Man with a vision: Movidius' CEO talks Project Tango (Q&A)
Remi El-Ouazzane says his chip firm Movidius is more than just another partner in Google's mobile 3D-mapping project -- it's at the center of a revolution in how computers process visuals.
When Google's mobile 3D-mapping effort Project Tango hit the dance floor this week, it did so fueled in large part by the Myriad 1, Movidius' unique computer vision processing chip. CEO Remi El-Ouazzane says that Movidius sees a future forever changed by the firm's hardware and companion software.
A graduate of Harvard Business School and most recently the worldwide general manager of the Open Multimedia Applications Platform group at Texas Instruments, French native El-Ouazzane has a long history of involvement in Android. That made Google's Project Tango, developed by the Advanced Technology and Projects group formerly of Motorola Mobility and led by former Defense Advanced Research Projects Agency chief Regina Dugan, a natural partner for El-Ouazzane and Movidius' lofty goals of on-the-fly mapping. Potential applications for the technology include real-time directional guidance for the visually impaired and real-time, 3D environmental overlays for games.
El-Ouazzane is a relatively new hire by Movidius, a 7-year-old startup based in San Mateo, Calif., that's solely focused on building what its French chief executive described to me as a "vision processing unit." Just as the graphics processing unit, the GPU, joined the computer processing unit, the CPU, with Nvidia's benchmark-setting GeForce 256 in 1999, El-Ouazzane believes Movidius' VPU is the next big revolution in processors.
Different from the rare GPU synonym "visual processing unit," Movidius' VPU is unique because it allows mobile devices to process computer vision -- how your hardware interprets images seen through its camera lens -- in near real time, with far less drain on battery life than current chips offer. Its unique chip architecture is revolutionary, according to El-Ouazzane, because it solves the mobile computer vision processing problem of draining the battery while performing the complex calculations needed to turn images of the real world into data, and process that data into a usable product for the person at the end of the line.
El-Ouazzane, who speaks with a moderate accent and leans in when making a point, says Movidius' tech can be used in devices far beyond smartphones and tablets, to include wearables, robotics, autonomous cars, and array cameras.
This interview has been lightly edited from the original transcript.
Q: What was the biggest challenge in developing the Myriad 1? El-Ouazzane: Computer vision is a multistaged pipeline, highly complex in nature. It is very branchy, so you need to understand the computational complexity and the branchiness of the pipeline. When you have this figured out, the key in the processor business is to maximize parallelism and maximize locality of data.
So, we've built an architecture that maximizes locality and maximizes data. We've built an architecture that can deal with any kind of data types, and length of data types. It's all about the architecture memory fabric, there is a lot of customization there.
When the GPU was created, it was created as a simple instruction, multiple data engine for the purpose of rendering a pixel on the screen. Until somebody proves me contrary, it's the best architecture to build a pixel on the screen. It's a special purpose processor that does a bit more. We're the same.
Is the ancillary technology that would rely on the VPU at a place where it can extract information from the VPU? Is the ancillary technology at a point where it is ready today? El-Ouazzane: Yes. These use-cases [augmented reality, indoor navigation, 3D scanning, object tracking/recognition, immersive gaming, gesture recognition] are available now. The beauty about Tango is that our technology is not only embedded into a device but is with developers that can make sense of it. Google has that capacity. Johnny Lee [Google's Project Tango leader] and the developers will make sense of our technology.
We could not have hoped for something better than this.
How did the relationship with Google come about? El-Ouazzane: It doesn't happen every year that someone in the business comes along and adds to the CPU and GPU. The genius of Tango is that it's got the APIs for developers to use the information now.
The team that Johnny's part of is managed by Regina Duggan. This lady has a history of finding nuggets. That culture in that team led them to discover us, a very stealth company. An incredible commitment to bridge the gap between us and application is their foresight, their vision.
They've been able to make sense of all the assets that we've been developing.
In discussing the chip, it's low power consumption is a big deal. What does that mean? What's the difference between the hardware in my smartphone and what your chip uses? El-Ouazzane: It consumes hundreds of gigaflops. When we defined our mission seven years ago, we had to project ourself in time. Moore's Law is having diminishing returns, you don't get the power benefits. So, we knew that it would change from a general purpose processor to a special purpose processor.
Over multiple iterations, we developed an architecture optimized for computer vision. The architecture favors parallelism, and not frequency. We're running [the Project Tango phone] on hundreds of megahertz, using a mix of hardware and programmable resources.
On today's battery, running our processor will divide your battery power by a factor of 10. You will not even be able to perform that with the same pixel quality on your screen. Your frame-rate will be diminished, because your processor will enter a thermal zone that it should not be at.
It uses 10 times less battery power than today's phones? In all instances? El-Ouazzane: No, it only would save you battery in the context of where we're used. Similar tech is running on [Mars] Rover, but it runs on kilowatts of power.
Are you working with NASA? El-Ouazzane: I can't comment on that.
What kind of interest have you received from other chipmakers like Qualcomm and Nvidia? Do you think they're threatened by you? El-Ouazzane: Strong GPUs are our best friends. We're becoming best buddies. Visual computing relies more and more on vision computing. Visual is what you're looking at; vision is what you extract from the camera.
When Jen-Hsun [Huang] started Nvidia, his goal was to get this technology as widely adopted as possible. His goal was not to become a multibillion-dollar business. We have the same goal. We want to make a technology that can surpass what your eyes and cortex can do. That's a multigeneration process.
Our business is too tough, seven years is too long, to not have a multiyear vision.
What are your plans for the company? Do you want to sell? El-Ouazzane: We don't want to sell; we have big plans. We envision a world where devices surrounding us will continuously augment our daily lives, by providing intelligence in real time through visual sensing. The end user wants the perception of reality.
When will phones with Movidius chips reach consumers? What's the timeline? El-Ouazzane: It's in the coming next generation of devices.
I read a book over the weekend called The Second Machine Age. The author described the history of innovation with a graph. There was a hard 90 degree curve near the end. That hard shift was the steam engine.
Out of Project Tango, you've seen nothing yet. What's coming in the next 24 months is mind-boggling. We have passed the knee, we have passed the hockey stick [in the graph]. What's coming in the coming months, coming quarters, is very mind-boggling.