Canon's 250mp 1.3 crop sensor. Why not today?

Whisky-dave wrote:
RichA wrote:
Why is it these things get confined to ultra-expensive scientific cameras?

Because they're super-expensive to make.

They can obviously make them just like other sensors, so perhaps 9/10
of them are too flawed to sell?

I recon the cost of prodicing such a thing is probbaly what ensures that it
won't be in commercail cameras.

It comes down to cost from statistics. With a 250MP field, if one is satisfied
with only 99.99% of the sensor locations being good, that means that you
still still have 25,000 bad pixels.

Up the bar to 99.999 9% and you're now down to "only" 250 bad pixels per unit.

Now contemplate the manufacturing process to etch/wash/plate/etc a
wafer through all of its steps to get from raw material to finished product;
In a nutshell, its safe to say that there's going to be at least 40 steps.
Normalized, to make material that's 99.999 9%, you need each step
to perform at 99.999 997 5% (or better). Since that's not possible, you're
IMO easily looking at a scrap rate of over 4000:1 (on a good day) and
probably over $20,000* in scrap ... per unit built.

* - this is just the raw material costs and doesn't include any fab processing
costs .. probably safe to triple this number.


-hh

In article , -hh
says...
It comes down to cost from statistics. With a 250MP field, if one is satisfied
with only 99.99% of the sensor locations being good, that means that you
still still have 25,000 bad pixels.

Up the bar to 99.999 9% and you're now down to "only" 250 bad pixels per unit.

Now contemplate the manufacturing process to etch/wash/plate/etc a
wafer through all of its steps to get from raw material to finished product;
In a nutshell, its safe to say that there's going to be at least 40 steps.
Normalized, to make material that's 99.999 9%, you need each step
to perform at 99.999 997 5% (or better). Since that's not possible, you're
IMO easily looking at a scrap rate of over 4000:1 (on a good day) and
probably over $20,000* in scrap ... per unit built.

* - this is just the raw material costs and doesn't include any fab processing
costs .. probably safe to triple this number.

99.99% good pixels or even just 99.9% good pixels (1 bad pixel in 1000
good pixels) could suffice, as long as all bad pixels are not all in one
place. Bad pixels can be easily mapped out.

They make DRAM with billions of memory cells at very low manufacturing
prices, so it shouldn't be that difficult to make an image sensor with
250 million pixels.
--
Alfred Molon

Olympus E-series DSLRs and micro 4/3 forum at
http://tech.groups.yahoo.com/group/MyOlympus/
http://myolympus.org/ photo sharing site

Alfred Molon wrote:
-hh says...
It comes down to cost from statistics. With a 250MP field, if one is satisfied
with only 99.99% of the sensor locations being good, that means that you
still still have 25,000 bad pixels.

Up the bar to 99.999 9% and you're now down to "only" 250 bad pixels per unit...

99.99% good pixels or even just 99.9% good pixels (1 bad pixel in 1000
good pixels) could suffice, as long as all bad pixels are not all in one
place. Bad pixels can be easily mapped out.

True, they could be, although the key assumption here is the assumption that
there's no clustering. Given that flaws in the substrate material will cause
clustering, that assumption is problematic.

They make DRAM with billions of memory cells at very low manufacturing
prices, so it shouldn't be that difficult to make an image sensor with
250 million pixels.

True, but DRAM doesn't utilize one huge chunk of homogenous material to
scale up. As such, it can bin out the bad ones (and localized remapping to
pre-emplaced onboard spares).

FWIW, an alternative to having one big chunk of material would be to use a
a bunch of smaller ones ... which in the context of an optical sensor would
then mean that an already spatially combined signal needs to be broken into
discrete sub-pieces to then redirect to each sub-sensor assembly. Can be
done (brute force engineering), although it gets messy quite quickly.


-hh

In article , -hh
says...


True, they could be, although the key assumption here is the assumption that
there's no clustering. Given that flaws in the substrate material will cause
clustering, that assumption is problematic.

On what do you base your assumption that all defects are in one small
location?

True, but DRAM doesn't utilize one huge chunk of homogenous material to
scale up.

What does that mean?

As such, it can bin out the bad ones (and localized remapping to
pre-emplaced onboard spares).

??? Please explain what you mean.

--
Alfred Molon

Olympus E-series DSLRs and micro 4/3 forum at
http://tech.groups.yahoo.com/group/MyOlympus/
http://myolympus.org/ photo sharing site

Alfred Molon wrote:
-hh says...

True, they could be, although the key assumption here is
the assumption that there's no clustering. Given that
flaws in the substrate material will cause
clustering, that assumption is problematic.

On what do you base your assumption that all defects are
in one small location?

There's a general correlation between circuit defects and
material flaws, and the latter are known to be discrete.
As such, defects are more likely to cluster around flaws.

The general challenge with larger structures is that as
the size of the chip grows, the likelihood of the substrate
having a flaw increases too...and thus, the likelihood of
a circuit defect.


True, but DRAM doesn't utilize one huge chunk of homogenous
material to scale up.

What does that mean?

Simplistically, that a wafer full of DRAM's doesn't need to
all be perfect: you fab, test, dice & scrap the bad ones,
sending only the good ones forward for assembly into a DIMM.

In this context, you care about the yield rates on a wafer level,
not the chip level, because you're making & using multiples of them.

As such, it can bin out the bad ones (and localized remapping to
pre-emplaced onboard spares).

??? Please explain what you mean.

Binning is to removed bad chips.

Localized remapping is an option on some architectures, where
there's a "spare" circuit already onboard which can be employed
through mapping out of bad stuff to have the chip still meet
its specifications. Remapping only is viable when the spatial
location of the circuits in question are not critical, which
means that this doesn't apply well to photon receptor wells.


-hh