carstenw wrote:
Are you referring to any specific hard law of physics, or just invoking the hard law of physics argument in a hand-wavy kind of way to "support" your argument?
I am referring to shot noise (also known as photon noise) which is a consequence of the packetization of electromagnetic energy into photons and thermal noise (aka dark current) which is hard to get rid of without carrying vats of liquid nitrogen around with you. There is also going to be read noise, but it's more amenable to technology.
KaaX wrote:
If you look at DxO tests (and their sensor tests are much better and more believable than their lens tests), the Sony APS-C 16Mp sensor (as in, e.g. NEX-5N) is almost exactly 1 stop worse than the 5D2 sensor. Remember, the 5D2 sensor is four years old. And given the hard limits that physics imposes, I don't expect a m4/3 to get there ever.
An m4/3 system looks great if (1) you value small size and light weight; and (2) don't have to shoot in low light.
Note the Sony 16MP sensor is 2 years old as well, the revision time is offset between the Sony and Canon sensors. And frankly at ISO 1600 the Sony sensor is pretty much dead at the 30dB limit on DXOptics, which really is the point where noise is irrelevant, The 5DmII is better, but that's just giving it an extra stop before the 30dB limit is reached. Note that while the 30dB limit is arbitrary, it does track pretty well with the point where noise becomes a concern in general use in my experience (for the non-engineers it's where the Signal to Noise ratio is 1000 times).
The m43 sensors aren't there yet, but the Panasonic 16MP sensor shows performance just about dead on the Sony 12MP sensor, which is the generation previous to the 16MP sensor and supports my assertion that the next generation of Panasonic m43 sensors will have performance comparable to the current Sony 16MP APS-C sensor and therefore entirely acceptable performance at ISO 1600 (at that point I expect current APS-C sensors to be at or above the 30dB point at ISO 3200 and FF to be hitting 30dB at ISO 6400).
mawz wrote:
The m43 sensors aren't there yet, but the Panasonic 16MP sensor shows performance just about dead on the Sony 12MP sensor, which is the generation previous to the 16MP sensor and supports my assertion that the next generation of Panasonic m43 sensors will have performance comparable to the current Sony 16MP APS-C sensor and therefore entirely acceptable performance at ISO 1600 (at that point I expect current APS-C sensors to be at or above the 30dB point at ISO 3200 and FF to be hitting 30dB at ISO 6400).
So, do you think there are limits to this technological progress? If the next generation of FF sensors will have 30dB SNR at ISO 6400, will the one after that reach the same SNR at ISO 12800? and the next one at ISO 25600?
KaaX wrote:
So, do you think there are limits to this technological progress? If the next generation of FF sensors will have 30dB SNR at ISO 6400, will the one after that reach the same SNR at ISO 12800? and the next one at ISO 25600?
There certainly are physical limits, I just don't think that we're hitting them yet, or even close. We certainly aren't even close to the process limits (which are in the 10nm range these days, sensor wells are 3 orders of magnitude larger).
Frankly people were saying we were near the limits back in the 12MP FF sensor days. We weren't and we aren't now. Once we see 2-3 sensor generations without any significant changes then we'll be at the limits.
Note that we may see a move away from pure Bayer sensors though, there are technologies available which increase the luminance sensitivity of sensors like Kodak's interspersed L pixels.
According to the laws of physics, man isn't able to fly. I've just spent a few days at the Singapore Airshow and seen not one but several indications that, with the help of modern science, man may indeed be able to fly. I might even try it myself this afternoon instead of hiring an elephant to get back to Bangkok. If it's successful, I'll probably buy an E-M5 as well later this year, in spite of the laws of physics
Jorgen Udvang wrote:
According to the laws of physics, man isn't able to fly. I've just spent a few days at the Singapore Airshow and seen not one but several indications that, with the help of modern science, man may indeed be able to fly. I might even try it myself this afternoon instead of hiring an elephant to get back to Bangkok. If it's successful, I'll probably buy an E-M5 as well later this year, in spite of the laws of physics
Only Newtonian physics. Modern physics describes fixed-wing flight very well (Helicopters on the other hand remain problematic).
mawz wrote:
There certainly are physical limits, I just don't think that we're hitting them yet, or even close. We certainly aren't even close to the process limits (which are in the 10nm range these days, sensor wells are 3 orders of magnitude larger).
Frankly people were saying we were near the limits back in the 12MP FF sensor days. We weren't and we aren't now. Once we see 2-3 sensor generations without any significant changes then we'll be at the limits.
Note that we may see a move away from pure Bayer sensors though, there are technologies available which increase the luminance sensitivity of sensors like Kodak's interspersed L pixels. ...Show more →
The conversion efficiency of the material that converts light to electrical charge hasn't changed significantly over the last 3 decades. It was possible to get between 50% and 80% conversion efficiency about 25 years ago. We are still in that area. Then we have the percentage of the sensor area that contains the conversion material, the size of the photosite relative to the sorrounding circuits, and how the light is conducted into them. There we have seen a significant progress for small pixels. There has also been a progress in the color filter development, but often at the cost of color fidelity. But even with perfect technology with 100% efficiency in all steps, we cannot get any more than about 2.5 stops improvement from where we are now.
Well, this is one huge lens in a stable of little lenses, whereas for NEX you have one little lens and the remaining all being huge!
alemmo wrote:
this camera does look nice but I don't think the M4/3 crowd can make fun of the "huge" Nex kit lens anymore. That 12-50mm is massive looking on the OM-D
alundeb wrote:
The conversion efficiency of the material that converts light to electrical charge hasn't changed significantly over the last 3 decades. It was possible to get between 50% and 80% conversion efficiency about 25 years ago. We are still in that area. Then we have the percentage of the sensor area that contains the conversion material, the size of the photosite relative to the sorrounding circuits, and how the light is conducted into them. There we have seen a significant progress for small pixels. There has also been a progress in the color filter development, but often at the cost of color fidelity. But even with perfect technology with 100% efficiency in all steps, we cannot get any more than about 2.5 stops improvement from where we are now. ...Show more →
Got any links to material on that? It's a claim distinctly at odds with what I've heard or read (and frankly, it's the exact same argument, just about down to those numbers, that I was reading 6 years ago, and about 4 stops of high ISO performance ago). Frankly, the only aspect that hasn't been changing is the conversion efficiency of the pixel well. It's possible that we're getting close to the limits of CMOS-based sensors but I'd like to see some papers on it (feel free to link to IEEE Xplore, I've got access) before I accept it.
It's also provably incorrect from a whole-imager perspective (even accepting it as correct regarding the silicon). Kodak's mixed luminance/colour filter array gives 1 to 2 stops on its own (depending on the exact pattern chosen) via changes to the CFA. That leaves 0.5 stops of your claim for actual improvement of microlenses and fill factor, a hardly believable claim. If your claim is accurate for the silicon, non-withstanding the CFA implementation, then there's more like 3.5-4.5 stops actually possible via alterations to the CFA, which we're seeing start to crop up now as the other limitations of Bayer become noticable. I don't expect that the 2x2 Bayer Matrix will remain the standard for more than a few more years now that the onboard processing power of most high-end cameras has reached a point where larger pixel arrays are viable.
KaaX wrote:
Sure, we'll see *slightly* better sensors.
On the other hand, I shoot at lot at ISO1600 f/1.2 using the four-year-old sensor of my 5D2. I don't anticipate being able to shoot m4/3 in the same light in the foreseeable future...
I see your point, and I also doubt we'll see any (AF OEM) f/1.2 lenses for MFT anytime soon (if ever), and that's why I also mentioned in my post that for those situations I would grab my D700 and not my MFT system
mawz wrote:
Got any links to material on that? It's a claim distinctly at odds with what I've heard or read (and frankly, it's the exact same argument, just about down to those numbers, that I was reading 6 years ago, and about 4 stops of high ISO performance ago). Frankly, the only aspect that hasn't been changing is the conversion efficiency of the pixel well. It's possible that we're getting close to the limits of CMOS-based sensors but I'd like to see some papers on it (feel free to link to IEEE Xplore, I've got access) before I accept it.
It's also provably incorrect from a whole-imager perspective (even accepting it as correct regarding the silicon). Kodak's mixed luminance/colour filter array gives 1 to 2 stops on its own (depending on the exact pattern chosen) via changes to the CFA. That leaves 0.5 stops of your claim for actual improvement of microlenses and fill factor, a hardly believable claim. If your claim is accurate for the silicon, non-withstanding the CFA implementation, then there's more like 3.5-4.5 stops actually possible via alterations to the CFA, which we're seeing start to crop up now as the other limitations of Bayer become noticable. I don't expect that the 2x2 Bayer Matrix will remain the standard for more than a few more years now that the onboard processing power of most high-end cameras has reached a point where larger pixel arrays are viable....Show more →
Well, peak quantum efficiency is somewhere around 80%, I believe (it greatly depends on the wavelength and so is considerably less for red and blue) and there doesn't seem to be any reason to expect it to improve much in the near future.
Note that all our discussion about noise implies unchanging pixel density and so unchanging resolution. You can, of course, gain a lot of SNR by making pixels larger and sacrificing resolution for that. A mixed luminance/color filter array would do a very similar thing -- it would sacrifice color resolution to get better SNR. But this is not a pure win -- it's a trade-off. You will be able to get massively better SNR by ditching the CFA completely and making a luminance-only sensor (for a while B&W photographers had hopes for something like this) but this again is a different game.
Yet another way to improve the SNR is to increase the transparency window of the CFA filters but there's a price here as well -- you "muddy up" the colors because you can't distinguish them as well now. It's a balancing act and another trade-off. Same thing for replacing the Bayer matrix with some other kind (e.g. Fuji's 3x3 matrix) -- you gain something, but you also lose something. TANSTAAFL, I'm afraid.
What I think we'll see is clearer specialization -- some sensors will explicitly go for low noise and other sensors will go for other things, notably high resolution. But the easiest way to make this trade-off is still the simplest: change pixel density.
wjmeyer wrote:
I see your point, and I also doubt we'll see any f/1.2 lenses for MFT anytime soon (if ever), and that's why I also mentioned in my post that for those situations I would grab my D700 and not my MFT system
I would have made such an argument, but unfortunately for it there's this:
Oh yes, I forgot about those beautiful Nokton's They sure are nice, my mind was set on AF OEM glass... will Olympus make a 25mm f/1.2 to equal the Canon L version, I doubt it, but for manual glass, well, it's pretty cool to see some of this stuff.
mawz wrote:
Got any links to material on that? It's a claim distinctly at odds with what I've heard or read (and frankly, it's the exact same argument, just about down to those numbers, that I was reading 6 years ago, and about 4 stops of high ISO performance ago). Frankly, the only aspect that hasn't been changing is the conversion efficiency of the pixel well. It's possible that we're getting close to the limits of CMOS-based sensors but I'd like to see some papers on it (feel free to link to IEEE Xplore, I've got access) before I accept it.
It's also provably incorrect from a whole-imager perspective (even accepting it as correct regarding the silicon). Kodak's mixed luminance/colour filter array gives 1 to 2 stops on its own (depending on the exact pattern chosen) via changes to the CFA. That leaves 0.5 stops of your claim for actual improvement of microlenses and fill factor, a hardly believable claim. If your claim is accurate for the silicon, non-withstanding the CFA implementation, then there's more like 3.5-4.5 stops actually possible via alterations to the CFA, which we're seeing start to crop up now as the other limitations of Bayer become noticable. I don't expect that the 2x2 Bayer Matrix will remain the standard for more than a few more years now that the onboard processing power of most high-end cameras has reached a point where larger pixel arrays are viable....Show more →
I guess our disagreement on the remaining improvement potential being 2.5 or 3.5 stops boils down to what we define as the current state. We have some data sheets for sensors with about 25-30% response for green light, but I can't find data sheets for sensors with noise based QE estimates around 60%. These estimates include fill factor. So if you claim that there are 3.5 to 4.5 stops remaining, the estimated QE of 60% must be entirely wrong.
Noise based QE estimates ( you have probably seen it):
wjmeyer wrote:
I see your point, and I also doubt we'll see any f/1.2 lenses for MFT anytime soon (if ever), and that's why I also mentioned in my post that for those situations I would grab my D700 and not my MFT system
f1.2? There's already 2 f0.95 lenses in the system
alundeb wrote:
I guess our disagreement on the remaining improvement potential being 2.5 or 3.5 stops boils down to what we define as the current state. We have some data sheets for sensors with about 25-30% response for green light, but I can't find data sheets for sensors with noise based QE estimates around 60%. These estimates include fill factor. So if you claim that there are 3.5 to 4.5 stops remaining, the estimated QE of 60% must be entirely wrong.
Noise based QE estimates ( you have probably seen it):
Thanks for the links, they make for some interesting reading.
Note that aside from the D3s with its QE of 57% (and its legendary high ISO performance), all the large sensor cameras there are at or under 50% QE. 40-45% seems to be typical for the best performing cameras. Based on those numbers it does look like there's some significant areas available for improvement in high-MP FF cameras, which tend to have lousy minimum read noise and QE. From that data I'd say we're probably getting near the limits of what APS-C and low-MP/high ISO oriented FF sensors can do but nowhere close when it comes to high-MP FF where the QE is typically 30-35% and the read noise is typically around 4-5e or higher.
Also remember that I'm getting 1-2 stops of that 3.5-4.5 claim via altering the CFA's to add luminance pixels, those 1-2 stops are available irregardless of the limitations of the silicon underneath. CFA's have a fairly significant effect on the light sensitivity of the sensor. We're finally starting to see workable alternatives to the 2x2 Bayer and I expect we'll see more of those soon enough.
They look really good to me as well, I can't wait to get the E-M5 in my hands to play with on my own and do my own testing. It can be so hard to tell from other image samples when you don't know much about the situation, but when you use it in real life situations (for you) that is the best way to tell.
You know that gets me thinking about a lot of the discussion that has gone on in this thread, at the end of the day, what matters most is not all the technical specs and how a sensor or a lens is made, what matters is how you feel about the images that you take. If you are unsatisfied with your images, either your technique is off, your equipment is not on par with your expectations and/or the equipment can't produce what you are looking for, but ultimately you have to decide what works best for you, and to me, that's part of the fun of it all
