A Third Type Of Processor For VR/AR: Movidius' Myriad 2 VPU

[scolaner]

Emerging tech such as VR, AR, drones, smart robots and more require vision processing, and Movidius aims to get in on the ground floor.

A Third Type Of Processor For VR/AR: Movidius' Myriad 2 VPU : Read more

 

[bit_user]

Google used Movidius' first gen product (along with Tegra K1 (A15 version)) in their 7" Project Tango tablet. I'm not sure whether it was really necessary (the upcoming Tango Phone looks to be Atom-based - let's see if it also uses a Myriad VPU). Since Tango is a prototyping platform, I think Google took more of a kitchen sink approach to give software developers more HW to experiment with.

To me, this doesn't seem fundamentally different than a GPU. For power-efficiency, fixed-function blocks are probably the main advantage (and by-and-large, they don't seem terribly interesting or sophisticated). But there's nothing to stop GPUs from adding those. And I don't buy the predication argument, because it's easy enough to emulate it by masking results of vector operations. If branching is more than a couple levels deep, then you need a different approach, altogether (basically, more threads).

I think Hololens probably does much more in fixed-function blocks. My guess is that Movidius probably intended richer fixed HW, but found a lack of consensus about what was really needed. Maybe that will change, in gen-3.

To be honest, what I find most intriguing is their choice of SPARC ISA. I haven't heard of anyone doing anything with SPARC in a while. Would be interesting to find out what's behind that decision.

 

[jasonelmore]

good luck competing with Nvidia, they have made it their priority to make VR, IoT, and Auto processors.. They completely ditched mobile and doubled down big time on VR and IoT.

As usual, they will probably use more power, but performance will be out of the roof.

I can't wait to see the 2nd and 3rd gen VR headsets a few years from now.

 

[bit_user]

...

Thanks for covering this, BTW. I appreciate the details, as well. If you get a chance, please ask them: "why SPARC?"

It would also be interesting to know if they've made any public statements about the number of design wins they had with the first gen or have lined up with gen 2.

 

[Shankovich]

So it's a GPU with a RISC CPU, interesting. If AMD was in better shape I could see them either imitating this or buying it.

 

[photonboy]

jasonelmore,
I'm not sure what this has to do with NVidia. This is a low-power processor for use in the VR unit itself, whereas NVidia's GPU's would be in the main computer.

I'd expect the Oculus Rift to have something like this though perhaps not quite as advanced for their first commercial release.

I wouldn't be surprised to see problems with PATENTS stall this or similar products.

 

[MobileEditor]

And I don't buy the predication argument, because it's easy enough to emulate it by masking results of vector operations. If branching is more than a couple levels deep, then you need a different approach, altogether (basically, more threads).

I did a little more digging and found this in an Nvidia whitepaper about Fermi: "In the Fermi ISA, the native hardware predication support used for divergent thread management is now available at the instruction level. Predication enables short conditional code segments to execute efficiently with no branch instruction overhead."

So any performance difference between Movidius' VPU and an Nvidia GPU on conditional code is not because one can do predication and the other cannot. There could be implementation differences, however.

If I'm understanding what I'm reading in Nvidia's literature, warps will serially execute both branch paths if a data-dependent conditional is encountered. Perhaps this incurs a higher performance penalty using SIMT (which would leave many threads stalled) than Movidius' SIMD approach. I'm not advocating one method over another (they both have their strengths), just trying to further the discussion.

- Matt Humrick, Mobile Editor, Tom's Hardware

 

[bit_user]

jasonelmore,
I'm not sure what this has to do with NVidia. This is a low-power processor for use in the VR unit itself, whereas NVidia's GPU's would be in the main computer.

First, AR products, like Microsoft's Hololens and Google's Project Tango have no "main computer". Second, Nvidia makes their Tegra SoC's for use in such products. So, in my mind, the question of how this compares to a GPU (and the competitive threat posed by/to Nvidia) is very relevant.

 

[MobileEditor]

Thanks for covering this, BTW. I appreciate the details, as well. If you get a chance, please ask them: "why SPARC?"

Movidius said cost was a factor in its decision to use SPARC instead of ARM.

- Matt Humrick, Mobile Editor, Tom's Hardware

 

[Urzu1000]

"There is a whole new market that is a sea of lava, gurgling to the surface, about to blow the top off of a mountain."

It's actually called magma prior to reaching the surface. I'll just show myself out now.

 

[bit_user]

Movidius said cost was a factor in its decision to use SPARC instead of ARM.

And I'm guessing they didn't even consider MIPS, since Imagination Technologies is sort of a competitor (they make mobile GPUs and computer vision processors). So, that would narrow the options, a bit.