.760 a sauna of a room in the summer which sounds like a pretty bad time to me so 00:02:44.720 --> 00:02:50.280 this is where our AI card comes into 00:02:47.760 --> 00:02:54.440 play this is an ASUS designed product but all of the important AI bits are 00:02:52.640 --> 00:02:58.000 actually made by Google of all people under the brand Coral we've got both of 00:02:56.720 --> 00:03:03.640 them linked down in the description below let's take this thing apart this 00:02:59.879 --> 00:03:05.799 specific card has 16 onboard Edge tpus 00:03:03.640 --> 00:03:10.560 as Google calls them and the TPU stands for tensor Processing Unit in the most 00:03:08.360 --> 00:03:14.720 simple terms these tpus are Hardware processors that are physically designed 00:03:12.599 --> 00:03:18.200 and optimized to run a specific application you may have heard of 00:03:16.120 --> 00:03:22.519 something called an Asic as something that's really good at mining as specific 00:03:20.200 --> 00:03:26.200 cryptocurrency well these are the same idea except that in instead of mining 00:03:24.560 --> 00:03:31.680 Bitcoin they're for running machine learning inference calculations so what 00:03:28.879 --> 00:03:37.840 we're looking at here is essentially just a PCI Express 16x interface here 00:03:35.560 --> 00:03:43.560 going you can actually see the traces going right into this PCI Express switch 00:03:40.439 --> 00:03:45.879 which appears to be just splitting out 00:03:43.560 --> 00:03:52.920 these Lanes into an m.2 2x interface for each one of these 00:03:49.480 --> 00:03:56.040 little dual tpus here huh I I was 00:03:52.920 --> 00:03:58.720 expecting something less kind of modular 00:03:56.040 --> 00:04:04.599 and science fair projecty now each of the EDG tpus of which there are two on 00:04:01.480 --> 00:04:06.640 these m.2 cards can do around 4 trillion 00:04:04.599 --> 00:04:12.959 operations per second with a power draw of only two Watts so 2 Watts * 16 tpus 00:04:10.720 --> 00:04:18.239 brings us to 32 watts for the card you add in your PCIe switch and fans and 00:04:15.200 --> 00:04:21.120 that brings you to 52 Watts which is 00:04:18.239 --> 00:04:25.639 well within the 75 Watts that this PCI Express slot could power but for some 00:04:23.320 --> 00:04:30.880 reason it still requires a six pin power connector and it's got this gigantic 00:04:28.720 --> 00:04:36.320 copper fin cooler on it now you're probably thinking wow that's really cool 00:04:33.759 --> 00:04:39.440 Linus but most of this sounds like the boring technical stuff that you were 00:04:37.720 --> 00:04:42.720 talking about earlier but if you were following closely while we were 00:04:41.199 --> 00:04:47.919 researching better ways to detect presence in my new house we stumbled 00:04:44.800 --> 00:04:52.199 upon some software called frig that uses 00:04:47.919 --> 00:04:54.759 cameras plus this Hardware to really 00:04:52.199 --> 00:04:59.280 efficiently detect presence it's really freaking cool let's take it for a test 00:04:56.639 --> 00:05:02.400 drive I got to put this back together and we're up this test bench is running 00:05:00.960 --> 00:05:08.800 the latest version of UNR and it's going to stand in for my home NZ now you might 00:05:05.320 --> 00:05:11.360 notice there is also a GPU installed in 00:05:08.800 --> 00:05:16.360 it so we need that as a video output because this is a non Apu ryzen 00:05:14.000 --> 00:05:21.240 processor but it's also useful for offloading the camera video decode 00:05:18.560 --> 00:05:25.680 process from the CPU now in my home deployment I'll probably just use the 00:05:23.280 --> 00:05:29.319 CPU because I'm going to have 24 epic cores that are otherwise going to go 00:05:27.039 --> 00:05:31.840 unused but for now this is the better solution 00:05:30.160 --> 00:05:35.960 now once un raid's booted up you can see that we have just a single array disc 00:05:34.319 --> 00:05:39.600 with no redundancy that's not what I'd recommend as a production deployment but 00:05:38.080 --> 00:05:43.800 it's good enough for this video now we just pop over to the community apps Tab 00:05:41.720 --> 00:05:48.440 and installed apps so we've got our Coral accelerator module drivers we 00:05:46.240 --> 00:05:52.680 don't need this if we're using a USB version of Google's AI processor we only 00:05:50.479 --> 00:05:56.919 need it for the PCI Express version and then we've also got our NVIDIA drivers 00:05:54.639 --> 00:06:03.280 which are for our GPU now it should be noted that you can use your CPU for frig 00:05:59.960 --> 00:06:05.680 but performance is not good next we 00:06:03.280 --> 00:06:09.639 install frig now we can't just launch it we've got to actually configure frig 00:06:07.680 --> 00:06:14.080 otherwise it's going to have no idea what devices it's supposed to be using 00:06:11.240 --> 00:06:19.479 so step one we're going to copy the GPU uu ID from our NVIDIA driver and paste 00:06:16.639 --> 00:06:27.479 that there then for our TPU mapping we've mapped it to our Apex or0 device 00:06:23.840 --> 00:06:29.280 and that is only one of the 16 tpus that 00:06:27.479 --> 00:06:32.840 are on that card because we're only going to have a couple of cameras we're 00:06:31.240 --> 00:06:38.199 not going to bother mapping all of them but we did map five of them so that that 00:06:36.560 --> 00:06:43.880 should be enough for our demo for now wow this is a great password Jake I like 00:06:39.880 --> 00:06:45.360 it I see I don't get overwhelmed most of 00:06:43.880 --> 00:06:49.280 this is included in the template but there are a couple things that we had to 00:06:46.880 --> 00:06:53.400 tweak so one is we had to set up our camera so you can see we've pulled our 00:06:50.919 --> 00:06:57.319 camera from The Writer's den and we've got our network path for it right here 00:06:55.520 --> 00:07:01.639 we're just using the one camera for now we actually didn't take all five of our 00:06:59.800 --> 00:07:05.080 npus we've just got four of them configured here again this is total 00:07:03.360 --> 00:07:09.759 overkill for what we're doing and then we've also got this right here so we are 00:07:07.319 --> 00:07:16.080 limiting our frames per second for the camera feed that we're using to 5 FPS 00:07:13.280 --> 00:07:20.280 that is all we're going to need in 99% of situations to have functional 00:07:17.520 --> 00:07:24.360 detection and adding more processing to this would serve no real purpose one 00:07:22.759 --> 00:07:28.360 other critical thing back up at the top here is that you can Plum this into an 00:07:25.879 --> 00:07:33.080 mqtt server which is basically its way of communicating back to your other home 00:07:30.360 --> 00:07:39.080 automation devices So This mqtt Server 10.2.0 71 is running on our office home 00:07:36.199 --> 00:07:42.520 assistant instance hey 5 FPS of green I can see why you said we didn't need more 00:07:40.479 --> 00:07:47.759 FPS there's not a whole lot of movement there it's peanut butter demo time 00:07:45.400 --> 00:07:51.800 peanut butter demo time peanut butter demo time where you at where you at so 00:07:50.520 --> 00:07:56.680 that one camera is set up we have another one too but let's just okay hi 00:07:54.240 --> 00:08:01.440 there oh wait hold on I screwed up oh all right do we not have the detection 00:07:58.560 --> 00:08:07.039 on uh we but I I didn't turn on like the the bounding boxes bounding boxes are 00:08:03.680 --> 00:08:08.960 just little boxes that show us what the 00:08:07.039 --> 00:08:12.879 software is detecting and the degree of certainty that it has for what that 00:08:10.800 --> 00:08:16.680 object is humans aren't the only thing that it can detect but they're kind of 00:08:14.800 --> 00:08:18.960 the only thing that we care about for home automation presence detection 00:08:18.159 --> 00:08:24.599 you're telling me we don't have to detect how many apples you have in your fruit bowl oh yeah oh yeah oh you were 00:08:23.240 --> 00:08:28.440 there buddy oh it's kind of detecting your feet does it know who I am that 00:08:26.720 --> 00:08:34.959 part we have not done yet what if it only sees like this of me yeah it still 00:08:31.919 --> 00:08:37.640 thinks you're there what if my legs show 00:08:34.959 --> 00:08:42.719 oh it lost you it lost me what about a hand oh yeah it sees you again no really 00:08:40.640 --> 00:08:47.959 okay what if it just saw an elbow oh yeah it's detecting you what really yeah 00:08:45.760 --> 00:08:51.640 okay can it can it see me it's confused right now it's a little confused what 00:08:49.760 --> 00:08:55.640 about if there's just like a casual like hand there it's kind of freaking out so 00:08:53.600 --> 00:08:59.120 sometimes it actually sees two people um when he's kind of obscured but that's 00:08:57.320 --> 00:09:04.320 totally fine oh like it thinks I might be like two people having a sex orgy 00:09:01.279 --> 00:09:05.560 behind the curtain uh yeah sure but it's 00:09:04.320 --> 00:09:11.399 fine because all we need to know is if there's anyone right right yeah and sex 00:09:09.600 --> 00:09:15.399 orgy aside you don't want the heat to cut out during your sex orgy every get 00:09:13.560 --> 00:09:20.120 cold man you can do so much cool stuff with this but wait there's more what if 00:09:18.360 --> 00:09:24.040 you have a common area like a kitchen or a recck room and you want to be able to 00:09:21.640 --> 00:09:27.200 apply personalized preferences when you're present like say have your 00:09:25.640 --> 00:09:31.320 Spotify music with your specific playlist play when you're in the room 00:09:29.120 --> 00:09:35.200 and not when ivonne's in the room right oh I never even thought of that you 00:09:32.959 --> 00:09:39.959 could walk around the house and have Spotify playing here oh I'm over here 00:09:37.760 --> 00:09:43.640 now and now there's my Spotify over here oh my God that would be so cool turns 00:09:42.160 --> 00:09:47.560 out someone had that exact thought and created Double Take which is a piece of 00:09:45.440 --> 00:09:53.800 software that allows you to determine the face this is more 00:09:49.600 --> 00:09:55.680 creepy yeah but it's all local I mean we 00:09:53.800 --> 00:10:01.560 could just configure this to not log any of this right yeah did a Linus appear oh 00:09:58.760 --> 00:10:05.680 wow yeah nice now Double Take isn't available on the unraid community app 00:10:03.040 --> 00:10:09.160 store at least not yet so we need to manually set up the docker container 00:10:07.079 --> 00:10:14.120 ourselves then once it's set up we need to configure it to listen to frig it 00:10:11.480 --> 00:10:18.360 over mqtt it should be noted that double take itself doesn't do facial 00:10:16.160 --> 00:10:22.959 recognition it's just an API and interface to make hooking into and 00:10:21.120 --> 00:10:27.480 training facial recognition software easier so the first thing we need to do 00:10:25.320 --> 00:10:31.040 is grab some pictures of me which are going to be training data for our power 00:10:29.560 --> 00:10:34.920 facial recognition so I guess I need to go over there right uh we could go over 00:10:33.399 --> 00:10:40.600 there or we can just upload some photos we already have Double T currently supports deep stack compre face and face 00:10:38.880 --> 00:10:43.800 box they're all machine learning based facial recognition models deep stack is 00:10:42.240 --> 00:10:46.920 on the community App Store so that's the easiest to set up but I tried it out and 00:10:45.480 --> 00:10:50.760 it wasn't really that great with the limited training data I had compri face 00:10:48.800 --> 00:10:55.160 on the other hand it requires Docker compos have to run in a VM but it's 00:10:52.760 --> 00:10:59.320 scary good I put a couple photos of lonus and secretly kind of had him walk 00:10:57.440 --> 00:11:03.560 by without telling him and it was like lus and then I walked by it was like 00:11:00.760 --> 00:11:07.920 Jake with like 10 photos of each of us all right let's try it why do you have 00:11:05.160 --> 00:11:14.200 so many pictures of me on your hard drive well you make my my drive hard why 00:11:11.560 --> 00:11:19.680 CPU when I can see you pe why talk to you when I can leave can I zoom how do I 00:11:17.040 --> 00:11:24.880 use a Mac do you know no these Stills are horrible it's scary good Lord to see 00:11:22.399 --> 00:11:32.680 you without a beard oh wow yeah yeah yeah oh my God 95 00:11:29.920 --> 00:11:37.720 5.8% how can it be that sure on unbearded Linus training there was only 00:11:35.279 --> 00:11:43.480 one that it is actually confident about I mean 94% seems pretty damn confident 00:11:41.000 --> 00:11:48.040 to me we can do is take all of these yeah and say train off of those data now 00:11:46.720 --> 00:11:52.399 you want to walk over there again yeah for sure okay that's plenty there's so 00:11:49.920 --> 00:11:55.160 many luses now Linus Linus Linus Linus there's a couple times it thinks it's 00:11:53.360 --> 00:11:58.760 James I think like once you trained a good model with compr face I'd probably 00:11:56.839 --> 00:12:02.000 say that like it would only accept it as as that person person unless the 00:11:59.839 --> 00:12:07.399 confidence was like 99% some of these other models Madison I also made with 00:12:05.360 --> 00:12:10.760 people's photos with them wearing a mask so it like literally has the eyes okay 00:12:09.279 --> 00:12:18.519 go back run over there all right I'm going I'm going I'm going see if it makes more mistakes I'm going to pull up 00:12:14.880 --> 00:12:21.519 a chair lonus style it's very confident 00:12:18.519 --> 00:12:23.399 about lonus now what who else would 00:12:21.519 --> 00:12:26.959 behave like this I don't I don't think that's what it's looking for but man so 00:12:25.040 --> 00:12:31.199 you could just kind of wander around your house for 10 minutes find any 00:12:29.199 --> 00:12:34.839 anything that's wrong boom train it part of it is also going to be specific 00:12:33.120 --> 00:12:37.839 cameras because you know each camera looks a little different one thing I 00:12:36.399 --> 00:12:42.800 asked Jake about and he said it doesn't have this functionality for now is can 00:12:39.959 --> 00:12:46.800 it morph over time and can it continue to self-train the thing with facial 00:12:44.639 --> 00:12:50.760 recognition is like the more data you feed it the more uncertain it becomes 00:12:49.279 --> 00:12:56.040 all of this leaves us with some questions do you even need this thing 00:12:53.920 --> 00:13:02.480 could a bog standard CPU do the same calculations the answer is yes it could 00:12:59.440 --> 00:13:04.240 but way slower to put it in perspective 00:13:02.480 --> 00:13:09.800 your average quad core CPU would probably only be able to handle a few 00:13:06.240 --> 00:13:13.000 frames per second on a quad core but 00:13:09.800 --> 00:13:15.320 because these Edge tpus are so optimized 00:13:13.000 --> 00:13:21.440 for running these specific calculations they are stupid fast at them a single 00:13:18.680 --> 00:13:27.560 Coral Edge TPU can handle around 100 frames per second of people detection in 00:13:23.959 --> 00:13:30.199 frig and we've got 16 of them so if my 00:13:27.560 --> 00:13:33.639 rough math is correct that that's around 1,600 frames per second we might 00:13:32.240 --> 00:13:38.800 actually be pushing that once we've got all the cameras in the new place we're only taking 5 FPS from each still it's a 00:13:36.720 --> 00:13:43.279 lot of cameras No No 5 FPS yeah five okay yeah man that's crazy so a single 00:13:41.240 --> 00:13:48.320 one of these tpus could probably handle 15 to 20 cameras at 5 FPS another 00:13:46.519 --> 00:13:53.120 question this raises is are we worried about the security implications 00:13:50.560 --> 00:13:59.720 fortunately since we have all of our own hardware and all of this is open source 00:13:56.320 --> 00:14:01.440 and locally hosted none the data goes 00:13:59.720 --> 00:14:06.519 outside of my house so I'm actually pretty happy with the level of security 00:14:03.360 --> 00:14:08.639 and safety as for being worried about AI 00:14:06.519 --> 00:14:14.240 taking over no one's asking that question it's in here anyway the answer 00:14:11.279 --> 00:14:19.440 is no leaving only one question then who's our sponsor freshbooks freshbooks 00:14:17.600 --> 00:14:23.199 is the simple accounting solution that is designed specifically with you in 00:14:21.240 --> 00:14:27.000 mind the small business owner it features built-in automation that allows 00:14:24.800 --> 00:14:30.680 you to spend less time tracking projects and more time growing your business so 00:14:28.800 --> 00:14:34.800 whether you're a trades person creative agency or a YouTuber you can choose the 00:14:33.440 --> 00:14:38.240 plan that's right for you with freshbooks they have an award-winning 00:14:36.440 --> 00:14:41.800 Toronto based support team who are always happy to help you if you need it 00:14:40.240 --> 00:14:45.800 and you don't have to take my word for it you can try fresh books for free for 00:14:43.720 --> 00:14:50.959 30 days today with no credit card required at freshbooks.com 00:14:48.399 --> 00:14:53.880 Linus if you enjoyed this video maybe go check out the video where we built the 00:14:52.360 --> 00:15:00.519 server that is going to be running all this stuff we're going to have to get a little bit creative with PCI Express 00:14:57.639 --> 00:15:04.240 expansion on it um but that's the machine that's going to somehow run this 00:15:01.959 --> 00:15:07.759 full height card yeah yeah so good luck everybody are we going to do an external 00:15:05.600 --> 00:15:10.560 PCI box we should totally do it yeah we should totally do it yeah that'll be a 00:15:08.800 --> 00:15:13.240 video get subscribed we can water cool it and everything