Earlier this year, Movidius came out with Myriad-1, its first “Vision Processing Unit”. It was adopted by Google for its Project Tango initiative, which has generated a lot of interest worldwide. Today, Movidius announces Myriad-2, its next-generation VPU, which is set to hit the market in products next year.
"MYRIAD-2 IS ALL ABOUT PERFORMANCE PER WATT" While Myriad-1 was targeting developers and was aimed at providing an early taste of the functionality that VPUs can bring to bear, Myriad-2 is all about performance. It is backwards compatible with Myriad-1, so any existing software will run without any modifications, but is much faster thanks to several deep hardware optimizations by Movidius.
The second-generation VPU from Movidius uses a mix of fixed-function hardware and programmable hardware, where the previous-Gen hardware only had programmable hardware. It is clear that using small dedicated blocks of logic to handle very specific functions is the way to go to increase performance, and more importantly, performance/Watt through the roof. There’s no way that programmable hardware can beat dedicated logic blocks for very specific tasks.
When I spoke to Movidius CEO Remi Remi El-Ouazzane, he went over specific details of Myriad-2. The new chip features 12 custom cores, whereas the previous one had 8. In addition to the increase in cores number, their frequency was more than tripled, going from 180 MHz to 600 MHz.
"AGGREGATE COMPUTATIONAL POWER: 2 TERAFLOPS" The aggregate hardware (cores + fixed functions) computational power of the chip is 2 Teraflops says Movidius’ CEO (this is a very theoretical number, but still a point of reference), and the 12 cores peak at 200 GFlops, which is not unlike Desktop PC CPUs launched in the past 5-6 years, except that Myriad 2’s power consumption is around 500mW (0.5W). That is precisely what makes it so desirable for mobile devices. It’s all about power consumption.
The Myriad 2 chip’s power can be harnessed via the OpenCL API, and developers have been working on apps for many months now. The first goal for a VPU is to bring DSLR-like photo quality into a tiny form-factor like a phone. Since using a DSLR-lens is not an option, computational photography is the only option.
The idea is that DSLR cameras are using a brute-force approach by simply gathering as much information as it can with a large lens and sensor. It works great, but is not possible on mobiles. The solution is to compute or reconstruct the visual information by using algorithms and other data (multiple frames, secondary lens, statistics…). With millions of pixels per photo, you need a ton of raw processing power, but at the same time, you don’t want your battery to go “poof” each time to snap a picture. Not an easy task…
Recently, I played with Tango demos at Google I/O and by next year, there should be a number of interesting apps, including 3D scanning, 3D modeling, 3D reconstruction, etc. For instance, you could imagine an app thatsnapsp a photo and provides precise measurements of anything in it. Handy when you are home or furniture shopping.
"ANYTHING THAT HAS A LENS COULD EVENTUALLY BENEFIT FROM VPUS" Beyond these simple examples, the possibilities are endless once you get a good vision system in the hands of developers: surveillance, auto-safety, robotics, and entertainment. Nearly everything that you do with your eyes could be enhanced, automated or supplemented by better computer vision. My personal take on this, is that anything that has a lens could eventually benefit from VPUs.
What about graphics processors (GPUs)? Remi Remi El-Ouazzane expects the Myriad-2 chip to reside side by side with GPUs. In his view, GPUs are excellent for Rendering, while Movidius’ chips are good at vision. In some sense, one is for input (vision), the other is for output (rendering). To be fair, GPUs are capable of doing some of the same work (check Project Tango on the development K1 tablet), so are CPUs – but Myriad-2 should be able to perform using less power because of its unique architecture.
Movidius confirmed that “very large” OEMs are already working on handsets using this chip, although they could not confirm who. At the moment, the first samples have arrived at Movidius and once the bring-up is done (make sure that things are OK, and drivers work), customers will start getting samples in the summer. This means that Myriad 2 was finished (taped-out) in the past quarter, just a couple of months after Myriad-1 was announced. The next step for VPUs: show us what they can do in the real-world.