DSP: XL2 = 8 bit, DVX = 12 bit

[nullphonic]

Would this cause a significant difference in quality really?

Quote from dv.com...

"And 8-bit models, which constitute by far the great preponderance of DV cameras currently on the market, must make do with a 256 assignable values per sample.Of course, there is more to consider than bit depth alone when you evaluate camera performance. Other factors just as or even more significant also come into play.Despite the seeming logic, some prickly engineers dispute the benefit of increased processor bit depth with respect to DV camera performance. These engineers argue that the eye can resolve pixel variations no greater than 230 to 250 per sample, so there is little point to designing a camera that can resolve difference greater than that"

Another Quote...

"To these "8-bits-are-enough" folks, the highly endowed processors found in some DV models are little more than a marketing gimmick.Nevertheless, it is a fact a 12-bit processor can substantially increase a camera's dynamic range"

Here's the link to the article...

http://www.dv.com/print_me.jhtml?LookupId=/xml/feature/2003/braverman0303

(Go down to "Watch your DSP")

[Guest]

A quick word from the engineering perspective - in general, an 8-bit or 12-bit DSP refers to the register size, not the analog to digital converter. Different DSP's have different architectures which is important to consider from the algorithmic viewpoint, but may or may not affect the overall "quality" of the camera image. A quick example - Microchip based DSP's are 8bit architectures with 12bit ADC's. Analog devices DSP's are usually 16 bit architectures and offer a wide range of ADC depth. So the point - DSP architecture may not matter - ADC depth may. Hope that makes sense.

[nullphonic]

I know that ADC depth makes a huge difference in audio processing so the same applies somewhat (I'm gathering) i.e. DSP capture isn't as significant (relative) to your ADC depth. I know that the difference in an 8 bit ADC and 32 bit ADC in audio is painfully obviously (digital noice) when analyzed.

(I'm obviously no engineer, could you tell ;D, but that makes sense)

[jaredalv]

"To these "8-bits-are-enough" folks, the highly endowed processors found in some DV models are little more than a marketing gimmick.Nevertheless, it is a fact a 12-bit processor can substantially increase a camera's dynamic range"

Is that last sentence saying that a 12-bit processor will increase a camera's exposure lattitude? If so, that's an obvious "better" to chalk up to 12-bits.

Trying to wrap my brain around this,
Chico

[Dan_Lahav]

I did a side-by-side test of my friends "A" and the regular DVX with identical settings. Didnt see ANY difference whatsoever. I wish i would have kept those clips...

[alpi69]

dunno, but 8-bit to 12-bit might not be visible raw from tape. but when rendered a few times i am sure the more information a pixel had before the better it is. but i am also no engineer and will have to rely on people like adam wilt to make these tests.

[scharky]

I thought I would add my two cents to this little conversation. Sorry if the post seems wierd, I just woke up because I have an early morning meeting at 6am :(
I work at a photography studio doing all of the digital imaging and retouching. Many people have asked me why I prefer to color correct in 16bit space as opposed to 8bit in photoshop. Like others have said on these forums, because it is way more accurate and allows for greatrer changes. You actually have twice as much information two work with, so when I strethch something that is way under exposed, those 16bits will keep it from posterizing in the final product. Before I go out to print, I have to convert it to an 8bit file, so it seems that all of the extra info that was there is just lost, but you really can tell the difference in the final product. I assume that it works the same way with 8/12 bit processing. Even though we are not able to work with the footage itself in 12bit, the camera does have much more information to work with before it is compressed and layed down to tape. What will this give us. Well, like many have said, colors will be more accurate and have less (off pixels) that make them up. The latitude should as well be improved, since there are so many more scales of grey to work with in the initial footage. Low shadow areas will have more detail to be brought out after compression. Will you notice the difference in every shot? probably not, but in the end, be rest assured that it is there and is giving you the best results possible.

[Guest]

Consider this - to acheive the 16.7 million colors on a computer screen, each pixel has 3 parts - red, green blue, each with a value between 0 and 255. This means that each color component is only an 8-bit value. Therefore, in a 3-CCD camera (i.e. DVX, XL2, etc.) sampling each CCD pixel at 8-bits should be perfectly adequate. This only applies to the analog to digital convertors though.

This 8-bit sample is then processed in the DSP, where one cannot say that 8-bit, 12-bit, or 16-bit is better or worse. It all depends on the algorithms used, speed of the chip, memory available, etc.

[Scottdvx100]

Just as in Photoshop if your original capture is more than 8 bits then it allows things like the gamma curves to have more control and produce finer increments. If you start with 8 bits you will lose color resolution when you start to manipulate the color.

If you ccd's bring in a scene that goes from light gray to dark gray and your color gamma settings stretch it to full black and white- 8 bits means your original range might be 200 values so there will be duplicate values in the final 8 bits. If it's 12bit then you have 3200 values (in this example out of 4096) and can easily still have 256 discreet values even after changing the gamma or lookup.

Even if the original capture is 8 bits (and I don't knwo the detials of the A model) internal calculations done at a higher level are still a good thing since many match functions can clip at 8 bits.

[ChuckS]

In many cases, in a side-by-side comparison, most people probably could not see a difference between an 8bit and 12bit CCD. However, having said that, there is a potentially significant difference – when you choose a cine curve you have the ability to set the white point and black point on your curve, this in effect gives the operator the opportunity to capture the “best” 8bits. You can skew the picture towards the whites or the blacks depending on your preference. If you only have 8bits to work with there is very little headroom and it is more difficult to get the same quality.

[erikbryan]

Doesn't the DV format only support 8bits on the tape?

[ChuckS]

Which 8bits?

[Slapdragon]

Doesn't the DV format only support 8bits on the tape?

Actually, DV is a 4:1:1 scheme. In television YUV color format, luma (sometimes falsely called luminance, which is a different related measure of brightness) is generated by taking 30% of the Red CCD, 59% of the Green CCD, and 11% of the Blue CCD to form the Y component. Then color is figured by taking 100% red (U) component and subtracting the Y component (R-Y) and 100% of the Blue (V) component and subtracting the Y component (B-Y) resulting in our three components, Y, R-Y, and B-Y, called "YUV" (which is of course not correct but widely used) or Y-Cr-Cb. The color component assumes that knowing any two additive primaries, the third additive primary can be deduced.

The 4:4:4 RGB information is thus converted to the Y, R-Y, B-Y component, which in D1 standard is encoded 4:2:2. This means that in any cluster of 4 pixels, two pixels are sampled for each color space component, and all four pixels are sampled for luma. When recorded, this saves a lot of space over straight 4:4:4 sampling at the loss of some color acuity, but research has found that the loss is minimal since most humans do not have perfect color cuity in the first place. 4:1:1 sampling takes each cluster of 4 pixels and samples luma four times, and color once, making this type of recording take up only 50% of the room of an RBG image at the expense of noticeable loss of color acuity (it should be noted though that most film stocks likewise have less color acuity than digital video).

The advantage of a bigger DSP is that there is more room to do math and get the most out of the limits of your color space. Just like multiple pass encoding can be used in MPEG to average the results of several methods of encoding and get a better product, a bigger DSP can do more work, within limits, on the various encoding jobs, and handle more complex tasks at a give clock speed, essential to a real time computing task such as color encoding in a video camera. Within reason, a bigger DSP means a better picture, all other things being equal. Of course maybe one side has the edge on the software side of the DSP and does not need a bigger DSP to do the same job.

[Slapdragon]

Doesn't the DV format only support 8bits on the tape?

The DSP bit depth is the size of the bit register (its pipe width) and not an expression of how the final recording is done.

[Mitch_Ives]

I did a side-by-side test of my friends "A" and the regular DVX with identical settings. Didnt see ANY difference whatsoever. I wish i would have kept those clips...

There is. Other DVX users that got an A model have expressed the same sentiment when editing the footage.

Now, will that float your boat... maybe, maybe not.

[Mitch_Ives]

Would this cause a significant difference in quality really?

Nevertheless, it is a fact a 12-bit processor can substantially increase a camera's dynamic range"

When reviewing the 100A, I was surprised that Panasonic down-played the jump from 10bit to 12bit. Yes, this makes a huge difference in the larger cameras, not only in dynamic range, but color fidelity. Making an 8bit camera in 2004 seems a bit strange... whether it's a small one or a large one...

[ChuckS]

The digital film industry went through a similar argument in 1992 when Quantel tried to convince everyone that the 8-bit Domino was adequate for film out – Quantel was wrong and the Domino died a slow painful death.

More bits = better quality – Period! :D

[nullphonic]

I've contacted Canon by email, hopefully I'll get a response. All of the arguments and information are very well grounded and thanks to all for the input, I just can't understand if it's truly incomparable why there's such a fundamental exclusion here. I know that the real proof is in the footage but a blatant specification gap such as this is puzzling.

I'll post the hopefully-received response if/when I get it...

[Slapdragon]

I've contacted Canon by email, hopefully I'll get a response. All of the arguments and information are very well grounded and thanks to all for the input, I just can't understand if it's truly incomparable why there's such a fundamental exclusion here. I know that the real proof is in the footage but a blatant specification gap such as this is puzzling.



Remember though the Megahertz myth. Intel, in selling its processors against IBM and AMD, started this myth, which makes sense on the surface, that the biggest and most important rating any processor can have is its clock speed. This of course was a total myth, since Intel's high end chips spent years below 1 ghz an no one in the server industry complained. But the audience caught on, and Intel obliged by producing shallow data pipes at amazing clock speeds that simply did not perform well against slower AMD and IBM chips.

Once AMD and IBM come into striking distance with chips that have nearly the clock speed and are more powerful to boot plus are not heat and energy hogs like the Intel P4, Intel reverses course and announces that the megahertz myth actually exists, and claims it will stop listing processors by clock speed.

The point being that in raw number crunching, a 1ghz G4 and a 1.4 ghz AMD were superior to the 2ghz P4, but the number was not a good measurement of the quality and power issue. Sort of like rating cars by the horsepower they generate and trying to argue that this means speed, while off the line lots of other issues like rear axle mass, torque, and so forth matter, and horse power is in fact a detriment if it is achieved at the expense of all other factors.

[Kidster]

I find it hard to believe that a Company like Canon with as many of those high tech guru's that think up all of this technology would wait all this time to unveil their latest and greatest and leave out a thing like not enough bits for DSP to obtain high quality video. There has to be a reason for not including 12 bit DSP. Can't be a cost thing could it? If so, how much dramatically more would it be to upgrade from 8 bit to 12 bit . Perhaps the 4 bit difference, really doesn't make a difference. They copied the 24p from Panasonic, in order to compete on the DVX level. So why not copy the entire spec line of the DVX and just add a few extra features. I just don't think they would engineer their "last SD" camera and hope to get the milage out of it, if it were not atleast (on paper) as good or better than their direct competition. In addtion be more expensive than the DVX.

[nullphonic]

Slapdragon: Great point
Kidster: I agree

Good news: Someone got back to me quickly!
Bad news: I'll have to dig further (wrong channel, my fault)

Reply...

"Thank you for contacting Canon product support.

Decisions are made by our engineers and product designers in Japan,
based on a number of factors. Consumer demand, technological
limitations, and customer feedback are all taken in to consideration.
Unfortunately, I am unable to provide you with an exact reason why this
particular feature is what it is on the new XL2.

Thank you for choosing Canon.

Sincerely,

Johnny
Product Support Representative

Special Note: Certain issues are very difficult to resolve via e-mail. If your question remains unanswered after you have received this e-mail, the direct assistance of a Canon telephone support representative may be required."

I'll call then. (I'm an info geek)

[Slapdragon]

Dear Johnny,

Unfortunately, certian purchase choices are made elsewhere than in the frontal lobe of my brain. *If your organization wont let the sales guys talk to the engineer guys, my organization wont let my payment guys talk to my camera buying guys.

Thanks for nothing,
A consumer.

[nullphonic]

LOL ;D

[auteur0]

Hi all.

This is my first post on this forum, and it just happens to fall well within
my range of knowledge, as I am an engineer (h/w, s/w, f/w and a struggling filmmaker), and do have some knowledge of DSP and microprocessor architectures. Finally, a question I might be able to help with.

In general, a 12-bit DSP can be considered a "better" DSP than an 8-bit DSP if what you are talking about is the "width" of its instructions path, that is, the width in bits of the numbers used to represent its instruction set. I say in general, as in the end, it all boils down to the software that was written to run on the device and how good it is at crunching the incoming digital video data. In many cases, it is possible to write code that performs a set of algorithms well on a DSP or processor of any instruction width, that is to say, that depending on the requirements of the algorithm, one might be able to write code that will run equally well on an 8-bit, 12-bit, 16 bit, or 32-bit architecture. In those cases, the additional instruction width doesn't gain you much. The algorithm being implemented just doesn't need to make use of the extra bits.

So, why the additional bits? Often times it boils down to running those instructions more quickly, by packing more information into each instruction. At other times, having the additional bits means that you can have access to additional instructions to implement algorithms that would be prohibitively difficult to manage in a smaller architecture. I would venture, that the engineers at Canon decided that they could get all they're algorithmic processing needs met at an acceptable level with a smaller, cheaper architecture, and thus spend the engineering dollars on something that would have a greater marketing impact like a better, more expensive lens. Usually, there is a trade-off with speed, and power, and heat dissipation in this kind of decision as well. But you can rest assured, that if the engineers didn't feel that their s/w could run effectively on a smaller DSP, they would have argued for a bigger, "better" one.

As for the anecdotal evidence that a bigger DSP means a better picture, that isn't always the case. Again, it boils down to the software -- what you do in code with that extra instruction width. With a bigger DSP you might be able to write a finer, faster piece of code that implements a better imaging algorithm, and therefore, get a better picture, but you could also just take the same code that ran well on a lesser DSP, and find that it now just runs a little bit faster. Again, it depends on the algorithm being implemented, and the constraints of that algorithm.

More important than the DSP alone, is the width of the Analog-to-Digital Convertor, and the complexity of the algorithms you want to use. A wider ADC bus generally means more accurate sampling, and better dynamic range. Coupled with efficient visual processing algorithms, and you can get a pretty nice picture, processed quickly.

All things then considered, the DVX is an extraordinary piece of hardware, and I would venture, so is the new Canon.

Just my $.02.

-Omar Green

[hvpz]

Hi Auteur0,

12 bits means nothing if you don't precise 12 bits of what...

It could be the address bus, the size of registers or even the size to code the instructions.

For example, the good old motorola 68000 had a 24 bit adress, 32 bits registers and a 16 bit alu and data bus.

---

btw, keep on struggling for film making...

[auteur0]

Agreed.

that's exactly what I meant, and why I limited myself to talking about instruction width, and data bus width. Because we're talking about an embedded DSP, post 2K, chances are the engineers at Canon aren't describing the address bus width -- memory is pretty cheap these days.

Even though in general, the marketing departments of electronics firms like to confuse people (to sell more products), 12-bits of anything doesn't mean much, even after it's specified. What they should really be talking about is which DSP algorithms they've implemented, what the dynamic range of the ADC is, and what they're doing to eliminate noise.

They might also want to specify how much of the data captured from the CCD is really used in the algorithms. Anything less than that is just marketing fodder -- no disrespect to marketing folks, I'm soon to be one myself (in my day job).

BTW -- thanks for the encouragement. I did the starving artist thing in NYC for 4 years after 10 years in Engineering, only to leave that town with a manager, a couple of short films, and a couple of semi-successful rewrite jobs. Been in the Boston area for 18 months writing code, and now I'm off to Cambridge in the UK for 2 years. Better film center there than in Boston, and I'll finally be able to crew up like I want and shoot some off-kilter stuff with my DVX.

Any UK folks on this forum?

-Omar

[hvpz]

Shane Carruth was a software engineer before he said STOP.

He wrote his screenplay, shoot it (s16 color) and sent it to Sundance. He won.

Of course, things as not as simple as just saying : He sent it...

There is a lot of work, talent, a strong will behind his success.

But, this is a message of hope for all people trying to escape from
real world to live with an for their passion.

:)

[ecna]

as long as we're talking about bits not meaning much unless they're in a very specific context, I'd like to point out that the Atari Jaguar was the first 64 bit console, and had two processors. It also had crappy games, weird color problems, and I'm pretty sure the people who made the games and the people who made the hardware weren't on speaking terms.

that said, I own two, and about 20 games... sigh... classic gaming junky.

[ullanta]

Just to be clear, maybe simplify:

I assume that 12-bit DSP implies 12-bit sampling.

Gamma curves map CCD-sampled intensity levels to 8-bit DV intensity levels.

With selectable Gamma curves, there's much more flexibility and accuracy if you are downsampling from 12 bits than if you are distorting 8 bit samples.

Somewhere someone posted pictures comparing gamma curves on DXV100P vs. A... this clearly showed the differences in both highlight and shadow detail, which is what you'd expect from the 4-fold DSP resolution jump. Some of this may have to do with soft knees, too, but the pictures are compelling.

[nullphonic]

Another thread has started that ends this one I guess (XL2 is indeed 12 bit and not 8 bit)...

http://www.dvxuser.com/cgi-bin/DVX2/YaBB.pl?board=XL2;action=display;num=1092325895