Fun with Bayer interpolation

[Ron Parr]

Well, you may be disappointed with the 1st generation of Foveon
technology. Go to http://www.foveon.net/docs/F7_Datasheet.pdf . Look
at the part where it says "ISO: 100 Full Resolution mode". Think
on what the "variable pixel size" implies for resolution.

Amplification for higher ISOs is done off-chip for all sensors, so it wouldn't make sense for them to include higher ISO numbers in the specs for the chip.

But the SD9 does NOT support JPEG in camera! Again, look at the
Sigma page. You will have to make significant resolution tradeoffs
to get lower storage sizes or frame rates.

Foveon rates their chip at 2 FPS in full resolution mode. Sony's new 6MP CCD does 3, so you are indeed getting fewer frames per second off the chip. I can't predict what Sigma will do with their electronics once the data are off the chip. However, I will point out that a 3MP camera with jpeg compression must have enough bandwidth to handle 24 bit per pixel images as well since after demosaicking, the the image must be handed off to the jpeg compression algorithm in a full 3-dimensional colorspace. Thus, the bandwidth requirements once the data are off the chip should be the same - unless you're worried about a fast RAM buffer to store the data before they are processed in any way. 3FPS would be about 30 MB/S, or about the level of a low-end 486 processor using now obsolete memory technology.

--Ron ParrFAQ: http://www.cs.duke.edu/~parr/photography/faq.htmlGallery: http://www.pbase.com/parr/

 

[Erik Magnuson]

Amplification for higher ISOs is done off-chip for all sensors, so
it wouldn't make sense for them to include higher ISO numbers in
the specs for the chip.

Are you sure about this? The block diagram for the F7 shows the A/D is integrated into the device and you only get digital output (which is not the case for a typical Sony CCD.) Exactly where are you going to do the amplification?

Foveon rates their chip at 2 FPS in full resolution mode.

In what they call "rolling shutter mode". It's not clear what this means or if this rate is achievable if you are also using a mechanical shutter.

new 6MP CCD does 3, so you are indeed getting fewer frames per
second off the chip.

The poster that I was replying to has a 1D that has over twice that rate. He may have bought his 1D for other reasons than frame rate, but I was just pointing a situation where his perfromance would be degraded.

the same - unless you're worried about a fast RAM buffer to store
the data before they are processed in any way.

Fast RAM is easy, although the X3 (as in needing x3 more) applies here. Either more ram needed or fewer shots per burst.

30 MB/S,

Ok, so you have it in buffer RAM. Now what? You have to get it to a storage device. CF write speeds range from 2.5MB/s (1D) to .5 to 1 MB/s (D30) to .3 - .5 MB/s (995). You are finally agreeing with me that this camera is going to need a state-of-the-art or better I/O system.
--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Ron Parr]

Amplification for higher ISOs is done off-chip for all sensors, so
it wouldn't make sense for them to include higher ISO numbers in
the specs for the chip.

Are you sure about this? The block diagram for the F7 shows the A/D
is integrated into the device and you only get digital output
(which is not the case for a typical Sony CCD.) Exactly where are
you going to do the amplification?

Where? I'm looking at the PDF and I don't see this.

Foveon rates their chip at 2 FPS in full resolution mode.

In what they call "rolling shutter mode". It's not clear what this
means or if this rate is achievable if you are also using a
mechanical shutter.

The "rolling shutter" refers to the readout mechanism and means that it's not buffered. Using a mechanical shutter would help reduce distortion introduced by the rolling shutter when photographing fast moving objects. These two are completely compatible.

In any case, it's not clear from the datasheet if the rolling shutter mode is referrnig to the 25FPS mode or both modes.

Ok, so you have it in buffer RAM. Now what? You have to get it to a
storage device. CF write speeds range from 2.5MB/s (1D) to .5 to 1
MB/s (D30) to .3 - .5 MB/s (995). You are finally agreeing with me
that this camera is going to need a state-of-the-art or better I/O
system.

All cameras fill up their fast memory buffer and slow down when writing to memory. What we've shown is that buffering speed is not an issue and that jpeg compression (if done in a future Foveon based camera) will take exactly the same speed as jpeg compression for Bayer pattern sensors.

In the case of the SD9, which only does RAW, you're claiming that it will need "state of the art" I/O which seems to require the following:

1) The assumption that jpeg compression is not currently a bottleneck on other cameras.

2) Sigma will feel obligated to write to flash memory faster and that this will be expensive or difficult to do.

So, what's the meat of your claim: When writing RAW files with a Foveon based camera, it will take longer to flush the buffer (assuming no other bottlenecks, e.g. compression, and using the same parts) than it would it would if were writing RAW files from a Bayer pattern sensor.

For this, you are continually interrupting otherwise rational discussion?

--Ron ParrFAQ: http://www.cs.duke.edu/~parr/photography/faq.htmlGallery: http://www.pbase.com/parr/

 

[Erik Magnuson]

Where? I'm looking at the PDF and I don't see this.

That's because I'm wrong here. I thought I was looking at the F7 datasheet when I was actually looking at the F10. (It's the F10 that has the on-board A/D.)

However, Sigma only rates the SD-9 at ISO 400 max (and we've seen no Foveon samples from ISO other than 100.) Another odd fact is that Foveon claims the one of the advantages of X3 is that it can gather more light because there are no filters thus it will be MORE sensitive than Bayer cameras. So why is the 1st camera is actually slower than its competition?

In any case, it's not clear from the datasheet if the rolling
shutter mode is referrnig to the 25FPS mode or both modes.

It looks like it applies to both (and the amount of data xferred scales appropriately.) It's again just curious that Sigma does not have an FPS rate on the SD-9 specs.

All cameras fill up their fast memory buffer and slow down when
writing to memory.

And I've shown that the problem is X3 worse! I've never said that it was impossible to solve. I've only said it would either be more expensive or lower performance than equivalent Bayer cameras.

What we've shown is that buffering speed is not an issue

Who ever said it was? Can't you read? IT'S WRITE TO CF TIME that will be the issue.

and that jpeg compression (if done in a future Foveon
based camera) will take exactly the same speed as jpeg compression

Actually, we don't know this. I agree that it SHOULD be true and that is why I find the SD-9 RAW output only and the "killer app for a new computer" issues so puzzling.

1) The assumption that jpeg compression is not currently a
bottleneck on other cameras.

It's not an assumption. Look at the write times for current cameras. Show me one that is faster (FPS) for RAW than JPEG.

For this, you are continually interrupting otherwise rational
discussion?

Hmm, in a discussion about noise vs. ISO, and output frame rate of X3 vs Bayer, I point out that the published specs for X3 ships cameras seem to imply lower performance than current Bayer cameras in these areas. And you think this is irrational and disruptive? Oh, I get it. Anyone who disagrees with you cannot be rational. If you find it so irrational and disruptive then don't read my comments.
--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Erik Magnuson]

But I clearly see the
differences between "fundamental" problems vs. "implementation"
problems.

Wow. You can spot this w/o seeing a single implementation. Just PowerPoint slides and a bunch of press releases. I have to admit that's better than I can do.

Remember the new Foveons chips are still the first
generation of their kind.

I remember. And I try and label my speculations as such and make no claim as to whether they are intrinsic or not. (In case you were not keeping score, ISO limitations and processing times are speculations. The increased I/O bandwidth requirements are fact. The exact cost of such requirements is speculation, but with a pretty darn good correlation to current data points.)

--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Ron Parr]

However, Sigma only rates the SD-9 at ISO 400 max (and we've seen
no Foveon samples from ISO other than 100.) Another odd fact is
that Foveon claims the one of the advantages of X3 is that it can
gather more light because there are no filters thus it will be MORE
sensitive than Bayer cameras. So why is the 1st camera is actually
slower than its competition?

We discussed this in another thread. I agree that the first generation sensor may not be as sensitive as competing sensors. I proposed reasons for this too.

Do you have anything new to contribute here?

In any case, it's not clear from the datasheet if the rolling
shutter mode is referrnig to the 25FPS mode or both modes.

It looks like it applies to both (and the amount of data xferred
scales appropriately.) It's again just curious that Sigma does not
have an FPS rate on the SD-9 specs.

I don't see why you're harping on the rolling shutter. Do you understand what a rolling shutter is?

There's no FPS listed for the Nikon D100 either. What's your conspiracy theory for this? It's funny how unfinished cameras tend not to have FPS ratings.

All cameras fill up their fast memory buffer and slow down when
writing to memory.

And I've shown that the problem is X3 worse! I've never said that
it was impossible to solve. I've only said it would either be more
expensive or lower performance than equivalent Bayer cameras.

You've shown that in one particular case (filling the buffer) it might be worse. However, what you've claimed is, "this camera is going to need a starte-of-the-art or better I/O system."

What we've shown is that buffering speed is not an issue

Who ever said it was? Can't you read? IT'S WRITE TO CF TIME that
will be the issue.

Nope. I can't read. This is a much more plausible conclusion than the possibility that you aren't following the argument.

and that jpeg compression (if done in a future Foveon
based camera) will take exactly the same speed as jpeg compression

Actually, we don't know this. I agree that it SHOULD be true and
that is why I find the SD-9 RAW output only and the "killer app for
a new computer" issues so puzzling.

I agree that it's peculiar, but the explanation is very straightforward: They cut corners both in hardware costs and development time.

1) The assumption that jpeg compression is not currently a
bottleneck on other cameras.

It's not an assumption. Look at the write times for current
cameras. Show me one that is faster (FPS) for RAW than JPEG.

Look at Phil's full page comparison of write times for different media for the EOS 1D. In every single case, the bandwidth is lower for JPEG than RAW. This shows that JPEG compression is slowing things down, the exact opposite of what you are claiming.

What it does suggest is that below 6X the slowness of jpeg compression doesn't hurt too much, but that it becomes an increasing factor with faster storage devices. Note that the fact that performance scales when you increase the speed of the storage device is great news for Foveon/Sigma because it means that in-camera electronics are probably not the limiting factor and that they will automatically benefit from the continuing improvment in flash storage speed.

Hmm, in a discussion about noise vs. ISO, and output frame rate of
X3 vs Bayer, I point out that the published specs for X3 ships
cameras seem to imply lower performance than current Bayer cameras
in these areas. And you think this is irrational and disruptive?

A dispassionate presentation of the facts would be most welcome.

A systematic misinterpretation of the facts in a way that is biased towards a particuar viewpoint, combined with your zeal for overstating the case is a bit disturbing.

Oh, I get it. Anyone who disagrees with you cannot be rational. If
you find it so irrational and disruptive then don't read my
comments.

I think that in the flood of stuff you have been posting, you have raised some valid issues. These aren't new, but if you presented them in a more even-handed way, you might generate fruitful discussion.

The way things have been going, however, I have a genuine concern that you will mislead others.

 

[Erik Magnuson]

We discussed this in another thread. I agree that the first
generation sensor may not be as sensitive as competing sensors. I
proposed reasons for this too.

Do you have anything new to contribute here?

I don't know. I did not participate in that other thread. If you are so concerned about repetition, stop now.

I don't see why you're harping on the rolling shutter. Do you
understand what a rolling shutter is?

Who's harping? I just said that they way I read the spreadsheet it applied to both. The typographical convention of putting the comment so that it's next to both specifications usually indicates this. Why are you so defensive that you feel the need to attack me?

There's no FPS listed for the Nikon D100 either. What's your
conspiracy theory for this? It's funny how unfinished cameras tend
not to have FPS ratings.

We have a good base of experience with what the frame rate of a D100 class camera will be (and will raise heck in Nikon Talk if it's any less). We don't have any experience with X3 cameras and to anyone who's ever built such systems, these are exactly the type of issues that tend to be hard or expensive.

You've shown that in one particular case (filling the buffer) it
might be worse. However, what you've claimed is, "this camera is
going to need a starte-of-the-art or better I/O system."

Well, I feel I've proved the case to anyone who cares to do the math and has any idea of what I/O throughput or even CF write time means. If you remain unconvinced, that's not my problem.

I agree that it's peculiar, but the explanation is very
straightforward: They cut corners both in hardware costs and
development time.

That's one explanation. I've presented reasons why it may be inconsistent with the state of the art in current cameras, but you remain unconvinced or even able to entertain that there might be other reasons.

Look at Phil's full page comparison of write times for different
media for the EOS 1D. In every single case, the bandwidth is lower
for JPEG than RAW. This shows that JPEG compression is slowing
things down, the exact opposite of what you are claiming.

Now you show your lack of reading comprehension. I said time not throughput. Look at http://www.dpreview.com/reviews/canoneos1d/page12.asp

2496 x 1662 RAW 1.8s 2.7s 3,500 KB
2464 x 1618 JPEG FINE 1.2s 1.8s 2,200 KB

It takes 1.8s to write the RAW file. It takes 1.2s to write the JPEG. 1.8 is 50% greater than 1.2. I can't use smaller words.

Why is the throughput less for JPEG? Most I/O operations consist of a fixed time overhead operation (initializing the device, checking for free space, etc.) and a data dependent time. The total throughput for shorter operation is always lower. You're a computer scientist: this is a standard latency/throughput relation. I don't know why you would think it doesn't apply here (or just won't admit it.)

device is great news for Foveon/Sigma because it means that
in-camera electronics are probably not the limiting factor and that
they will automatically benefit from the continuing improvment in
flash storage speed.

Here the problem is not reading but interpreting the data. The 1D vs D30 vs 995 results show that the in-camera electronics ARE the limiting case for all but the most expensive cameras--which has been my point all along. However, even the firewire reader is only 3MB/s for writing. Flash memory write times are not decreasing at any significant rate (e.g. slower than the expansion of data sizes.) So the benefit will be slow in coming. (Do you know how flash memory works? Have you ever implemented an I/O system for flash? Do you know of some breakthrough technology that the rest of us don't?)

A systematic misinterpretation of the facts in a way that is biased
towards a particuar viewpoint, combined with your zeal for
overstating the case is a bit disturbing.

Funny. Given your track record above in misunderstanding or misinterpreting the facts, that's exactly what I was thinking about you. Anytime I've presented facts and speculation, I've also incuded enough references and calculations for others to come to their own conclusions. I don't insist they trust me. They might even correct me as you did w.r.t the A/D on the F7 data sheet.

I think that in the flood of stuff you have been posting, you have
raised some valid issues.
The way things have been going, however, I have a genuine concern
that you will mislead others.

If I have raised valid issues, how can I mislead anyone? We're mostly adults here, capable of deciding things for ourselves. I don't think you have to be so protective. However, this has become repetitive and a waste of space. From now on, when you want to present your tired old arguments on why Bayer interpolation is so expensive and degrading, just include a URL to these threads. If I want to rebut, I'll do the same. It'll save us and everyone else so much time.--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Ron Parr]

We discussed this in another thread. I agree that the first
generation sensor may not be as sensitive as competing sensors. I
proposed reasons for this too.

Do you have anything new to contribute here?

I don't know. I did not participate in that other thread. If you
are so concerned about repetition, stop now.

Please stop with the lies and misrepresentations.

http://www.dpreview.com/forums/read.asp?forum=1000&message=2334512

I don't see why you're harping on the rolling shutter. Do you
understand what a rolling shutter is?

Who's harping? I just said that they way I read the spreadsheet it
applied to both. The typographical convention of putting the
comment so that it's next to both specifications usually indicates
this. Why are you so defensive that you feel the need to attack me?

You started off by implying that the rolling shutter was bad in some way. Explain how, retract your comment, or stop complaining about being attacked.

There's no FPS listed for the Nikon D100 either. What's your
conspiracy theory for this? It's funny how unfinished cameras tend
not to have FPS ratings.

We have a good base of experience with what the frame rate of a
D100 class camera will be (and will raise heck in Nikon Talk if
it's any less). We don't have any experience with X3 cameras and to
anyone who's ever built such systems, these are exactly the type of
issues that tend to be hard or expensive.

In other words, you're acting from a position of ignorance, but nevertheless holding Foveon to a higher standard.

You've shown that in one particular case (filling the buffer) it
might be worse. However, what you've claimed is, "this camera is
going to need a starte-of-the-art or better I/O system."

Well, I feel I've proved the case to anyone who cares to do the
math and has any idea of what I/O throughput or even CF write time
means. If you remain unconvinced, that's not my problem.

I can't possibly imagine how you've convinced anybody of anything other than your penchant for exaggeration.

Look at Phil's full page comparison of write times for different
media for the EOS 1D. In every single case, the bandwidth is lower
for JPEG than RAW. This shows that JPEG compression is slowing
things down, the exact opposite of what you are claiming.

Now you show your lack of reading comprehension. I said time not
throughput. Look at
http://www.dpreview.com/reviews/canoneos1d/page12.asp

2496 x 1662 RAW 1.8s 2.7s 3,500 KB
2464 x 1618 JPEG FINE 1.2s 1.8s 2,200 KB

It takes 1.8s to write the RAW file. It takes 1.2s to write the
JPEG. 1.8 is 50% greater than 1.2. I can't use smaller words.

Why is the throughput less for JPEG? Most I/O operations consist of
a fixed time overhead operation (initializing the device, checking
for free space, etc.) and a data dependent time. The total
throughput for shorter operation is always lower. You're a
computer scientist: this is a standard latency/throughput relation.
I don't know why you would think it doesn't apply here (or just
won't admit it.)

Your initial claim was that a Foveon based camera would require a "starte-of-the art or better" I/O system. We explored this and showed that in fact, the only case where it would make a difference is in flushing the memory buffer. I said that you were assuming that JPEG compression was not a bottleneck in this process and you suggested that we check Phil's data.

Fortuitously, Phil has data on exactly the question we're asking:

http://www.dpreview.com/articles/mediacompare/page2.asp

They show that up to 6X speed, CF writes are the bottleneck, not the camera's internal systems. After 6X jpeg compression slows things down. Thus, the camera's internal I/O systems are not the bottleneck for RAW: CF write speed is. There's nothing that sigma can do about this other than wait for faster CF.

Here the problem is not reading but interpreting the data. The 1D
vs D30 vs 995 results show that the in-camera electronics ARE the
limiting case for all but the most expensive cameras--which has

Regarding flash memory speeds: Look at the chart. There's a 2X difference between the older, slower stuff and the newer faster stuff. However, I must say that I don't know what the timing looks like for the CF protocol, so I can't say what the upper bound on the speed improvements will be.

What we know is that the I/O system on the 1D seems to max out CF write speed and that around 6X, jpeg compression starts to slow things down. I haven't seen similar graphs for other cameras, so I'm not going to make a wreckless claim about how hard it was for Canon to achieve this.

Funny. Given your track record above in misunderstanding or
misinterpreting the facts, that's exactly what I was thinking about

Name one.

I think that in the flood of stuff you have been posting, you have
raised some valid issues.
The way things have been going, however, I have a genuine concern
that you will mislead others.

If I have raised valid issues, how can I mislead anyone? We're
mostly adults here, capable of deciding things for ourselves. I

So, let me get this straight: What you're saying is that if you accidentally say something correct in the flood on nonesense we've been getting that's all for the better, but it's up to everybody else to separate the wheat from the chaff?

Well, sorry, as long as I have stomach for it (which admittedly may not be much longer), I'm going to point out the problems in your posts.

--Ron ParrFAQ: http://www.cs.duke.edu/~parr/photography/faq.htmlGallery: http://www.pbase.com/parr/

 

[Erik Magnuson]

Please stop with the lies and misrepresentations.

http://www.dpreview.com/forums/read.asp?forum=1000&message=2334512

Oh, I'm sorry. I was looking for the thread where you said something about how I/O limits would be addressed by future X3 implementations. The only thing discussed in the referenced thread was sensor chip cost. Thus my confusion. I assumed you had some pearls of wisdom I had missed. Perhaps one day you will understand that a camera's electronics cost is more than just the cost of the sensor chip.

You started off by implying that the rolling shutter was bad in
some way. Explain how, retract your comment, or stop complaining
about being attacked.

It's a caveat that was applied to the specs. According to Kodak http://www.kodak.com/US/plugins/acrobat/en/digital/ccd/Shutter-Operations-CCD-and-CMOS-PS-0259-Rev-1.pdf , one possible disadvantage of rolling shutter mode is motion artifacts. They claim to have patented a better mode for this that seems to be more appropriate for still cameras.

In other words, you're acting from a position of ignorance, but
nevertheless holding Foveon to a higher standard.

Yes, I am holding Foveon to a higher standard. What's your point? When you are claiming to improve on existing systems, you must show results or face doubts. Especially when your design makes unusual demands on system throughput.

I can't possibly imagine how you've convinced anybody of anything
other than your penchant for exaggeration.

Hmm, just because you cannot imagine something, it cannot happen? What a limited world you must live in.

http://www.dpreview.com/articles/mediacompare/page2.asp

They show that up to 6X speed, CF writes are the bottleneck, not
the camera's internal systems. After 6X jpeg compression slows
things down.

Slows things down compared to what? Compared to writing an imaginary data file that is the same size as the JPEG but contains all the same data? It doesn't matter if JPEG takes some time as long as the time spent doing the compression is noticably less than the additional I/O time of writing the uncompressed data! I've never said JPEG takes 0 time. I've only said that the total SYSTEM time (from pressing the button to having a useful image on the flash card) will be significantly less (50% or better from the examples.) And bigger output files only increase the disparity given the current write-speed to processing speed ratio (a ratio that is not likely to improve given historical trends.)

Thus, the camera's internal I/O systems are not the
bottleneck for RAW: CF write speed is.

Ahh. Here is the part where our terminology differs. How does the data get to the CF if not via the camera I/O? What other "output" is there that matters? You seem to think that just writing data to the internal buffer is enough. I assume that you are taking pictures to do something with them and thus need to save them somewhere. But maybe you always delete your pictures immediately. If so, the bit bucket is faster than any memory.

There's nothing that sigma
can do about this other than wait for faster CF.

Well, they could support a JPEG mode like everyone else. Because, absent a sudden breakthrough in storage technology, this is not likely to happen fast enough to matter.

However, I must say that I don't know what the timing looks
like for the CF protocol, so I can't say what the upper bound on
the speed improvements will be.

So, admitting complete ignorance of the limits or operation of the underlying technology, you are going to assume that there are significant improvements just waiting to be exploited? Hmm, you do have an imagination after all.

Name one.

Actually, the one that galls me is your complete misconceptions about what system throughput means and how much each stage contributes to it. No existing camera seem to work the way you think it does.

So, let me get this straight: What you're saying is that if you
accidentally say something correct in the flood on nonesense we've
been getting that's all for the better, but it's up to everybody
else to separate the wheat from the chaff?

Yep, that's life. All reponsible adults should haveworking BS detectors. Until someone makes you God or forum moderator, you will just have to put up with people disagreeing with you.

--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Rinus Borgsteede]

Thanks Ron for an excellent explanation.

This issue has generated a lot of posts. People difffer in opinion but the truth is that the Bayer pattern or for that matter any other interpolation is still vulnerable to errors.

Ultimately a chip like Foveon with the hexagon diagonal photosites as seen in the Super CCD may provide some real filmlike resolution.
Rinus

 

[Ron Parr]

Please stop with the lies and misrepresentations.

http://www.dpreview.com/forums/read.asp?forum=1000&message=2334512

Oh, I'm sorry. I was looking for the thread where you said
something about how I/O limits would be addressed by future X3
implementations. The only thing discussed in the referenced thread
was sensor chip cost. Thus my confusion. I assumed you had some
pearls of wisdom I had missed. Perhaps one day you will understand
that a camera's electronics cost is more than just the cost of the
sensor chip.

??? Still chaning topics when you're wrong, hmm?

In other words, you're acting from a position of ignorance, but
nevertheless holding Foveon to a higher standard.

Yes, I am holding Foveon to a higher standard. What's your point?
When you are claiming to improve on existing systems, you must show
results or face doubts. Especially when your design makes unusual
demands on system throughput.

But we've established that it doesn't except for the case when you fill the buffer.

I can't possibly imagine how you've convinced anybody of anything
other than your penchant for exaggeration.

Hmm, just because you cannot imagine something, it cannot happen?
What a limited world you must live in.

I assume people are intelligent. I know it's optimistic, but I cling to the hope.

Thus, the camera's internal I/O systems are not the
bottleneck for RAW: CF write speed is.

Ahh. Here is the part where our terminology differs. How does the
data get to the CF if not via the camera I/O? What other "output"
is there that matters? You seem to think that just writing data to
the internal buffer is enough. I assume that you are taking
pictures to do something with them and thus need to save them
somewhere. But maybe you always delete your pictures immediately.
If so, the bit bucket is faster than any memory.

Is it that you don't understand that this is only issue only when the buffer fills, or are you intentionally being obtuse?

However, I must say that I don't know what the timing looks
like for the CF protocol, so I can't say what the upper bound on
the speed improvements will be.

So, admitting complete ignorance of the limits or operation of the
underlying technology, you are going to assume that there are
significant improvements just waiting to be exploited? Hmm, you do
have an imagination after all.

Pathetic drivel.

Name one.

Actually, the one that galls me is your complete misconceptions
about what system throughput means and how much each stage
contributes to it. No existing camera seem to work the way you
think it does.

Sorry. They all do.

On the other hand, you've claimed that the camera's internals would need to be something special in order to deal with Foveon type data and you've been wrong on every point. The two limiting factors are the throughput off the chip and bandwidth to CF, both of which ensure that ordinary camera guts will suffice.

We've verified that capturing large images indeed involves pulling them off the chip and sticking on memory and that (shock!) when the buffer gets full, you need to wait for it to drain.

What's incredible is that despite total wrongness about the true sources of the bottlenecks and their remedies, you still manage to decry what you perceive as ignorance in others.

So, let me get this straight: What you're saying is that if you
accidentally say something correct in the flood on nonesense we've
been getting that's all for the better, but it's up to everybody
else to separate the wheat from the chaff?

Yep, that's life. All reponsible adults should haveworking BS
detectors. Until someone makes you God or forum moderator, you will
just have to put up with people disagreeing with you.

Indeed!

--Ron ParrFAQ: http://www.cs.duke.edu/~parr/photography/faq.htmlGallery: http://www.pbase.com/parr/

 

[Ron Parr]

Is in-camera review of pictures you've taken.

Getting good interactive response times when thumbing through pictures you've taken could be a problem. Having a buffer doesn't help here if read time off CF is slow.

They could store little thumbnails to help, but if you want to zoom in or do a hisogram, you'll need to wait for the whole thing to load and then be converted for display.

--Ron ParrFAQ: http://www.cs.duke.edu/~parr/photography/faq.htmlGallery: http://www.pbase.com/parr/

 

[Erik Magnuson]

??? Still chaning topics when you're wrong, hmm?

Maybe, but since I'm not wrong here, your assertion is moot.

But we've established that it doesn't except for the case when you
fill the buffer.

Oh, who is the royal "we" here? For some reason, Phil thinks this time is important enough to include in his reviews. It must have some meaning.

Is it that you don't understand that this is only issue only when
the buffer fills, or are you intentionally being obtuse?

OK. I press the button. A photo is captured. It gets written to the buffer. However, by your logic since the buffer is not full it never gets written to flash. What camera works like this? Ok, you say you want to be able to take another photo while the first one is being written to flash. But now you need dual-ported memory to handle the simultaneous operation. And either the buffer holds fewer shots because the file sizes are larger or the buffer is has to be considerably larger. And add a bigger, faster processor to manage it all and keep those I/O lines humming. None of this impossible, it's just not cheap. It's even more expensive to do quickly or your buffer will always be full. Which brings us full circle to my point from the very beginning: TO ACHIEVE COMPETITIVE PERFORMANCE, AN X3 CAMERA WILL NEED MORE EXPENSIVE SUPPORT AND I/O CIRCUITRY THAN A MOSIAC CAMERA. Is that acute enough for you? Did the all caps help?

Pathetic drivel.

Ooh, I'm cut to the quick.

Sorry. They all do.

Sure. Bayer interpolation and JPEG compression are the performance bottleneck for all cameras. That's why those TIFF and RAW write times are so blazing. And fast I/O systems and big buffers are cheap. The only reason they are not in less expensive cameras is that the mfg is trying to preserve the model differentiation between their pro lines and other lines. (Today is opposite day.)

On the other hand, you've claimed that the camera's internals would
need to be something special in order to deal with Foveon type data

I believe my actual words have been "state-of-the-art" (to achieve lower performance compared to a "comparable" mosaic sensor camera) or "better" (to match the performance of said mosaic sensor camera.) I've said more expensive, not impossible. Do you think you might be guilty of misrepresenting what I said just a wee bit here? This seems to be a common pattern in your many of your posts. My claims are actually quite mild to anyone who examines them dispassionately.

and you've been wrong on every point.

...and there you go again. Are you just mad because I've shown how you are all wet w.r.t JPEG being a performance bottleneck? Or have you come up with that imaginary image data format that can be computed in 0 time and yet written out faster than JPEG?

The two limiting factors are
the throughput off the chip and bandwidth to CF, both of which
ensure that ordinary camera guts will suffice.

Ordinary. I don't know what this means, there is too much variability. Pick an "ordinary" camera and we can calculate what impact of replacing the current sensor with a X3 sensor would be. I'll start: the D30 using a microdrive using Phil's numbers. D30 RAW size 3MB. X3 RAW file size: 8MB. I/O rate 836 kb/s. The D30's buffer is 9MB in size (can hold 3 RAW, 9 JPEG).

Go!

1st photo. Done.

2nd photo. Oops, can't do it, buffer full. Have to wait. It will take 9-10 seconds to write that 8MB file. (Ok, maybe only 8 if the camera is real smart about memory management and does not mind overwriting a partially flushed file.)

Anyone think that will suffice? Oh, you think that a 3x buffer size is "ordinary". OK, now we get the same number of RAW shots buffered, but the flush time is still 7-8 seconds (assuming we were gettting 2FPS off-chip.) OK, now you say you will do 2x write speed. (Hmm, currently only $5000+ cameras can do that. We're beginning to push the limits of ordinary.) But wait, the D30 can also hold 9 JPEG shots in that original buffer. The flush time now is only 2 seconds with 3x the shots before full. And if we put that same 3x buffer and 2x write in the D30, it's suddenly a lot faster again than the X3 version! Magic stuff, huh?

We've verified that capturing large images indeed involves pulling
them off the chip and sticking on memory and that (shock!) when the
buffer gets full, you need to wait for it to drain.

And, duh, it takes longer to do so. (And suddenly inexpensive cameras that did NOT do parallel I/O now require it.) Hence my performance claims.

What's incredible is that despite total wrongness about the true
sources of the bottlenecks and their remedies, you still manage to
decry what you perceive as ignorance in others.

Yep, that's me. I call 'em as I see 'em. How's your stomach holding up?

--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Ron Parr]

You've correctly percieved that I'm getting sick of this discussion.

Here's what it seems you've established in your previous email:

• You didn't realize that essentially all modern cameras let you take shots as you fill the buffer. Indeed, that's why they have buffers that are larger than one shot.

• You were implicitly assuming all along that the buffer would be the same size despite the fact that files are larger. I can see why you're so confused. If you make stupid assumptions, you get stupid conclusions. (Please don't tell me the cost of memory, which the Korean's have been essentially giving away, will be an issue on a $3000 camera.)

--Ron ParrFAQ: http://www.cs.duke.edu/~parr/photography/faq.htmlGallery: http://www.pbase.com/parr/

 

[charlie gast]

All the mathematical GARBAGE you say are in his posts ARE in your camera.

 

[Erik Magnuson]

• You didn't realize that essentially all modern cameras let you

take shots as you fill the buffer. Indeed, that's why they have
buffers that are larger than one shot.

You still don't get it. Filling the buffer is NOT an important performance measure. That's like saying I have a really fast camera - if there is no film in it. If the output is not "fixed", it's not a camera (unless it's a camera obscura). The performance measure that everyone but you thinks is important is the time before the image is written to some form of non-volatile storage. Why do you think that this time measured and reported in all the reviews? Only high-end cameras offer the ability to continue to take photos while the buffer is being written to flash (precisely because it's relatively expensive to implement.)

• You were implicitly assuming all along that the buffer would be

the same size despite the fact that files are larger. I can see
why you're so confused. If you make stupid assumptions, you get
stupid conclusions.

Well all I did was assume that "state-of-the-art" means "the best that is available today" and show how that's not quite good enough. I did not realize that in parr-land it means something else. Still, the rising tide lifts all boats. Any improvements you posit to the "state-of-the-art" will increase the performance of mosaic sensor cameras too.

(Please don't tell me the cost of memory,
which the Korean's have been essentially giving away, will be an
issue on a $3000 camera.)

Again, you are betraying your ignorance of what it takes to build these types of systems. If you think that high-performance dual-ported RAM in a small form factor costs the same as a stick of SDRAM, you would be sadly dissappointed.
--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Erik Magnuson]

All the mathematical GARBAGE you say are in his posts ARE in your
camera.

The point is that most of us use our cameras to take actual photographs for the purpose of being observed by other humans. The most significant measure of goodness is usually "pleasing". Mathmatical precision is only important if it leads to this goal. Sharpness, resolution of fine detail, and color accuracy have a high correlation with "pleasing" which is why we seek cameras that have these attributes. But there are always tradeoffs. (Do you use an 8x10 view camera and drum scan the output? No? Why not? Oh, yeah, sometimes other little things like flexibility and convenience enter the picture and the results are still good enough.) Once you agree that there are tradeoffs to be considered, the rest is just quibbling over how to draw the line and what's important to each individual.

In the end, it does not matter if the camera output is produced by mathmatical approximations or a small demon inside with a paint brush. It either produces good pictures or it doesn't. All photographs are just approximations to the ideal image.
--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Ron Parr]

You still don't get it. Filling the buffer is NOT an important
performance measure. That's like saying I have a really fast camera

• if there is no film in it. If the output is not "fixed", it's not

a camera (unless it's a camera obscura). The performance measure
that everyone but you thinks is important is the time before the
image is written to some form of non-volatile storage. Why do you
think that this time measured and reported in all the reviews? Only
high-end cameras offer the ability to continue to take photos while
the buffer is being written to flash (precisely because it's
relatively expensive to implement.)

The discussion was about frames per second for the camera. FPS ratings for digital cameras are always measured while the buffer is filling, not in terms of throughput to CF. Check the FPS rating for any camera reviewed by Phil. All of these are rated at a particular speed until the buffer fills.

Perhaps the hidden message in what you're saying is in the slow response times issues I raised above. If you want to review your photos on the LCD and they're being written from the buffer or even if they're already on CF, it may take a long time to flip through. However, this is really a different question from the FPS issue. If you want to beat this drum, I may agree with you. I don't know about Nikon, but Canon embeds jpegs in their RAW files to facilitate this. One has to wonder what Sigma will do about this.

• You were implicitly assuming all along that the buffer would be

the same size despite the fact that files are larger. I can see
why you're so confused. If you make stupid assumptions, you get
stupid conclusions.

Well all I did was assume that "state-of-the-art" means "the best
that is available today" and show how that's not quite good enough.
I did not realize that in parr-land it means something else. Still,
the rising tide lifts all boats. Any improvements you posit to the
"state-of-the-art" will increase the performance of mosaic sensor
cameras too.

Having a larger buffer won't change the FPS rating of the camera. All it will change is the length of time at which it can sustain that rate.

(Please don't tell me the cost of memory,
which the Korean's have been essentially giving away, will be an
issue on a $3000 camera.)

Again, you are betraying your ignorance of what it takes to build
these types of systems. If you think that high-performance
dual-ported RAM in a small form factor costs the same as a stick of
SDRAM, you would be sadly dissappointed.

I wasn't going to get into to this, because I was hoping that you would realize dual ported memory was unnecessary and just drop it, but here goes:

You don't need dual ported memory to implement a buffer of the type we're discussing. The first and obvious thing you can do is just pause writing CF when you're feeding in a new image from the sensor. This is what the D30 does.

If you're a wee bit more sophisticated, you can realize that memory is much faster than the bandwidth you can pull off the sensor and much, much faster than what you can write to CF. If you have memory that's fast enough, you can just interleave both operations. Now, I DO realize that the high bandwidth acheived by modern systems come the fact tha memory transfers are pipelined and that interleaving might not be possible depending upon latency. However, you can also just put a small write buffer in front of the CF and occasionally pause to fill it up if needed. the write buffer can be filled VERY quickly, so you would get the effect of dual ported memory without ever needing it. This is a perfect example of why I/O devices have buffers in the first place.

At this point, I would like to suggest that we tone down the rhetoric a bit. I noticed this post on the Canon SLR forum:

http://www.dpreview.com/forums/read.asp?forum=1019&message=2353253

and I felt that it was timely.

Perhaps there have been misunderstandings between us in our dialogue, but I don't think that any of this justifies the tone that things have taken. It makes me angry when I read things such as, "Again, you are betraying your ignorance of what it takes to build these types of systems" which were clearly written out of anger and are neither correct nor constructive. It's taking some determination for me to not respond in kind, but I'm making an effort and I hope that for the sake of forum you will do the same.

--Ron ParrFAQ: http://www.cs.duke.edu/~parr/photography/faq.htmlGallery: http://www.pbase.com/parr/

 

[Erik Magnuson]

The discussion was about frames per second for the camera. FPS
ratings for digital cameras are always measured while the buffer is
filling, not in terms of throughput to CF. Check the FPS rating
for any camera reviewed by Phil. All of these are rated at a
particular speed until the buffer fills.

No, he (and everyone else) carefully reports all three times (FPS until full, FPS when the buffer is full, and buffer flush time. If you think the buffer flush time is sooo unimportant, go over to the Oly forum and read the flame wars about the E10/E20. It has the very X3-like characteristic that the fill time may be fast but the flush time is slow. (It also buffers the same number of frames whether in RAW or JPEG.) See how popular this design choice is. Some people see it as a fatal flaw, others say they can live with it as they don't need this kind of performance.

Perhaps the hidden message in what you're saying is in the slow
response times issues I raised about.

You still keep reading FAR more into my claims than I've ever said. All I've ever said was either SLOWER or MORE EXPENSIVE I/O. Not unusable or unacceptable (although perhaps dissappointing to some.) The inevitable consequence of producing X3 more data. I've also said that based the cost/performance curves we have from current cameras, this means that any sensor cost advantage may be negated.

Having a larger buffer won't change the FPS rating of the camera.
All it will change is the length of time at which it can sustain
that rate.

Finally, we agree on something. (Is the world ending?) But all cameras have 2 FPS rates: one burst and one sustained. Every review measures both (unless the camera is too cheap for them to be any different.) You keep ignoring or dismissing the second one as unimportant. Fine. It does not matter to you. It obviously matters to a large number of other photographers.

You don't need dual ported memory to implement a buffer of the type
we're discussing. The first and obvious thing you can do is just
pause writing CF when you're feeding in a new image from the
sensor. This is what the D30 does.

Of course. That's likely one of the reasons the D30 is slower than the 1D. Stopping and restarting has a price on both ends. BTW, I consider the D30 "state of the art" for it's price class. If you are saying you need something at least as good in the SD9, then you are in the unhappy position of agreeing with me again. (Actually, I suspect we agree 95% of the time. It's just when we disagree, we let it get out of hand.)

Perhaps there have been misunderstandings between us in our
dialogue, but I don't think that any of this justifies the tone
that things have taken.

Well, I did get a chuckle over the idea that this forum somehow needed to be protected from me. I spent a lunchtime plotting on how I could use my evil influence to take over the world. Fortunately, I forgot how I was going to do it a few hours later, so you are all safe now.

I don't take it personally and harbor no hard feelings. We'll know a heck of a lot more once cameras actually get into reviewer's and user's hands. Until then, it's all just hot air.--ErikFree Windows JPEG comment editor http://home.cfl.rr.com/maderik/edjpgcom

 

[Tau]

Funny, the initial subject name of this thead. I watched this instead of TV today. Loved it.
Jack.

The discussion was about frames per second for the camera. FPS
ratings for digital cameras are always measured while the buffer is
filling, not in terms of throughput to CF. Check the FPS rating
for any camera reviewed by Phil. All of these are rated at a
particular speed until the buffer fills.

No, he (and everyone else) carefully reports all three times (FPS
until full, FPS when the buffer is full, and buffer flush time. If
you think the buffer flush time is sooo unimportant, go over to the
Oly forum and read the flame wars about the E10/E20. It has the
very X3-like characteristic that the fill time may be fast but the
flush time is slow. (It also buffers the same number of frames
whether in RAW or JPEG.) See how popular this design choice is.
Some people see it as a fatal flaw, others say they can live with
it as they don't need this kind of performance.

Perhaps the hidden message in what you're saying is in the slow
response times issues I raised about.

You still keep reading FAR more into my claims than I've ever said.
All I've ever said was either SLOWER or MORE EXPENSIVE I/O. Not
unusable or unacceptable (although perhaps dissappointing to some.)
The inevitable consequence of producing X3 more data. I've also
said that based the cost/performance curves we have from current
cameras, this means that any sensor cost advantage may be negated.

Having a larger buffer won't change the FPS rating of the camera.
All it will change is the length of time at which it can sustain
that rate.

Finally, we agree on something. (Is the world ending?) But all
cameras have 2 FPS rates: one burst and one sustained. Every review
measures both (unless the camera is too cheap for them to be any
different.) You keep ignoring or dismissing the second one as
unimportant. Fine. It does not matter to you. It obviously matters
to a large number of other photographers.

You don't need dual ported memory to implement a buffer of the type
we're discussing. The first and obvious thing you can do is just
pause writing CF when you're feeding in a new image from the
sensor. This is what the D30 does.

Of course. That's likely one of the reasons the D30 is slower than
the 1D. Stopping and restarting has a price on both ends. BTW, I
consider the D30 "state of the art" for it's price class. If you
are saying you need something at least as good in the SD9, then you
are in the unhappy position of agreeing with me again. (Actually, I
suspect we agree 95% of the time. It's just when we disagree, we
let it get out of hand.)

Perhaps there have been misunderstandings between us in our
dialogue, but I don't think that any of this justifies the tone
that things have taken.

Well, I did get a chuckle over the idea that this forum somehow
needed to be protected from me. I spent a lunchtime plotting on
how I could use my evil influence to take over the world.
Fortunately, I forgot how I was going to do it a few hours later,
so you are all safe now.

I don't take it personally and harbor no hard feelings. We'll know
a heck of a lot more once cameras actually get into reviewer's and
user's hands. Until then, it's all just hot air.
--
Erik
Free Windows JPEG comment editor
http://home.cfl.rr.com/maderik/edjpgcom

--The significant problems we face can not be solved at the same level of thinking we were at when we created them.(A.E.)