Monday 19 February 2007

Sony's new 6 Megapixel CMOS Video Sensor

Sony’s new CMOS imaging sensor seems to be attracting lots of ooh ahh comments on the Internets, this is Engadget's take for example

This is an incredible sensor but something smells wrong to me. It's not that this isn't a totally cool piece of tech hardware but it didn't seem to add up for me.

Anyway - taking a look at the PDF.

On the surface it’s definitely a 60fps 6.4 Megapixel sensor. That’s 384 Megapixels per sec. We could do some digging but lets just run with 24bit RGB for now for how much data we need per pixel. If this sensor is able to output 6.4Megapixel full colour video at 60fps it needs to be able to output 9.2 Gigabits per second. That’s a lot of bandwidth – it’s pushing out a DVD in 4 seconds for example - and it's almost triple the theoretical maximum capability of the best commercial hard drive interface on the market today (3Gbps SATA).

Anyway ploughing on with the PDF and you find that what they actually say is that the 6.4Megapixel 60fps gets ouptut at “10 bit data at a pixel level”. That would then be 6.4Meg*10bits per pixel * 60 frames per sec == 3.8Gbps. They also say that the interface uses a 10bit LVDS interface @ 432Mhz which would provide just slightly more bandwidth than is strictly necessary. It's still an astonishingly fast interface but now at least it all adds up from that point of view.

However that’s all still only at 10 bits per pixel which is nowhere near good enough for colour unless you are talking about GIF\PNG (10-11bpp typically) or JPEG (<10bpp)>

Assuming that the intention is to use a Bayer array or something like that and some fancy demosaicing then they will be interpolating this into a 3.6Megapixel full colour data stream albeit with some loss of precision buried in there.

The sensor size (2.5micron) also seemed awfully tiny to me and I suspected that it might be pushing the edge in terms of diffraction limits. It is quite interesting to look at the Airy disk size for this sensor: go to http://www.hdrsoft.com/resources/dri.html , plug in the base values and this shows that the sensor could be noticeably diffraction limited even at f/stops as wide as F4. Adjusting for the Bayer array and basing it on 3.6Megapixel sensor that still kicks in at F8 which still isn’t very good for a still camera but is probably not an issue for video unless the viewers intend to have their noses glued to the screen or plan to watch ultra slow motion on your brand new 50" 1080p flat screen.

Anyway so the fairly high resolution "on the fly" still capability is interesting and not totally shabby but the key capabilities start to become obvious when you look into the additional output formats listed in the final table as these quite obviously pitch this (quite sensibly) as a sensor for proper HD (ie 1080p & 720p) Digital Camcorders; the 2x2 binned 1GBps 10bpp 1440x1080 60fps output format and a 3x3 10bpp 960x720 60fps mode are for shooting widescreens (1.3333:1 ) onto a square sensor array so that they can be scaled back to full 1080p (1920x1080 with square HD pixels) or 720p (1280x720 square pixels) formats respectively.

The 2x2 binning mode then supports the highly desirable "Full HD" 1080p video (2Megapixel) at 60fps. And yes I suppose it is obvious that that is what they need to be making these days but the PR pitch about 6 Megapixel video is a bit of a red herring, don't you think?

1 comment:

Iain McClatchie said...

Joe,

Data rate: It is really doing 60 fps, 6.4 MP, 10 bit A/D per pixel. This is not unprecedented. Micron has had a 4MP 200 fps chip with 10 bits per pixel for quite some time.

You won't be storing the raw data coming off this Sony chip. They'll build you a nice compression engine that will get 50:1 or more video compression, so that you'll be down to 8-10 MB/sec. A 4GB SDHC card will give you almost seven minutes of super-high-res slo-mo recording. Downsample to HDTV and your card will give plenty of recording time.

Color: It's a Bayer pattern sensor, just as you suspect. That means each pixel sees just one of red, green, or blue, and you have to interpolate.

Pixel size: a 2.5 um pixel is not amazingly small -- the 8 MP micron sensor has 1.75 um pixels, for example. Small pixels are bad because they are limited by photon shot noise. See http://www.clarkvision.com/imagedetail/does.pixel.size.matter/

Aperture: with pixels this small, the lens for a camera based on this sensor will be f/2.0 when full open. They probably won't do f/1.4 because it is very expensive to make such a large relative aperture lens with acceptable abberations. You are right that tiny pixels like this get diffraction limited at quite modest f/#s. What can I say, small pixels suck, but hey, they are cheap.

The other shoe: Sony doesn't say anything about quantum efficiency, so that might be bad, since it's usually spec'd. It could be that they haven't figured out the microlenses yet (need a few more months for that), so they are showing off a sensor without microlenses, which will have terrible QE.