I can't watch those videos at the moment, but the NNA doesn't really compete with GPU inference. GPU inference relies on large batch sizes for efficiency. It's geared towards data center inference. This NNA is for edge devices. A better comparison would be with Intel's Movidius Myriad X or NVIDIA's upcoming DLA.
I don't think we've heard from too many companies that need AI that they would like to take advantage of 1-bit precision computing. I imagine it would only be useful for some very specific tasks, but not the majority of AI inference tasks.
This chip is designed for a very specific task: image recognition using convolutional neural networks. Which work with 1 bit precision with moderate accuracy drop (improving every few months).
This is interesting as one of the main reasons Apple left them for their own designs has been said to be because Imagination refused to work with Apple on this very thing. In addition, Microsoft left because of a refusal to work with them as well on a specialized machine learning processor.
So this announcement is very strange, particularly since Apple has hired some many of their engineers away over the last 18 month, and this is a very small company to begin with.
Other than Apple, they had no other top tier customer for their GPUs. I wonder how well this will work out for them.
Although its sad when a company goes under, these types of projects are great for employees since they can leverage it on their resume when they interview at Google, Apple, Intel, QCOM, Nvidia, AMD, etc.
This is a field that has a lot of interest and money pouring into it.
Yes, it's why there's all this coming out, announcements, and open source. It's a competitive field with a relatively small labor pool that understands the subject.
Was the issue that Imagination didn't want anything to do with neural network hardware, or that they didn't want their IP tied to Apple or MS? Maybe what they wanted was the ability to sell it to other customers, and without exclusivity windows or having to pay royalties to those guys.
IMO, the only thing strange about this announcement is timing. They are late to the party.
Interesting article but some of your pictures do not load (either here or if you open them independently). The same happened with the Thunderbolt 3 dock article..
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
18 Comments
Back to Article
KristofBeets - Thursday, September 21, 2017 - link
Couple of their videos help to put performance into context:GPU inference: https://youtu.be/QGWNJM9DJzE
NNA inference: https://youtu.be/ut948L1YL0g
Yojimbo - Thursday, September 21, 2017 - link
I can't watch those videos at the moment, but the NNA doesn't really compete with GPU inference. GPU inference relies on large batch sizes for efficiency. It's geared towards data center inference. This NNA is for edge devices. A better comparison would be with Intel's Movidius Myriad X or NVIDIA's upcoming DLA.iwod - Thursday, September 21, 2017 - link
>In conjunction with today’s PowerVR Series 9XE and Series 9XM announcement,Am i suppose to Google this myself?
yannigr2 - Thursday, September 21, 2017 - link
Yes you are. If you are lazy, just hire someone to google for you.name99 - Thursday, September 21, 2017 - link
You could hire LouiseHolt! She seems to spend her life cruising AnandTech boards looking for work...Threska - Thursday, September 21, 2017 - link
Say it's booming but the only field Imagination plays in is mobile. Cloud and autonomous vehicles would be someone else's playground.p1esk - Thursday, September 21, 2017 - link
I'm surprised they didn't go all the way to 1 bit precision for both weights and data.Krysto - Thursday, September 21, 2017 - link
I don't think we've heard from too many companies that need AI that they would like to take advantage of 1-bit precision computing. I imagine it would only be useful for some very specific tasks, but not the majority of AI inference tasks.p1esk - Thursday, September 21, 2017 - link
This chip is designed for a very specific task: image recognition using convolutional neural networks. Which work with 1 bit precision with moderate accuracy drop (improving every few months).jospoortvliet - Monday, September 25, 2017 - link
This core can't do voice recognition or text prediction?melgross - Thursday, September 21, 2017 - link
This is interesting as one of the main reasons Apple left them for their own designs has been said to be because Imagination refused to work with Apple on this very thing. In addition, Microsoft left because of a refusal to work with them as well on a specialized machine learning processor.So this announcement is very strange, particularly since Apple has hired some many of their engineers away over the last 18 month, and this is a very small company to begin with.
Other than Apple, they had no other top tier customer for their GPUs. I wonder how well this will work out for them.
webdoctors - Thursday, September 21, 2017 - link
Although its sad when a company goes under, these types of projects are great for employees since they can leverage it on their resume when they interview at Google, Apple, Intel, QCOM, Nvidia, AMD, etc.This is a field that has a lot of interest and money pouring into it.
Threska - Thursday, September 21, 2017 - link
Yes, it's why there's all this coming out, announcements, and open source. It's a competitive field with a relatively small labor pool that understands the subject.mode_13h - Thursday, September 21, 2017 - link
Was the issue that Imagination didn't want anything to do with neural network hardware, or that they didn't want their IP tied to Apple or MS? Maybe what they wanted was the ability to sell it to other customers, and without exclusivity windows or having to pay royalties to those guys.IMO, the only thing strange about this announcement is timing. They are late to the party.
vladx - Thursday, September 21, 2017 - link
They are not late at all since neural networking is still in its' infancy especially in the mobile and embedded sectors.Tams80 - Friday, September 22, 2017 - link
That would make sense. I could certainly see the likes of Apple and Microsoft demanding some sort of exclusivity.Santoval - Thursday, September 21, 2017 - link
Interesting article but some of your pictures do not load (either here or if you open them independently). The same happened with the Thunderbolt 3 dock article..ilogic135 - Friday, September 22, 2017 - link
Interesting Article!