DACs - die pictures

[mawyatt]

1µm magnified 20x is 20µm on the senor. Find out how many pixels that is and you will have an idea if you are limited by the camera or by the optics. Probably optics, because sensors and image processing algorithms have no trouble resolving lines spaced only a few pixels apart as you can confirm yourself by photographing any random thing with a known good lens.

The Canon 60D has a pixel size of 4,3µm. So with 20x that would allow me to distinguish lines separated 200nm. OK 200nm is more than we need but the dslr sensor is not perfect and the lenses are not perfect too. So 20x could be a good factor.

This is beyond the capability of conventional techniques since diffraction and Nyquist come into play. Green has a wavelength of ~550nm. There are numerous discussions about this, details over at PM. Here's just 1 discussion.

https://www.photomacrography.net/forum/viewtopic.php?t=41213&highlight=

In a two lens / infinity system, magnification can be adjusted independently of the primary lens by changing focal length of the second lens. Of course, aperture and quality of the primary lens still limit available detail and resolution. Vignetting may occur if you try too low magnification.

I´m a bit afraid of too much glass. Every part could add degradation.

Not necessarily, consider the fine vs. infinite microscope objectives. Th infinite objective tend to have better image quality than finite objectives, yet have more glass, the tube lens!

With a single 160mm 40x objective applying your lighting method will be rather difficult. Those objectives have dimensions similar to my webcam lens - take any die bigger than 1mm and the center will be dark.

I agree with that. Perhaps reflecting the light around the lens would also be possible... 


I use external lighting with Strobes, some use continuous LED lights. I'm working on modifying the IKEA Jansjo LED Lamps for much higher optical output and using pulses, see here.

https://www.photomacrography.net/forum/viewtopic.php?t=41464&highlight=

And also modifying a high powered Video LED for macro use.

https://www.photomacrography.net/forum/viewtopic.php?t=41353&highlight=

Some objectives have low working distances which makes lighting very difficult, try and use lenses with longer working distances. A simple small styrofoam cup placed over the subject and illuminated with a couple speedlights or strobes works well, of course more elaborate setups are also effective. One thing to remember, that's contrary to some folks thinking, is the closer the light source to the subject the better and softer the light. If you are diffusing the light source, like with the cup, then the cup becomes the light source as "seen" from the subject and the actual optical source can be moved further away.

Anyway, all the questions you have and likely to ask, have been asked by many before, and well answered over at the mentioned PM site. Spend quite a few hours/days there studying the techniques, lenses, cameras, focus rails, illumination sources, fixtures, setups and post processing. There are also lots of theoretical discussions, like the limits of resolution, lens design, de-convolution image processing and so on.

Best,

[magic]

Regarding illumination, we want to produce coloration of the IC by iridescence so we illuminate from the same angle as we shoot from. If one wants to avoid heavy focus stacking it is necessary to shoot at 90° angle and therefore light must come from the direction of the lens. Professional metallurgical microscopes enable this by TTL illumination, Noopy invented an effective and low cost alternative which is bouncing light off the lens.

This is what we are talking about and AFAIK it hadn't been done by anyone before. It's simple, it's fast and it works well, but it only works with bulky lenses or very small dice.

https://www.photomacrography.net/forum/viewtopic.php?t=41213&highlight=

That's a good post (even if not entirely beginner-friendly, perhaps). At any rate, it seems to confirm what we said that a few pixels per line pair ought to suffice.

I humbly offer my own proof by example. The attached picture is a 100% crop which demonstrates that my system clearly resolves dark lines spaced 5 pixels apart in any orientation, does a half-assed job at 4 pixels and totally fails at 3 or less. Some of that might be due to the lenses used rather than effects of sampling/antialiasing/demosaicing alone.

[Noopy]

Thanks mawyatt for all your input!
You are right that we can´t discuss the whole background here but you gave me some very interesting points to think about.
Your Mitutoyo seems to be a very nice lens but it´s also very expensive! 

As magic explained it´s very important for us to get these nice coloured pictures which is possible with a "cheap" DSLR if you put your light behind the die and let it bounce back from the lens of the objective. Sound crazy for a normal photographer but it works well after some trial and error.
But perhaps I can mix the the microscope lens with the right illumination... 


...I know it´s "not possible" to get a resolution of 200nm with normal light, it was more a theoretical figure... 

[Ranayna]

Having an actual Chipdesigner here allows me to finally ask this question, especially since you pointed it out yourself:

Are such extra features, like your tennis raquet, just arbitrary, or do they have some kind of secondary use as well, like usage as some kind of fiduciary, or for quality checks?
Would your boss (I assume you have one ) let you put anything (within reason) in spare areas?

[mawyatt]

Regarding illumination, we want to produce coloration of the IC by iridescence so we illuminate from the same angle as we shoot from. If one wants to avoid heavy focus stacking it is necessary to shoot at 90° angle and therefore light must come from the direction of the lens. Professional metallurgical microscopes enable this by TTL illumination, Noopy invented an effective and low cost alternative which is bouncing light off the lens.

This is what we are talking about and AFAIK it hadn't been done by anyone before. It's simple, it's fast and it works well, but it only works with bulky lenses or very small dice.

https://www.photomacrography.net/forum/viewtopic.php?t=41213&highlight=

That's a good post (even if not entirely beginner-friendly, perhaps). At any rate, it seems to confirm what we said that a few pixels per line pair ought to suffice.

I humbly offer my own proof by example. The attached picture is a 100% crop which demonstrates that my system clearly resolves dark lines spaced 5 pixels apart in any orientation, does a half-assed job at 4 pixels and totally fails at 3 or less. Some of that might be due to the lenses used rather than effects of sampling/antialiasing/demosaicing alone.

OK, I see you are using a different type illumination than from all around the subject like we use, which requires high levels of diffusion. You might post this interesting technique over at PM to see what those folks have to say, there are some techniques called bright field, dark field and epi that folks use, some require polarized source and such. I don't know enough about these other techniques to comment.

Yes realistically it takes 3 pixels or more, and the more the better to resolve that level of detail. There has been much discussions about this recently if you follow the threads and references. This also follows along the discussions on de-convolution in post processing (AI sharping), even without the detailed lens/camera functions which is called blind de-convolution, folks are producing some really nice images that don't look too doctored up. I have yet to try any de-convolution sharping, but looking forward playing with this soon....if I can get all the other projects out of the way

Your fine detailed image looks great, nice work!! 

What are the feature sizes?

Best,   

[mawyatt]

Thanks mawyatt for all your input!
You are right that we can´t discuss the whole background here but you gave me some very interesting points to think about.
Your Mitutoyo seems to be a very nice lens but it´s also very expensive! 

As magic explained it´s very important for us to get these nice coloured pictures which is possible with a "cheap" DSLR if you put your light behind the die and let it bounce back from the lens of the objective. Sound crazy for a normal photographer but it works well after some trial and error.
But perhaps I can mix the the microscope lens with the right illumination... 


...I know it´s "not possible" to get a resolution of 200nm with normal light, it was more a theoretical figure... 

Yes the Mitutoyo's are somewhat expensive, but worth it if you do this type of imaging a lot. I wouldn't recommend starting with one, but later you'll probably end up with 1 or 2...maybe more Some Nikon's are really good also.

Later you will also find you'll want something that's really good around 1~2X, much better than standard macro lenses, something that can out-revolve say a Nikon D850 at 1X The Mitiutoyo's below 5X aren't the better choices. Nikon produced some superb film reproduction lens (think these were for motion picture film replication) called Printing-Nikkors, the 105mm F2.8 version is a brilliant optical design, but are rare and expensive. Another lens that's found repurposing is the lens from an old Minolta DiMage 5400 scanner, when setup properly this lens is really good from 1.5~4X. It's price has jumped since other folks are finding out about it.

https://www.closeuphotography.com/minolta-dimage-scan-elite-5400-lens

Robert's site is a wealth of information on lenses and such.

https://www.closeuphotography.com

It's kinda fun to construct a lens from surplus/scrap parts and end up with a jewel of an overall lens assembly for studio macro use (reminds me of getting a couple Tektronix 2465 scopes off eBay and fixing them, same for a couple HP34401A DVMs), there are lots of options to about ~5X but then you start wanting the Mitutoyos for 5X and beyond.

Please keep posting your beautiful chip images, really enjoy seeing these masterpieces of silicon displayed is such wonderful fashion.

Got to finish up my voltage reference design built around the LM399, which your images and detailed circuit were superb and very helpful!! I had created a SPICE model for the LM399 and also a 6.2V Zener with NPN (2N3904), had to "tweak" the LM399 resistors to get the temp curve to look OK, but after seeing your image and schematic understood why things needed changing.

One circuit you might be interested in is a ultra-precision voltage divider that doesn't require precision components. If you use a CMOS FF and tie resistors between the Q and Qbar output, the center of the resistors is shunted to ground with a filter cap. The center voltage will be exactly 1/2 the CMOS FF Vdd voltage independent of the 2 resistor values, they don't even have to match!! What's happening is the Q and Qbar resistors from a simple voltage divider, say Vdd(R1/(R1+R2)), where R1 is tied to Q and R2 to Qbar. On the next clock edge Q and Qbar swap state and now the voltage divider is Vdd(R2/(R1+R2)). The capacitor averages these two voltages to exactly 1/2 Vdd. In real life the small Rdon of the NMOS and PMOS comes into play but can be swamped by R1 and R2 values, or use paralleled inverters on Q and Qbar outputs. This can produce results  around a ppm with 1% resistors We patented (5030848) this concept way back and used it in numerous designs, but the patent has long since run out.

If you have a 74AC74 or other FF and a 74AC04 or other inverter CMOS parts, and a couple 10K, 50K or 100K, or just about any resistor value and any cap. Just clock the FF at say 1~10KHz and supply 5.00000 volts as Vdd, then measure across the cap. You will get something like 2.49999 volts once you factor the DVM impedance with the Thevin eqv. impedance from the voltage divider. You can change one of the resistor values by 1~20% and output won't change!! Of course you can buffer the result with an op-amp (which we usually did) for lower output impedance.
 
Anyway, a fun little circuit you can put together on a plug-in board.

Best,

[mawyatt]

Having an actual Chipdesigner here allows me to finally ask this question, especially since you pointed it out yourself:

Are such extra features, like your tennis raquet, just arbitrary, or do they have some kind of secondary use as well, like usage as some kind of fiduciary, or for quality checks?
Would your boss (I assume you have one ) let you put anything (within reason) in spare areas?

Yes they are arbitrary

I was the Chief Scientist/Engineer (before retiring last year) and so did ask my "boss", she said it was OK

I have taken some images for chips from other companies and many have all sorts of cute little cartoons imbedded in them, one had a cartoon character for every designer on the chip, 15 I recall!!

Best,

[Noopy]

...

As I have already written: Thanks for all your input! 

Please keep posting your beautiful chip images, really enjoy seeing these masterpieces of silicon displayed is such wonderful fashion.

I will do so. I really enjoy it. 

Got to finish up my voltage reference design built around the LM399, which your images and detailed circuit were superb and very helpful!! I had created a SPICE model for the LM399 and also a 6.2V Zener with NPN (2N3904), had to "tweak" the LM399 resistors to get the temp curve to look OK, but after seeing your image and schematic understood why things needed changing.

I´m glad to hear that! 

One circuit you might be interested in is a ultra-precision voltage divider that doesn't require precision components. If you use a CMOS FF and tie resistors between the Q and Qbar output, the center of the resistors is shunted to ground with a filter cap. The center voltage will be exactly 1/2 the CMOS FF Vdd voltage independent of the 2 resistor values, they don't even have to match!! What's happening is the Q and Qbar resistors from a simple voltage divider, say Vdd(R1/(R1+R2)), where R1 is tied to Q and R2 to Qbar. On the next clock edge Q and Qbar swap state and now the voltage divider is Vdd(R2/(R1+R2)). The capacitor averages these two voltages to exactly 1/2 Vdd. In real life the small Rdon of the NMOS and PMOS comes into play but can be swamped by R1 and R2 values, or use paralleled inverters on Q and Qbar outputs. This can produce results  around a ppm with 1% resistors We patented (5030848) this concept way back and used it in numerous designs, but the patent has long since run out.

If you have a 74AC74 or other FF and a 74AC04 or other inverter CMOS parts, and a couple 10K, 50K or 100K, or just about any resistor value and any cap. Just clock the FF at say 1~10KHz and supply 5.00000 volts as Vdd, then measure across the cap. You will get something like 2.49999 volts once you factor the DVM impedance with the Thevin eqv. impedance from the voltage divider. You can change one of the resistor values by 1~20% and output won't change!! Of course you can buffer the result with an op-amp (which we usually did) for lower output impedance.

Never heard of this! Interesting... 

Best regards!

[magic]

there are some techniques called bright field, dark field and epi that folks use

I believe what we are doing is epi-brightfield, i.e. direct reflection from the specimen. In darkfield (scattering by the specimen) ICs turn out pitch black, I suppose the surface must be mirror flat. The novel part is not as much the manner of illumination itself as the way it is realized without dedicated hardware.
https://resnicklab.wordpress.com/tag/darkfield/

You can produce poor man's condenser-less epi-darkfield by mounting thin SMD LEDs to the bottom of the lens right around the glass. I found 0603 size in 0.3mm thickness and managed to get them within 1.5mm of the optical axis and 1mm above the die, for almost 45° angle. No good, all the light reflects to the opposite side and the die looks exactly like in the example above.

This also follows along the discussions on de-convolution in post processing (AI sharping), even without the detailed lens/camera functions which is called blind de-convolution, folks are producing some really nice images that don't look too doctored up.

That's interesting. I may have a look because I'm still a fanboy of compact cameras in 2020 and they are often diffraction limited. I tried something along those lines of thinking once: I took a raw and sharpened it with radius set to the estimated diffraction radius and intensity "to taste". It kinda looked plausible. I wonder what processing the cameras themselves perform. I mean, you can now buy 20mpx 1/2.3" sensors coupled to lenses that are f/6 at telephoto - this has to involve some black magic or outright cheating

By the way, a convenient diffraction limit calculator for photographic lenses. I suppose it also applies to reversed lenses, then we are talking resolution on the object's surface.
https://www.cambridgeincolour.com/tutorials/diffraction-photography.htm

Your fine detailed image looks great, nice work!! 
What are the feature sizes?

It would look better with higher contrast and less noise from contrast enhancement
That's a jellybean bipolar chip, resolution on this image is some 1.5~2 pixels per micron.

I just realized there is another problem with this image: IIRC, the area around the second metal contact from the left was supposed to be yellow. Indeed, the camera has trouble capturing narrow yellow areas elsewhere, as shown below. This is probably due to low density of red pixel on Bayer arrays. Next time I will have to shoot raw to see what it really looks like and whether different algorithms would do a better job

[TheUnnamedNewbie]

I know it is a bit beside the point, but I would like to point out that you can actually image structures significantly smaller than the wavelength! I know of two methods, one a bit more reasonable at optical than the other.

First is to use materials with high refractive index. This shortens the wavelength, and thus allows you to image smaller things (you just need to 'expand' the field before it exits the high refractive index materials). This was/is used in IC manufacturing to pattern, with water lenses as final focus mechanism.

There are also near-field methods that can work (somewhat similar to what lytro does but in far-field, I think). However, these are kinda not reasonable at optical frequencies since it would require stuff like complex optical metamaterials.

Just wanted to share for those who are interested. Very interesting thread, am keeping an eye on it!
Modern DACs tend to be far more boring since they are more and more digital since we can scale digital more easily, and (digital - not trimming) calibration becomes more and more reasonable to do as we can fit more and more digital on a certain area.

[mawyatt]

I know it is a bit beside the point, but I would like to point out that you can actually image structures significantly smaller than the wavelength! I know of two methods, one a bit more reasonable at optical than the other.

First is to use materials with high refractive index. This shortens the wavelength, and thus allows you to image smaller things (you just need to 'expand' the field before it exits the high refractive index materials). This was/is used in IC manufacturing to pattern, with water lenses as final focus mechanism.

There are also near-field methods that can work (somewhat similar to what lytro does but in far-field, I think). However, these are kinda not reasonable at optical frequencies since it would require stuff like complex optical metamaterials.

Just wanted to share for those who are interested. Very interesting thread, am keeping an eye on it!
Modern DACs tend to be far more boring since they are more and more digital since we can scale digital more easily, and (digital - not trimming) calibration becomes more and more reasonable to do as we can fit more and more digital on a certain area.

Yes the semiconductor folks passed the diffraction barrier long ago, and today with 7nm features well beyond what any "sane" engineer/scientist would think possible , so hat's off to those folks  

Some of the microscope folks use high magnification lenses designed to have the front lens element and subject in oil, this certainly helps the image rendered because of the oil higher index. You could try that with chips, but I haven't, maybe someone can comment.

I've imaged one of the most advanced high speed DACs available a few years ago, it's not available as a die though and highly proprietary. The process used was SiGe BiCMOS and over 3/4 of the die (22mm by 18mm) is covered with digital and very boring indeed. The new DAC under development will likely be even more boring since it's all CMOS. ADCs are the same I believe, just a massive sea of boring CMOS.

Best,

[Noopy]

Today I have a Datel DAC-HZ12 for you:

https://richis-lab.de/DAC05.htm

It´s very similar to the Burr-Brown DAC80.







Can´t explain the ground wires going to the four MSB in the bottom left Corner. 

It works with the ICL8018 also used in the DAC80.




It seems that the trainee did the bonding... 




Same interesting buried zener as in the DAC80. 


 

[Noopy]

I have a new DAC for you, a AD565A:







It´s no big problem to identify the different blocks.
It works similiar to the DAC80 (https://www.richis-lab.de/DAC02.htm)




The emitter areas of the current sources show the ratio (4):8:4:2:1 you need for a good stability. 




And here the AD-buried-zener as you can find it also in the AD587 (https://www.richis-lab.de/REF06.htm).


A lot more pictures here:
https://www.richis-lab.de/DAC06.htm


If you want to know something special which google translate didn´t tell you ask me! 


 

[Noopy]

I have decapped a C565:

https://www.richis-lab.de/DAC08.htm






Interesting to see how accurate HFO copied the AD565. 




In my view they went without the buried zener. It seems this structure is a normal zener. Probably the buried zener was to complex.

 

[Noopy]

What do you think about a 18Bit-DAC with a accuracy of 18Bit? 






It took me some time to analyse the whole circuit. I won´t write down everything what I have already written in german:
https://www.richis-lab.de/DAC07.htm
But feel free to aks me whatever you want.




The AD1139 uses a AD588 (https://www.richis-lab.de/REF12.htm) to generate the references:
-10V for the ref-output
5V for the bipolar offset and the 14Bit-DAC for the 12LSBs
-5V for the 6 MSBs




The -5V is very important. Analog didn´t use the AD588-opamp (1µV/°C offsetdrift) but a OP27 (0,2µV/°C).
They also didn´t use the output amplifier of the OP27 but an external transistor. I assume Analog wanted to keep most of the power dissipation out of the opamp.
There is also a resistor divider. I assume they wanted to bring the -5V star near to -5V so that the transistor can regulate the voltage with lower currents.




The OP27 is a nice opamp! 
By the way: It seems that OP27 and OP37 once had different dies. PMI then developed a die that can handle OP27 and OP37.




Also interesting: The OP07 is clearly related to the OP27. A bit smaller and more simple.
The specifications of the OP07 are a bit worse but it has a smaller bias current (+/-1,2nA with bipolar input).




The resistor array is laser tuned. There are different structures offering different tuning intensities.




It looks like Analog tuned some bigger resistors then coated the array leaving some openings and then tuned the resistors to exactly the correct value. Interesting...




The MSB-DAC generates its currents with three 4053 analog multiplexers. Two 4042 are latching the data.




The -5-potential generates the currents for the output.
The three MSBs use the -10-potential to sink the current flowing from the output to the -5V-node. That´s good for accuracy because it relieves the -5-reference.







Analog needed a small regulator to generate a -5V voltage for the analog multiplexer HC4053.




A 14Bit-DAC generates the 12 LSBs. I´m pretty sure it´s a AD7535.
A AD712-opamp buffers the signal ground because the AD7535 causes a code dependend ground current.
The second AD712-opamp generates a voltage out of the current of the AD7535 to sum it at the resistor array.






Nice!




The MOS-switches are getting smaller and smaller with lower currents. Beginning with row 7 the transistors get longer and longer.




Why did they integrate protection diodes on the input of the output opamp?
The OP27 has protection diodes at the input allowing the flow of +/-25mA. The resistor should do enough current limiting. 

 

[magic]

What do you think about a 18Bit-DAC with a accuracy of 18Bit? 

I think it wasn't cheap
That's a lot of nice stuff.

BTW, regarding bias currents, OP27 has more because it runs much more current through its input stage to achieve low voltage noise. OP07 uses lower current so its current noise is better but voltage noise is rather lousy.

[Noopy]

Someone donated me two of them. 

Really some nice engineering! 

Thanks for you explanation regarding the bias currents! 

[doktor pyta]

Very interesting analysis.
It was a pleasure to watch and read.
Thanks!

[Vgkid]

That ad1139 gave a interesting teardown, thanks.

[Noopy]

Thank you for your positive feedback! 

[Noopy]

Following the AD565 and the C565 I have the TF536 for you:



The TF536 was built by HFO because in the late 1980s east germany needed a 16Bit-DAC.
In 1990 the development was canceled due to the liquidation of the company.
The TF536 uses the technology of the C565 (AD565).




The accuracy of the TF536 was specified with 13Bit. It was planned to produce it as C536 for CD-Players and as C5360 and C5361 with a digital correction to achieve the full 16Bit accuracy.
The correction is quite smart. The current generated by a sweep generator is subtracted from the output of the DAC and the result is compared with the ground potential. The algorithm uses only small parts of the ramp and the logic determines the time step from one trigger to the next. That´s quite accurate.
The TF536 has a so called Carry-Bit which is a second Bit9. That´s necessary to get enough correction clearance in positive and negative direction.
The developers wanted a Bit17 and a Bit18 to compete rounding errors but the TF536 doesn´t have them yet.
It was also not easy to construct a comparator good enough for more than 16Bit. That one is also missing in the TF536.




The die of the DAC is quite big: 6,14mm x 4,65mm
The digital die (a second one) would have added 7,7mm x 7,5mm




The 16Bit are divided in four 4Bit-parts. Bit1 to Bit12 are quite similar to the C565. Bit13 to Bit16 are built with individual current sinks.
There is no reference integrated on the die.




In contrast to the C565 the TF536 has the posibility to adjust the threshold of the digital inputs.




It seems that they had the possibility to compensate leackage currents by "switching" some current sources that have to be connected to the output.


More pictures here:

https://www.richis-lab.de/DAC09.htm

 

[Noopy]

Today I have the DAC08 for you:




This DAC08 was built by Raytheon.




The die is 2,45mm x 1,86mm.




Sorry, german... 




Every digital input signal controls a differential amplifier. The pad Vlc adjusts the threshold.




The current switches are getting smaller with lower currents so the current density stays the same.




The circuit is a little bit different than the schematic shown in the datasheet.
The reference current ist generated with a 500 resistor, not with a 1k resistor.
The currents 1 to 4 are generated as shown in the shematic using a R2R ladder.




The currents 5 to 8 are generated with less effort. To half the second I7 current Raytheon just connected two diodes (transistor-BE) parallel to the current switches.


https://www.richis-lab.de/DAC11.htm

 

[Mecanix]

So much gold! The thread's content, I mean 

[Noopy]

Thanks! 

[Noopy]

Today I have a older, less complex DAC371 for you:






It´s potted with something like silicone...




With some patience I was able to clean the circuit.




They needed a small wire because one line on the board had a cut.




The DAC371 generates a reference voltage. The eight resistors convert the voltage in binary weighted currents. If not shorted outside over one of the diodes each current travels to the common-base transistor. The low input resistance doesn´t influence the resistors and the collector isolates the resistors so that the voltage at the output doesn´t interfere with the resistors.




A datecode on a resistor, nice! 




The bonds on the pins are not very beautiful...
It seems they had kind of a fence in the package to place the die in the right place.




And at least one nice die picture... 


More pictures here:

https://richis-lab.de/DAC10.htm