Nvidia buys chunk of video enhancing outfit

NVIDIA HAS JUST announced a "strategic partnership" with video software application firm Motion DSP, in which it has also bought a stake.

Motion DSP makes software that fixes video, allowing users to make videos from cell phones, still cameras, camcorders, or the Interwibble significantly less shoddy-looking using video enhancement and reconstruction technology.

This admittedly, isn’t really new. What is new, however, is that Motion DSP has now ported about 80 per cent of its algorithms to the GPU using Nvidia’s Cuda, which the company claims has significantly accelerated things.

By using the plethora of parallel cores on Nvidia’s GeForce GPUs and using Cuda on high-end cards like the GTX280, the firm’s software can actually reach real-time performance, something still not yet attainable with contemporarily available CPUs.

Motion DSP’s other software, Ikena Reveal, is already used by clients like the CIA and the military who reckon the apps have "CSI-style" video surveillance potential – for example, enhancing a blurry numberplate on a vehicle or helping police see more detail on CCTV footage.

The firm’s new software, dubbed 'Carmel', purportedly uses complex multi-frame methods to track every single pixel across dozens of video frames (up to 30) and then reconstruct better quality video from even rubbish resolution ones. The company boasts that by working on uncompressed bits it can not only significantly reduce compression and sensor noise, but even correct for bad lighting.

Carmel’s algorithm takes tens of video frames and combines all of their attributes to reconstruct each frame at a higher resolution than it was before.

"Our software runs almost five times faster on the GTX280 [than on an Intel quad core]" says Dr Sean Varah, CEO of MotionDSP. He added that this was true even when comparing GPU and CPU performance on exactly the same processor parameters.

When asked by the INQ if branching wasn’t a problem when using this kind of technology on a GPU, Varah noted that, although it was true that GPU streaming processors were less suitable for branching than CPUs, GPUs could still run code that had branches. He added, however, that the best approach to GPU programming was to avoid branching altogether if possible and that overall, video processing algorithms parallelise well, which was where GPUs really excelled.

"We were able, even without Nvidia, to start porting most of our code to run on a GPU thanks to their Cuda development kit," noted Varah, adding that ATI, despite having cards with equal amounts of computing oomph (about a teraflops), just doesn’t have something like Cuda yet.

Varah said his company was striving toward the "holy grail" of enhancing Youtube videos in real time, and admitted the firm had considered going down the Intel path. But, Varah noted, "If we stayed with Intel, we weren’t going to get there until Intel had at least eight-core chips out for consumers. By using Cuda to port to the GPU, we were really leapfrogging Moore’s law by 18-24 months and getting to real-time sooner."

Explaining why Nvidia had decided to buy a stake in his company, Varah explained that Nvidia knew it had to branch out and stop catering only to gamers if it wanted to survive. Asked by the INQ if the size of the stake Nvidia had bought was significant, Varah replied "It’s significant to us!".

Motion DSP says it hopes a retail version of its software can be released by the first quarter of 2009, but that there should be a demo out within the next month or so. µ

L'Inqs
See some video examples of how Carmel works
Basic overview of the tech