This is NOT a Graphics Card

[DJAKO]

17 hours ago, daddysmackdown said:

What cameras are being used? 

14 hours ago, jhoglund218 said:

I am also curious about this. Last I remember he was running ubiquity cameras, but I’m not sure if that’s still the case. 

I run all the programs mentioned in this video with Eufy, Wyze v2, and Reolink cameras. As long as your camera can output a RTSP stream, you should be fine.

 

18 hours ago, Hobox said:

How much does something like this cost? Did I miss a mention in the video? Didn't see any rough numbers when googling

I run all the programs mentioned on an Intel NUC8 with 16GB of RAM and I use two Google Coral USB Accelerators with Frigate which take a lot of the load off the NUC. You can definitely run all this on a sub $1,000 machine.

[Kyrle]

Would there be any special considerations for running this in a VM? For example, would you need to do a passthrough for the the TPU?

[GodAtum]

3 hours ago, DJAKO said:

I run all the programs mentioned in this video with Eufy, Wyze v2, and Reolink cameras. As long as your camera can output a RTSP stream, you should be fine.

 

I run all the programs mentioned on an Intel NUC8 with 16GB of RAM and I use two Google Coral USB Accelerators with Frigate which take a lot of the load off the NUC. You can definitely run all this on a sub $1,000 machine.

Don't you need a GPU as Linus said in the video?

[Egg-Roll]

Linus: me and you

It's "You and I"... When talking about others and yourself 99% of the time the correct way is for you to be last in the sentence.

Typically I don't go crazy over grammar but that for whatever reason trigger me...

 

Also what has ASUS done giving Linus AI  I do want this product however, might be the first ASUS product I'm willing to buy in years

23 hours ago, Radium_Angel said:

(reference: https://dilbert.com/strip/1995-02-22)

 

I thought that sounded familiar lol...

 

  

9 hours ago, Ninja123 said:

As demoed in the video, MANY thousands of dollars. The featured AI accelerator alone is over $1k.

 

Granted, if engineered to a more reasonable price point, you could use the $25 M.2 version in an unpopulated M.2 A+E slot of a spare PC, and have similar results at a fraction of the cost and power consumption.

Uh, the over 1K price includes 8 of those $25 units I'm sure (it's a all in one package), also someone correct me if I'm wrong but all this board actually is, is a glorified PCI-E (actual) 16 lane card splitting the lanes into 2 lanes per slot so each TPU has 1 PCI lane as each M.2 unit has 2 TPUs. If this is the case theoretically can't one just buy some cheapo M.2 addon board?

 

  

2 minutes ago, GodAtum said:

Don't you need a GPU as Linus said in the video?

Nope, Linus said you don't need one as well, just be smarter to use one.

[Kid.Lazer]

22 hours ago, cm360 said:

For HVAC efficiency, a much lower polling rate would be fine - perhaps poll every minute, but turn off HVAC if no person is detected in the last 15 mins.

This. The system also definitely needs some sort of hysteresis. You don't want the HVAC kicking on for a room when the person just walked in for 3 seconds to put their shoes in the closet, never to return.

[TomazP]

Probably a dumb question to some of you - what other OS that unraid you can use? Something of a more free version?

[AndreiArgeanu]

The AI accelerator linus shows in his video seems overpriced. It's 1800$. The M.2 dual Edge TPU's are 40$ at MSRP and there 8 of them in the GPU making it 320$, 128mb of SPI flash which shouldn't cost more than a few dollars and the PCIE switcher which you can buy for about 200$, however it's much cheaper if you buy in bulk so chances are they're actually paying around 150$ a pop. And I can't imagine the cooler, power delivery and other miscellaneous chips and manufcaturing being worth more 200$ tops bringing the estimated manufacturing cost to about 670$ while they're charging nearly 3 times that for it. Is there something I'm missing? Is this just normal?

 

[Ninja123]

2 hours ago, Kyrle said:

Would there be any special considerations for running this in a VM? For example, would you need to do a passthrough for the the TPU?

Depends on whether you're running a full VM or just a container, but for a VM yes, just hardware passthrough of the device.

 

36 minutes ago, GodAtum said:

Don't you need a GPU as Linus said in the video?

You don't need it per se. They were primarily using it to get video output from a non APU, non server board. It can also serve double duty to help decode video from the cameras (same way a GPU or APU is leveraged in Plex for video encoding) without bogging down the CPU.

 

10 minutes ago, TomazP said:

Probably a dumb question to some of you - what other OS that unraid you can use? Something of a more free version?

Unraid just uses Docker underneath, so your favorite linux distro + docker would work well, and is free.

[DJAKO]

1 hour ago, GodAtum said:

Don't you need a GPU as Linus said in the video?

The Edge TPU does a lot of the hard work on the Frigate side for object detection and that can be offloaded to a $60 USB Accelerator. The only other process really running at that point is FFMPEG.

 

I run CompreFace on a 6 year old MacBook Pro and results come back in less than a second. For facial recognition those times work fine for me.

[Kyrle]

5 hours ago, Ninja123 said:

Depends on whether you're running a full VM or just a container, but for a VM yes, just hardware passthrough of the device.

 

It would be a full dedicated VM running in hyper-v. So, basically just passthrough the TPU and proceed as normal? If so, cool! I just found my next project!

[drqi]

I think it would be awesome if LTT published a docker image and config files that could be used to replicate what they did here.
They make a lot of videos that are basically "Look at this cool thing we made". It seems like a way to step it up a notch would be "Look at this cool thing we made, and here's everything you need to make one yourself". Ya, most people wouldn't bother actually doing it, but the way that Level1Techs or Jeff Geerling often publish the source code or detailed instructions makes it feel like the time I spend watching their videos is somehow more productive.

It depends on the subject of the video obviously, but in theory in this case there could be a link in the description to a docker compose file or something that a viewer could get up and running just on a linux desktop with a webcam attached in like 10 minutes without needing a ton of of experience beyond how to use a linux terminal.

[Hexenhammer.]

Guys, im new to AI acceleration, so i have a question: Why this cant be done on  Nvidia GPU? Both 3000 and 2000 series have Tensor cores which are meant exactly for AI acceleration.

[Ninja123]

1 hour ago, drqi said:

I think it would be awesome if LTT published a docker image and config files that could be used to replicate what they did here.
They make a lot of videos that are basically "Look at this cool thing we made". It seems like a way to step it up a notch would be "Look at this cool thing we made, and here's everything you need to make one yourself". Ya, most people wouldn't bother actually doing it, but the way that Level1Techs or Jeff Geerling often publish the source code or detailed instructions makes it feel like the time I spend watching their videos is somehow more productive.

It depends on the subject of the video obviously, but in theory in this case there could be a link in the description to a docker compose file or something that a viewer could get up and running just on a linux desktop with a webcam attached in like 10 minutes without needing a ton of of experience beyond how to use a linux terminal.

LTT didn't post a full writeup, but the docker compose configs are readily available on the respective install sections of Double Take and Frigate:

version: '3.7'

volumes:
  double-take:

services:
  double-take:
    container_name: double-take
    image: jakowenko/double-take
    restart: unless-stopped
    volumes:
      - double-take:/.storage
    ports:
      - 3000:3000

 

version: "3.9"
services:
  frigate:
    container_name: frigate
    privileged: true # this may not be necessary for all setups
    restart: unless-stopped
    image: blakeblackshear/frigate:<specify_version_tag>
    shm_size: "64mb" # update for your cameras based on calculation above
    devices:
      - /dev/bus/usb:/dev/bus/usb # passes the USB Coral, needs to be modified for other versions
      - /dev/dri/renderD128 # for intel hwaccel, needs to be updated for your hardware
    volumes:
      - /etc/localtime:/etc/localtime:ro
      - /path/to/your/config.yml:/config/config.yml:ro
      - /path/to/your/storage:/media/frigate
      - type: tmpfs # Optional: 1GB of memory, reduces SSD/SD Card wear
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - "5000:5000"
      - "1935:1935" # RTMP feeds
    environment:
      FRIGATE_RTSP_PASSWORD: "password"

 

[Ninja123]

28 minutes ago, Hexenhammer. said:

Guys, im new to AI acceleration, so i have a question: Why this cant be done on  Nvidia GPU? Both 3000 and 2000 series have Tensor cores which are meant exactly for AI acceleration.

A GPU could be used in place of a TPU for machine learning, but GPUs are typically more expensive, less power efficient, and slower. For this specific application, there's no sense in wasting a 3000 series GPU when a $40-60 TPU could probably handle the workload at a fraction of the cost and power envelope.

[reptarded]

So besides the application in the video, what else can this be used for?

 

I need to decide whether or not to cancel a drunk purchase...

[DJAKO]

On 10/10/2021 at 2:55 PM, reptarded said:

So besides the application in the video, what else can this be used for?

 

I need to decide whether or not to cancel a drunk purchase...

I run this with my doorbell and garage cameras to know when people are detected and if I want who that person may be. It's allowed me to stop paying for the Nest subscription because I can do the recognition myself!

[Middcore]

[Ninja123]

On 10/10/2021 at 11:55 AM, reptarded said:

So besides the application in the video, what else can this be used for?

 

I need to decide whether or not to cancel a drunk purchase...

AI accelerators like the one featured in the video are designed to accelerate AI and machine learning, usually neural networks or machine vision.

[Flamingi]

Hey @jakkuh_t, did you manage to set this up with multiple TPUs? I'm currently doing that, and whenever I define more than one in the Frigate config, it will crash. And it's only for the last one it tries to detect, so if I put 2/3/4 there, it will find 1/2/3 and on the last one it crashes. (Issue in more detail here: https://github.com/blakeblackshear/frigate/issues/2156)

[MultiGamerClub]

So much hassle for a "GpU"

[danwat1234]

@LinusTech "Now in my home deployment, 5:34 I'll probably just use the CPU because 5:36 I'm going to have 24 epic cores 5:38 that are otherwise going to go unused."

 

Folding@home and BOINC needs CPU cores as well not just GPU compute!