Canon EOS 5D Mark IV To Feature 4K Video?
- Thread starter rs
- Start date
msm said:
The actual physical prototype 120mp APS-H sesnor that Canon actually produced, tested, and gathered data for quantify the real performance.
Upvote
0
msm said:
From the horses mouth:
This is one of many a few press releases, actually. There is another one that describes some of the more technical aspects, describing on-die processing and the like. I haven't found that one yet. To break out the important parts:
The technology Canon developed to increase readout was even increased parallelism. The description used at the time this was first announced, as well as the subsequently granted patent for DS-CP-ADC, describe something VERY similar to Sony Exmor, which uses one ADC unit per pixel column, which is different than past sensor designs, which used one ADC unit per group of columns. For example, in a camera with 8 readout channels, and 4000 columns of pixels, every ADC would be responsible for processing 500 columns of pixels, which when you factor in the row count, is hundreds of thousands to millions of pixels per ADC. With column-parallel ADC, each ADC unit is only responsible for processing a few thousand pixels. With the ADC units on the sensor, the distance between pixel and ADC unit is greatly shortened, which allows them to solve the timing issues. Thanks to hyperparallelism, each ADC unit has to do less work, so you can actually achieve faster readout at a lower frequency, this improving readout performance without hurting IQ.
Even if each pixel itself was noisier than Canon's 16mp APS-H sensor, it doesn't matter. Noise is related to total sensor area, quantum efficiency, and in small part to read noise (that only affects the deep shadows). Ignoring dynamic range for a minute (there is no information about the DR of the 120mp APS-H sensor, so I honestly cannot speak to it), if you downsample a 120mp APS-H image to the same dimensions as a 16mp APS-H image, the per-pixel noise is going to average out. Since the two sensors have the same total area, there is unlikely to be any measurable differences. Given the 120mp sensor used a better readout system, I'd be willing to bet good money that it actually had the better noise characteristics. And there is absolutely no question it would have much sharper, clearer details.
I'll see if I can find one of the other press releases, the more technical one, and share it. I'm not trying to mislead anyone. I really try not to assume, whenever possible everything I say is based on some fact or piece of official data somewhere. I have a bit more depth of knowledge than just what this one press release offers because I've read everything there is to read about things like Canon's 120mp APS-H sensor (and plenty more from other sensor manufacturers), so I have a larger body of knowledge to draw from.
I really honestly do believe that Canon's 120mp APS-H sensor, which does actually exist in prototype and uses some of Canon's still photography patents, could be one hell of a powerhouse for IQ. Not necessarily greatly reducing noise...but massively increasing detail and sharpness, either allowing photographers to print really large without having to upsample, or by allowing significant improvements in overall IQ simply by downsampling.
Upvote
0
dilbert said:
Canon used a smaller process for this. From what I recall about either a press release or some other specs listed somewhere, they sensor was actually stitched together from separately manufactured parts. (A lot of Canon's prototype sensors are actually made that way, by stitching multiple separate fabricated parts into a single device. Their ultra high sensitivity 0.1 Lux large format sensor, for example, is manufactured that way as a simple matter of necessity.) I suspect they fabricated it with the same fab they produce their small form factor sensors with. Chipworks verified years ago that Canon has had a 180nm copper interlink process, one even capable of producing sensors with light pipes, for a while now.
So yes, you are correct, Canon's current APS-C and FF sensors are built on a 500nm process, which wouldn't support a sensor like this (well, it could...it's just that the photodiodes would be really tiny and therefor the fill-factor, the total actual light-sensitive area, of the sensor would be lower than an identically sized sensor with larger pixels). However there was a rumor not long ago that Canon was revamping its fabs, moving to larger wafers. Either they are repurposing some of their existing fabs for the small form factor stuff, or they are building new fabs to expand their 180nm fab capacity.
Upvote
0
LetTheRightLensIn said:
I hope Canon is "smart enough;" they certainly need to be at this point in the game. They're losing so many to the Gh4 and A7s, and those still holding out for Canon's "answer" are going to jump ship if it doesn't deliver. But, like you, I wouldn't be remotely surprised if Canon fails epicly.
Conversely, they could just make the 1D C $3,500. That video set at that price point is what they need to give. The problem is that on the stills side, it's the 1d X. I'd be all for Canon dumbing down a 1D C on the stills side (to more of mark-III capabilities, maybe even less), and then rebranding it as a baby 1d C. Or, better yet, put it into a 5d body and call it the 5D C. I honestly don't need any new tech than what is already in the 1d C.
Upvote
0
jrista, I don't see how this press release quantify performance. It says nothing about SNR or DR etc, it doesn't even try to claim that the sensor is competitive in those regards. It may also require hardware around it which is not feasible nor practical in todays cameras. For instance to read out and process all that data would require a lot more readout channels and processing power than what you see in a 1DX today.
Press releases may also typically be written by PR or marketing personnel written for other purposes than to scientifically describe their findings.
As for patents, I don't read them but they don't actually give any data about how well actual their actual implementations perform do they? Without that data we can not tell if its awesome or not. What seems great on paper might be bad in practice.
For instance Foveon sensors seem like a great technology on paper does it not? No CFA wasting away 2/3rds of the light and no demosaic algorithm interpolating data and making images soft in 100% view. Yet in real life Foveon is outperformed by standard CFA sensors, it gives the resolution but does not perform well in other aspects. Real life performance is what counts and Foveon sensors don't have it (yet, would like to see that change).
Press releases may also typically be written by PR or marketing personnel written for other purposes than to scientifically describe their findings.
As for patents, I don't read them but they don't actually give any data about how well actual their actual implementations perform do they? Without that data we can not tell if its awesome or not. What seems great on paper might be bad in practice.
For instance Foveon sensors seem like a great technology on paper does it not? No CFA wasting away 2/3rds of the light and no demosaic algorithm interpolating data and making images soft in 100% view. Yet in real life Foveon is outperformed by standard CFA sensors, it gives the resolution but does not perform well in other aspects. Real life performance is what counts and Foveon sensors don't have it (yet, would like to see that change).
Upvote
0
Lee Jay said:
Not really. You have to take it in proper context. For large sensors, APS-C and larger, I don't know of any that use a process smaller than 180nm. The ultra tiny form factors, the sensors that are a fraction the size of a fingernail, are the sensors that use very small fabrication processes, but even those aren't eight generations more advanced. I think a 65nm process is used for sensors with 0.9µm and the upcoming 0.7µm pixel sizes. That would be three generations smaller transistor size (180nm: -> 130nm -> 90nm -> 65nm).
You also can't use CPU transistor fabrication technology as a basis of comparison. They are primarily on 22nm, with 14nm parts supposedly due this year (maybe they are already here, haven't looked into it.) But that is a whole entire different market. We know that Canon's DIGIC 5 used a 65nm process, manufactured by Texas Instruments I think. But that's still a processor. You can't mix that with sensor tech.
At the moment, I think 65nm is the current smallest fabrication process used for image sensors. The next step would be 45nm, and I've read a couple patents that describe sensors with 0.7µm (700nm) pixel sensors that would be fabricated with a 45nm process, but I haven't actually seen anything yet that indicates it's being done. Even if there were sensors being manufactured with 22nm gates, that is six generations...not eight. Given that larger form factor sensors don't even remotely need a 65nm gate to be highly efficient, I wouldn't say that a 180nm process is out of date for APS-C and FF sensors. If it is out of date, it would only be out of date by one generation, 130nm.
I did read about some high sensitivity sensors recently called SPADs, or Single-Photon Avalanche Diodes, which are designed for specialized purposes (medical imaging and such like PET, FLIM, etc., scientific imaging, astrophotography, etc.) These are pretty bad-ass CMOS devices with ultra high sensitivity (basically photon counters). They have been fabricated on 180nm, 130nm and 90nm processes. They generally seem to be 130nm parts, are usually fairly small sensors (smaller than APS-C), with pixel sizes maybe a little bit smaller than current APS-C parts. They have a specialized pixel structure, but overall are not all that much different than your average CMOS sensor. These are CUTTING EDGE devices...really cutting edge. They do their job extremely well, and really don't need transistors smaller than 90nm. It actually seems smaller processes actually make it more difficult to fabricate these high end sensors than the larger processes.
So I really don't think that 180nm is old and out dated, not for the size of sensors were talking about.
Upvote
0
msm said:
SNR and DR aren't the epitome of sensor performance, though. They are only factors of sensor performance. Both are heavily affected by readout noise, and it's been demonstrated that column-parallel ADC designs produce less read noise, by at least two companies now (Sony and Toshiba, and I believe other high end sensor manufacturers have similar designs in the works as well). Canon described some kind of hyperparallel on-die ADC for the 120mp APS-H.
Assuming the silicon process was the same generation as the cameras of the time it was released, it's logical to assume it has the same fundamental characteristics as the 1D IV. The 1D IV had around 45% Q.E. and the same DR limitations as all Canon cameras (due to read noise). I see no reason to assume this sensor would be significantly different in those fundamental statistics at worst, better if their highly parallelized readout offers similar improvements as Sony and Toshibas. Canon silicon hasn't really changed much over the years...the most significant improvements each generation are a few percent jump in Q.E.
Your also misunderstanding the point of using a column-parallel ADC. You actually DON'T need as much processing horsepower to read out more pixels faster when you hyperparallelize the ADC units. The problem with having too few units is each unit MUST be high powered enough to handle the hundreds of thousands or millions of pixels they have to process. That means higher frequency, and it also means more attention must be paid to the design of those units to limit the amount of noise they add to the signal (and even then, they are noisy parts because of the high frequency).
By using one ADC unit per column, each ADC can operate at a lower frequency. The lower frequency immediately offers a benefit in terms of read noise. Other techniques, such as moving the clock and driver off to a remote area of the die (like Exmor), you can reduce noise even further (Exmor took it one step farther, and used a digital form of CDS, which they claim was better than using analog CDS...however ironically they added analog CDS back into the mix with later version of Exmor for video cameras...now they do both analog and digital CDS). You trade die space for the ability to operate at a lower frequency and power. With a 180nm process, that's a no brainer. This HAS BEEN DONE...both Sony and Toshiba have working CP-ADC designs built into their CMOS sensors that are actually used in consumer products. Sony has a number of technical documents that explain how they achieved exactly what Canon describes in their 120mp APS-H papers and patents...low power high speed readout of high resolution sensors via hyperparallel ADC.
So, even though Canon's 120mp APS-H isn't in an actual consumer grade product that we can buy, it uses technology that mirrors products from other brands that we can buy, and that have been tested. The most telling are Sony security video cams that use Exmor sensors, which can operate at very high frame rates in very low light...they are not only doing high speed readout with very, very low noise and relatively high DR, they are also doing processing with image processors that are packaged to the bottom of the sensor, and wired directly to it.
To be strait, I am speculating a bit, but it's very educated speculation. It isn't like it's just 100% completely unfounded drivel.
msm said:
As for Foveon, I think your incorrect in your assessment. Foveon only "fails" at ONE thing: resolving power. There have been debates in the past on these forums where Foveon fans claim that because it has a 100% fill factor for all colors, that it has as much or higher RESOLUTION than bayer sensors. Those claims are wrong, as bayer sensors get largely the full benefit of the raw sensor resolution in terms of luminance...they only really suffer in color resolution and color fidelity (both areas where Foveon excels).
For what Foveon is, at it's REAL spatial resolution, they are actually very good. Their red channels are a little nosier, but their blue channels are less noisy than bayer. No surprise, given the layering order of color photodiodes in the Foveon. Even though image dimensions/resolving power for Foveon is lower than in bayer sensors, those smaller images usually exhibit high quality. I do think that color fidelity with Foveon cameras is superior to what I get with my Canon DSLRs (I just like my resolution too much to give it up
). So I think it's unfair to claim that the real-life performance of Foveon is bad or even poor. For what it is, it's real life performance is very good.The only drawback of Foveon is it's resolving power...and I truly believe that Sigma has done Foveon a big disservice by trying to upsell it as having more resolution than it really does, or somehow claim that because it gathers full color information per pixel that upsampling it somehow beats bayer sensors for resolution and detail. Actual real-world examples that do exactly that have proven otherwise. Foveon's problem isn't that it's bad technology...it's that Sigma owns it, and Sigma doesn't have the marketing power nor the R&D budget to really make Foveon shine and become a highly competitive alternative. Sigma is much more a lens company than a camera or sensor company, IMO. I do believe it COULD be highly competitive in the hands of a wealthier corporation that could more richly fund it's development.
Upvote
0
Similar threads
