Google’s New TPU Turns Raspberry Pi into a Supercomputer!
HTML-код
- Опубликовано: 14 окт 2024
- Well that escalated quickly…
coral.ai/
Google for local AI
Coral helps you bring on-device AI application ideas from prototype to production. We offer a platform of hardware components, software tools, and pre-compiled models for building devices with local AI.
We’re making AI more accessible
We’re creating a flexible development system that makes it easy for you to grow embedded AI products into reality.
We believe in privacy-preserving AI
We’re exploring new ways to implement AI by bringing it local, making it fast, reliable, and respectful of user data.
We believe that AI will help create a better world
By working with creators, designers, engineers, manufacturers, and industry, we can help build truly beneficial AI for our world.
Coral technology is enabling a new generation of intelligent devices
Coral provides a complete platform for accelerating neural networks on embedded devices. At the heart of our accelerators is the Edge TPU coprocessor. It's a small-yet-mighty, low-power ASIC that provides high performance neural net inferencing.
Learn more about Coral technology
Products flexible for prototyping and production
Prototype
Imagine what’s possible with local AI, then move right into proof-of-concept. Our products make it easy to take your idea from sketch pad to workbench.
Scale
Our products are designed to grow with you, integrating seamlessly into production processes at any scale.
Deploy
Robust modular components bring reliable AI to more applications than you ever dreamed possible.
Coral is brought to you by Google Research.
We’re bringing you the best of Google Research
Coral is a multi-year effort between many teams across Google, including Cloud and TensorFlow.
www.raspberryp...
Raspberry Pi is the name of a series of single-board computers made by the Raspberry Pi Foundation, a UK charity that aims to educate people in computing and create easier access to computing education.
The Raspberry Pi launched in 2012, and there have been several iterations and variations released since then. The original Pi had a single-core 700MHz CPU and just 256MB RAM, and the latest model has a quad-core CPU clocking in at over 1.5GHz, and 4GB RAM. The price point for Raspberry Pi has always been under $100 (usually around $35 USD), most notably the Pi Zero, which costs just $5.
All over the world, people use the Raspberry Pi to learn programming skills, build hardware projects, do home automation, implement Kubernetes clusters and Edge computing, and even use them in industrial applications.
The Raspberry Pi is a very cheap computer that runs Linux, but it also provides a set of GPIO (general purpose input/output) pins, allowing you to control electronic components for physical computing and explore the Internet of Things (IoT).
So nobody’s gonna talk about 73% barbell and 11% dumbbell?
Lol train that shit!!!
It's a heavy metal guitar.
73 percent weight lifting tool lier
@@tylisirnlier
@@Jblow-u2m agree! Have it no manners, Sir?
9% man not 91%
Although it wasn’t showed in the video, the top of the screen is what percentage of the item fills the frame. He is getting 91% from the number at the bottom of the screen where you can see the console. It is a bit cropped but you can see 91 then quickly changes to 90. Mostly his fault for not showing the correct percentage more.
@@ImStveL He needs to fix that man
No you idiot. 9% of the entire image is the guitar, the bottom says how certain it was!
@@ImStveLlol don’t know what you’re smoking. The screen is displaying 1 result, and the bottom is displaying 1 result when the program in his full video is classifying every frame. Not sure what drugs the video maker is doing to say 91% though, when that was the maximum result. The classifier was hovering around 70% most the time as electric guitar
@@kurtsaidwhathis explanation is also confusing, saying the pre build model is classification when the image clearly show object detection
Image recognition has been possible on low end devices for decades, without using tpus.
I remember seeing a face detection python library run perfectly smoot on a rpbi2
You can run a 90s computer and have image detection. It's very simple, it's all on your resolution. It's all on a lot of little things. If you look at the leaders of Vision and image detection for the factories of KEY ants and cognics, they're software and hardware is Really lame and they charge a lot of Money to be the industry standard. But the technology is really weak. It's like my old surveillance system from the year 2000 with motion and object detection. It's just software that overlays the image. And looks at the pixels and figures out. What is this making up in My Library?
Face detection does not need to use ai. It is a "solved" Problem there are algorithmus to detect faces that use some special filters.
You're conflating two things. Face detection and image recognition are very different things. Facial detection is much simpler because we have pattern recognition algorithms that are very generic which have been known for several decades. Old phones had this tech on them. However, looking at a high rez photo of anything and being able to name what object is featured in the photo is a wayyy more complicated problem.
I implemented it on a small scale robotic arm on my last year of school. Tiny shit followed you across the room using a raspberry Pi 2
@@TheHanutaXDthis isn’t “ai” either. It’s a machine learning model that has come up with… drum roll… an algorithm. This algorithm happens to be more capable in that it can identify a broader set of objects. But the underlying algorithm is pretty similar to the “decades old” one you knobs are yammering about. So you can save your neck-bearding
Image recognition can be classified as a lazy learning algorithm. It takes a fast gpu and lots of ram to train, but after the model has been trained, the usage is very lightweight.
Yes. (Pin this)
All of this hardware is almost half a decade old now and you need to run an old version of tflite to guarantee models run mostly on the TPU as most modern architectures including any form of transformer are completely unsupported.
My iPhone 10 could do this
@@Puppy__Nietzsche I mean sure, BUT.. your iPhone 10/"X" had a dedicated AI algorithm-crunching chip in it's very powerful (at the time) A11 processor. A Raspberry Pi is (supposed to be, if they weren't scalped constantly) a $35 SBC for hobbyists and students, not a $700-$1000 palm sized supercomputer. The point of this TPU, as he explains in the video, is to make a very cheap computer capable of doing some cool things you'd normally have to buy a much more powerful device for.
I know this is old but the coral came out in 2019 that's not a decade I currently use for objects and face detection on a nuc
@@outbakjakThe Raspberry Pi can already do ~5fps live detection in camera. it ain't good but it's still very usable, especially if you use SSD mobilenet
@@Puppy__NietzscheMy S8 could do this lol
You could use YOLO tiny for that. I used in my project and it absolutely ran in raspberry pi itself. You don't need any other devices attached to it
What kind of raspberry did you use? I’m trying to do the same with YOLO Tiny
You totally can but how else are shitty RUclips’s gonna make content?
Ah yes. 73% Barbell.
You can easily run image recognition on an ESP32, i would have loved to see an actual challenge.
oh god
@@realSkyfr oh God what
I do use ESP32 for ai-on-the-edge project, but it take 5min to process single image. Coral TPU is better for real-time recognition in my frigate NVR (90ms).
When his voice started sounding off I thought he was gonna do a reveal where he processed the video off of a raspberry Pi.
i also noticed his voice
Gotta be AI
TensorFlow lite has been able to do this on ARM microcontrollers like the pi or even a teensy for years. It works by offloading the training onto a more capable machine so the actual embedded device only needs to be able to execute the output. The Coral is basically only there for the training portion and once that's done, it's kind of unnecessary.
Google's Coral is an Edge TPU. It's a dedicated "low power" device designed for accelerating local ML inferencing. It's not meant for training.
A Rasp Pi with TF-lite will do the same thing, but it will be slower and eat more power than a Raspberry Pi with a coral device.
73% barbell🤣🤣🤣
"the most challenging computation" is not ML... there are many like weather prediction, protein unfolding etc.
protein unfolding and weather prediction also uses lot of machine learning underneath.
@@anshXRThey can run on a ML based algorithm, but protein folding and weather folding can be run as a raw simulation (MD for Protein folding as an example) which is a very demanding algorithm. ML algorithms learn patterns and utilise them , it's kinda like using approximations.
Even raw simulations use approximations, but there are levels of approximations. A low level approximation is highly computationally intensive, but high level approximation like ML easier to compute.
Both of those qualify as ML problems
Been using Frigate for my CCTV for years with the exact same USB TPU. They have a PCI version too incase you don't want the dongle dangling everywhere.
First used it on a Mini PC. Moved to a pi 4 years ago. Recently upgraded to pi5 to run even more along side.
"I totally understand this code" i can relate 🙂
This is the comment I was lookin for
6 years ago on Raspbery instaled (objects recognition) with camera instant. The software was more old.
Wow I didn’t realize how expensive Pis had gotten
The ones you're used to buying cheap are still available and for similar low price but newer raspberry Pi's have gotten more powerful as well, so like everything else, you have to pay more for the more you're getting,
as the hardware increases its capabilities.
Raspberry pi is a company has also brought forth a lot of new offerings that are low power and cheaper and not necessarily less powerful than a more expensive pi, depending on the task you want to utilize it for.
I think the last one I got was $15 retail to be used for its small form factor and it's dependability for a repetitious task.
@@JericHouse copy that. thank you for the details.
I think RPi5 is able to perform realtime object recognition in real time without the help of an external accelerator like this one. So I don't know what is the point of Coral in this case.
Yeah-- Try NP-Hard/Complete problems. Such as the The Traveling Salesman Problem, The Knapsack Problem, Boolean Satisfiability Problem, Graph Coloring Problem, Hamiltonian Path Problem, Subset Sum Problem.
But does it need internet for the coral to work? To gain its smart intelligence.. Doesn't need network to think or gather its data to begin its process. Or is it seriously isolated as a brick terminated?
Nice PRS!
Image recognition is absolutely not out of the question on a pi. If you have a trained model, object recognition is relatively light weight. For real world comparison: eufy‘s security cam base station runs on quad-core a55 that is slower and has less ram than most pis. And it does it with 4 cameras streaming simultaneously. You don‘t need a TPU for the trained model.
I've been using yolov5 on my pi 4 and it has no problem running it
That's awesome man, real time camera recognition will take a step further with this, you don't need a powerful not portable system to that, and thanks for the don't trip hat reference of Mac Miller, love his music.
You can already do this with a standard pi 4, and if you have a basic NVidia LP Graphics Card like the 1650 you can get like 4x more performance
@@SUPABROS 7w from tpu that do the half of 300w graphic card idk seems petty good and Fair
Why did bro alter his voice or use an ai transcript?
If you have a use-case in which the camera and the objects aren't moving fast (define fast for yourself), you can use this simple sounding trick:
Do the object detection in every 5th frame but keep the bounding boxes on the screen and the video regardless, by choosing a lightweight algorithm, you should be able to get around 30fps!
"I'm sorry, Hal, I can't do that until you shave."
The TPU usb accelerator is from 2017
I compiled opencv(took almost a whole day) in my pi 3 and its able to do realtime object detection without any tpu
Everyone knowns calculating digits of pi is the most computational task.
Once you have the edge tpu set up does it require an internet connection to operate?
Coral has been around for years
but the rz isnt doing the computing?
With yolo and cv2 you can also run object detection in videos without google coral on raspberry pi 4 and 3b
This was not released recently. Its a 4 year old product. But its been out of stock forever. Also comes as PCIE card if you want.
Good content man! You just earned a sub
Spot on "I totally understand this code" 😂👍🏻👍🏻
I've done this with a plain old pi3, it was slow, about 5fps, but could still identify things.
Not smooth but not new and i reckon a pi4 would fair much better
Whats the song you use towards the end of the video? Sounds good
I am an Android app developer
This is being done on smartphones since probably 5+ years (Both Ios and Android)
Yes locally on smartphones not cloud. I am talking about complex object detection, not just face detection etc
Ok but why would you spend 1000 on a phone to use as a security camera? You realize different hardware and formfactors have different use cases yeah?
@@meinbherpieg4723You can do it on a 100 dollar cheap Android phone
You should make a magic mirror
I have a slow computer and i'm learning machine learning, no budget for upgrade, can coral tpu enough for me to atleast get a grasp on machine learning fundamentals?
What’s the music where you inserted code like Usain Bolt? 😂
That is honestly so fkin cool they have an AI PU like that.
Im curious how well this would work if there was code for it to work like an ASIC and do crypto mining.
Is this different from computer vision?
Jetson nano's filled the rpi form factor machine learning area for a while.
Is audio was made by ia?
This should make their share price go through the roof!
Are you using ana ai voice for this video? It sounds... Unnatural.
That s really good. Add a voice , mutiple cameras . And youd have a amazing aid for blind people . U heard it here first .
If you attached it to a data base of imagery could it identify a person from the list as they walked into the camera ?
Would you recommend coral tpu still?
So which AI platform is this based on again??
Mac Miller reference with the Don’t Trip hat?
RUclipsr try not to use supercomputer incorrectly challenge (very hard)
I've seen that thing used in a bunch of videos and never knew what it was. I thought it was some pico/zero or something
After years of being hit with object detection ads, still have no clue what is it even for. You can buy cameras wth built in object detection, so whats this tpu is ever used for?
First, not sure when the video released, but i had one of these back in 2021, so its not new. Also, it only brought the frames up on a yolo v4 bounding box algo to like 32fps. Turns out my smart phone has a better camera, processor, screen, is smaller, etc than a rspbrry pi. So i remade my app in kotlin and called it a day.
Iicc wasn't there an Android app that did the same thing called "tensorflow"? Or something?
Didn't someone do something similar with a Nintendo Wii body motion device or the XBox motion device?
The product is old came out in March 2019
i took mine to the absolute limits too! i overclocked it to the absolute max, even watercooled it to play proper minecraft on it.
But did it say “Paul Reed Smith” or “Orange Amplification”?
amazing video! keep it up
So is this just good for image recognition?
First time I've ever seen fractional FPS
Better than the rabbit r1
Either way this Tech is really sweet!
I subscribed. Although, do you know of any better ai video detection packages or hardware, etc?
cv2 and a good ol' fashion GPU. My 15 year old lenovo can run object detection with no problems. Shit, an android phone can run it. Training the model is what is time and power consuming. As long as the model is trained virtually anything can run it.
That really changed everything……
Yeah uh... you can do that with a wisblock bro.
Wait, so you are telling me that if I take a rpi zero w 2, or rpi 4, and I design a neural network to recognize a specific animal on a live video feed, I won't be able to without this ?
how do you know its being used?
thats about the same performance as the pi itself?
Does it also upload whatever I do to Google?
my bushitmeter just went off limit. so many things you said are just wrong
I'm building a weird drone to do weird things. I know most high end drones can already do this but not open source...
They always want to convince people to stop trying.
This is just one more example of trying to get people demotivated from making their own controllable hardware.
Could I run a cluster of these?
Didn’t this come out like 4 years ago?
Yes
Awesome 👌
Compare the Raspberry vs the raspberry with TPU
This has been out for a really long time, and tons of raspi alternatives have TPU built in. If you want to be known gor high quality videos, use high quality information, not low quality lies
There's dual tpu module for two of em in one
Ah, a dedicated matrix multiplication calculator
Has anyone ever told you, you look like Ryan Reynolds at home?
Yeah Coral ran a code and classified him as RR with 91% certainty 😢
Suggest me some real time projects using raspberry pi
73% barbell and 7% electric guitar. Why saying something which is not correct. Even if it was not expected the concept was nice idea but falls information is always - for me.
Although it wasn’t showed in the video, the top of the screen is what percentage of the item fills the frame. He is getting 91% from the number at the bottom of the screen where you can see the console. It is a bit cropped but you can see 91 then quickly changes to 90. Mostly his fault for not showing the correct percentage more.
Google's TPU is not new. You're like 6 years behind
Rpis have been doing image recognition for ages....
quite literally just ran openCV on only an rpi5 in a dual thread with a gui
This isn't that new, but very cool!
This could be used to let security cameras be able to detect a gun
That's been out for years and years. My consumer cameras from 2022 have weapon detection built in. Knives, guns and crossbows are identified in sub 2 seconds.
anyone else thought the thumbnail was a hand rolling a joint?
What are the use cases?
Security cameras. Feeding description information into a development pipeline to do something more with what the AI recognized. Home automation. Car mods. Entertainment? Games? It's literally up to you to create something cool. This isn't an entertainment product. It's a tool.
it was more sure that the guitar was a dumbell than a guitar
This would be very useful if you have Security Cameras connected to it to make sure there's no skunks, raccoons, bears, or mountain Lions in your backyard before leaving your dog out at night to pee!
Now make it do captchas
This is great! I can see a wave of raspberry pi home security systems.
Yes it can
Do Realtwiprasa