Raspberry Pi AI Kit available now at $70 | Hacker News Reader

Raspberry Pi AI Kit available now at $70

raspberrypi.com

98 pointsTiredOfLife46 comments

I saw a demo of some software that can scan at 24 frame per second all of the workers behind a cafe counter, and classify them by which products they are working, their throughput, and special states such as using their phone, talking to a coworker, looking at the window.

what's something this thing could do in that same tier of hardware but would make people generally feel happy to see a demo of it? i know there are a thousand use cases that would feel good to hear, just literally spread a tiny drip of cheer with an anecdotal application for this

Probably a bird classifier. Aim the camera at a bird feeder, log what kind of birds are in view, save clips sorted by species. Make a graph of how many of each species there are each day over the year.

Coral has a smart bird feeder as one of their demo projects for their similar product: https://coral.ai/examples/

A smart lawn mower that navigates without fencing lines.

Gesture recognition integrated with Home assistant.

Some positive uses from the other day: https://news.ycombinator.com/item?id=40535636

(Funnily enough the post those comments are on used a Hailo accelerator and Pi cameras, but not actually a Raspberry Pi - it was before this kit was announced.)

> I saw a demo of some software that can scan at 24 frame per second all of the workers behind a cafe counter, and classify them by which products they are working, their throughput, and special states such as using their phone, talking to a coworker, looking at the window.

jeez. when the guillotines eventually come out I hope nobody is actually surprised by it.

We are all camgirls under surveillance capitalism.

somehow "classify employees by their productive behavior" doesn't seem sympatico with "spread a tiny drip of cheer" in my brain. maybe just me.

I think they were asking for a tiny drop of cheer because they saw the dystopian demo. As in, "please tell me there are ways this will do good and not just make life worse"

Hard to imagine a more pressing case for legal protections for workers against continuous workplace surveillance thaan the prospect of its being automated to this degree.

Also how do you throw your body onto the cogs and gears in the machine when they are this abstracted?

It’s not even a good demo. Even if someone is dumb enough to do this, they could do it offline on a recorded video feed.

Something like this is illegal in Switzerland and most likely many European countries.

It's nightmarish

Seems like there would be applications in sports.

Does this thing have a secret programming interface like the GPU does?

What is an NPU exactly anyway? A GPU without graphics hardware that works mostly in INT8? Or is the architecture different in other ways too?

An NPU is a type of ASIC for fast, parallel matrix multiplication. GPUs and FPGAs are much more multi-purpose (and more powerful overall typically) but at the cost of much worse power efficiency.

The GPU uses Mesa and KMS/DRM nowadays. Hardly secret.

"However, only a minor part of the driver was released as open source; all of the video acceleration is done using a firmware coded for its proprietary GPU, which was not open sourced. The entire SoC itself is managed by a ThreadX-based RTOS that is loaded into the VideoCore's VPU during bootup.[16]"

https://en.wikipedia.org/wiki/VideoCore#Linux_support

I saw that Hailo accelerator featured in another project on HN. Never heard of Hailo before, but if Raspberry Pi is reselling them, it must be good.

https://www.fpgadeveloper.com/multi-camera-yolov5-on-zynq-ul...

Hailo seems to be accelerating (ha! pun!) with funding, hw and sw

Suprised by the price tag. The board is probably <$10. 15tops should not cost $50.

I wonder whats the price for 1000 of these units. (hailo also has bigger units)

> The board is probably <$10

What are the alternatives at this price point? Given that a "Hailo-8 Starter Kit" is $179 (I'm not sure if that's more of less powerful than this Pi one), the Pi offering seems like a good deal. But to be honest I have no feeling for what 15 tops and a load of supporting software should cost.

https://hailo.ai/order-starter-kit/

The starter kit contains the Hailo-8 (26 TOPS), whereas the Pi AI kit contains the Hailo-8L (13 TOPS), so the pricing of the two seems more or less in line with each other at least

Right, so $80 worth of processing and the HAT itself ($10). $70 is decent.

The PCIe Coral TPU (with 4 TOPS compared to the Hailo-8L's 13) is still $25, and the USB 3 version is $60 — there aren't many low power alternatives and $70 for the bundle seems reasonable (with some profit margin) for what it has.

The article says they're providing GStreamer integration. Can someone tell me how a GStreamer pipeline to run the camera feed through an object detection ML model would look like? Am I free in the choice of model type? (I heard there are a few, one being something called ?Yolov?)

For me "available at" prices of Raspberry are unachievable. Either you can buy RPi for twice the price in my country, or you cannot buy at all. Depends on time of year or whatever

Which country? The pandemic pricing is over but Pis with any kind of upgrades at all are usually a poor value compared to to their competitors. Still a decent value if you keep them completely bare bones but they are losing their edge there.

What’s the competitor here?

ODroid made way more sense for my Home Assistant box, as I did not want to rely on an SD card that famously fails in Pis with a lot of writing. There are other various competitors in that form factor.

I've never bought one before but 1 Liter PCs and used equipment off eBay also are highly recommended. Depends on how powerful you need the box to be and how much power you care about it using.

Does anyone know how much memory the 8L has (or whether it depends on the host's system memory)? Hailo seems to think it's some kind of trade secret.

What can actually be usefully run on this?

Frigate, the open source CCTV NVR, can use the similar Coral TPU to do efficient AI object/person detection locally. I assume this could also do the job once software support is in place.

Can one run fast LLM inference on these?

Certainly doesn’t look like it.

Not really what the Pi is about, though. It’s a physical computing board.

Well, you could still build a lot with it, but the issue is it would be very expensive to have something that was able to do LLM inferencing.

Without the accelerator addon the Pi5 could do smaller models at passable speed, so I wonder if an accelerator can make this faster and allow larger models.

https://robleshermoso.com/posts/ejecutando-llm-raspberry-pi-...

Bit of a confusing typo there with 2.7 trillion instead of billion, but should check out.

There's an article about this stuff in The MagPi this month [0] -- and about a Stable Diffusion model trained only on truly Public Domain images.

I don't think the Foundation are asleep on the idea people will want to do this, but they clearly want to focus on ethical AI stuff, which is the right call if you ask me.

[0] Edit: last month, so the free PDF is available now:

https://magpi.raspberrypi.com/issues/141

How does this compare to a NVIDIA 4090?

Better comparison would be the google edge TPUs

This is true when considering similarity. However, most people considering such a device are likely more familiar with the performance of a 4090. Consequently, this comparison offers a more intuitive understanding of the potential decrease in performance.

4090 Founders Edition has listed Tensor TOPS at 1321 - assuming that it is actually a comparable metric

Roughly 5x cost per unit performance, at 1/100th the performance, with a more limited feature set.

But it's not like these are for inference or generative AGI or whatever.

> Roughly 5x cost per unit performance, at 1/100th the performance, with a more limited feature set.

And less than 5% of the price.

In other words; other than being some sort of processing units and associated with AI, they aren’t really that similar.

That seems similar to opencv on a tpu?

Think many will be disappointed when they realise this is a different flavour of ai than what the hype cycle is about

It’s in that ballpark, it would seem.

This is the right sort of AI for Raspberry Pi to be positioning themselves with: it’s edge-oriented, sustainable through the hype cycle, more interesting for a device that is aimed at DIY robotics and physical etc.

MagPi (the Foundation’s magazine) is running a series of articles about ethical on-device AI and has already covered LLMs and ethical Stable Diffusion models.

Reply on news.ycombinator.com