Hacker News new | past | comments | ask | show | jobs | submit login
Edge TPU: Hands-On with Google’s Coral USB Accelerator (fritz.ai)
49 points by rdeepc 42 days ago | hide | past | web | favorite | 20 comments

The Edge TPU looks like a cool toy. But I would never base a product on anything that Google makes (either hardware or software). It's here today, and might not be here tomorrow, at Google's whim. I think the value of Edge TPU is mostly educational.

And before someone waters this down with "but every company has whims and every company can discontinue products" — in electronics, companies will guarantee 10-15 year lifetimes, especially for automotive-grade parts, with end-of-life buy notices well ahead of sunsetting.

Even Raspberry Pi which started off as specifically educational does 5 year release cycles on their CM (Compute Modules) which is why companies like NEC use them in their solutions.

Yes — and I have actually designed a product that used an RPi Compute Module, checking their lifetime promises first.

> ... in electronics, companies will guarantee 10-15 year lifetimes.

That's true only on parts that are already in high demand, and this doesn't strike me as a high demand chip (at least not initially).

And even then, you're not guaranteed any kind of availability on the part. (Maxim comes to mind). Or it's only available with a 6 month lead time and minimum part quantities on the order, etc.

So this seems to be a problem of this particular space, and blaming Google for acting normal in this space seems silly.

(To be clear, the edge USB device works with the raspberry pi, but throughput is limited to USB 2.0 speeds... Because that's the speed of rpi USB ports. It's not clear to me what the author is asking for wrt full support for the pi.)

Thanks for pointing this out. The wording there didn't quite capture what it was supposed to. We've gone back and adjusted the language to reflect this (I helped edit this post with the OA).

The biggest problem is that they provide no specs whatsoever and you have to use their cloud compiler for your models. Even if you stockpile the hardware, if Google gets bored and shuts down the service, you can’t train anymore. Google also gets to see every model you ever deploy to the thing.

That is a big problem, for now, hope there will a converter script in the stable release

A benchmark test against Nvidia Jetson would be very interesting, especially if it compared power consumption as well.

There will be a comparison with Intel NCS

Here's a video I made of one of the included demos on RPi: https://youtu.be/6uQlCiAGWzc

I didn't yet succeed in converting the installation script so it would run on Arch (unsupported) but I'll try again over the weekend.

The support team have been responsive on a number of questions too. When they open up the range of models that can be run, it'll be a huge boost.

A curious detail I noticed:

the USB Id of the Coral Accelerator is for Global Unichip Corp, who are a fabless ASIC design service company (according to their Wikipedia page). I had been expecting it to be Google obviously.

That surprises me too - it's very unlike Google to contract out tech like this.

Given how the SOM version connects over PCI-E, maybe it is a USB3.0 to PCI-E bridge?

Are there comparisons between this and the google AIY vision kit with their vision bonnet?

I’d assume it was developed at google by the same team

There will be a comparison with Intel NCS

Hm but is it fast?

Its faster than any other products of its kind

I've love to see some performance comparison for the Google Edge TPU, the Intel USB widget, and anything else in that price bracket.

I have Edge TPU, Intel NCS, and vision kit hope to make a comparison among them

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact