Using oversampling to enhance SAR ADC performance

Started by Myron Mychal October 17, 2003
I'm working with the new 56F630x series of Motorola DSP's (with an
onboard 12-bit ADC).  My application could make good use of 14-bit
data, so I planned on oversampling by a factor of 16 and filtering
with decimation to get an extra 2 bits out of the ADC.

I recently spoke with someone from Motorola who suggested that this
might not work with their ADC because it doesn't ensure equal
probability within the encoded voltage "bins" of the ADC.  Is this
true?  If so, does this mean that I cannot use the
oversampling/filtering technique to get another 2 bits out?

Myron Mychal
DSP Consultant
I'm working with the new 56F630x series of Motorola DSP's (with an
onboard 12-bit ADC).  My application could make good use of 14-bit
data, so I planned on oversampling by a factor of 16 and filtering
with decimation to get an extra 2 bits out of the ADC.

I recently spoke with someone from Motorola who suggested that this
might not work with their ADC because it doesn't ensure equal
probability within the encoded voltage "bins" of the ADC.  Is this
true?  If so, does this mean that I cannot use the
oversampling/filtering technique to get another 2 bits out?

Myron Mychal
DSP Consultant
Mike Rosing <rosing@neurophys.wisc.edu> wrote in message
news:<3F8B359B.5060406@neurophys.wisc.edu>...
> Myron Mychal wrote: > > I'm working with the new 56F630x series of Motorola DSP's (with an > > onboard 12-bit ADC). My application could make good use of 14-bit > > data, so I planned on oversampling by a factor of 16 and filtering > > with decimation to get an extra 2 bits out of the ADC. > > > > I recently spoke with someone from Motorola who suggested that this > > might not work with their ADC because it doesn't ensure equal > > probability within the encoded voltage "bins" of the ADC. Is this > > true? If so, does this mean that I cannot use the > > oversampling/filtering technique to get another 2 bits out? > > I don't know, but I'd think an empiracle test would tell you > more that a data sheet. Put in a really slow ramp or triangle > and run the over sampling routine. It should be pretty easy > to compare the stability of the real signal to the measurement. > > By adding some noise to the ramp you can see how bad the averaging > actually is. If nothing else you can characterize the problem > and then take it into account with software. >
Myron and Mike, Mike's idea about an empirical evaluation is a good one. At least it will tell you what one unit will do and if you test a few of them you may have a good idea. Don't get caught in the trap that the test cases don't represent a production run variability. You may never be able to reproduce or not able to reproduce reliably..... A little digging with the supplier might go a long way in matching your application with the underlying characteristics of their product. Even so, be cautious when using a chip in an application it's not intended for. About software: Once I worked on a system that had been mostly analog with some digital logic for process management. The system was tightly integrated with sensors into the real physical world which provided many of the operational situations / parameters, etc. Microprocessors had been around for some time by then and the engineers decided it was time to upgrade to a "much better" product. They assured me it would be much better because it would "have a computer inside". They were never able to express just what the computer was going to do!! Can anyone imagine that I was skeptical about this proposal? So, gee, I have to be similarly skeptical that software will fix "the problem" Mike! Probably that's not what you meant, eh? Fred
Myron Mychal wrote:

> I'm working with the new 56F630x series of Motorola DSP's (with an > onboard 12-bit ADC). My application could make good use of 14-bit > data, so I planned on oversampling by a factor of 16 and filtering > with decimation to get an extra 2 bits out of the ADC. > > I recently spoke with someone from Motorola who suggested that this > might not work with their ADC because it doesn't ensure equal > probability within the encoded voltage "bins" of the ADC. Is this > true? If so, does this mean that I cannot use the > oversampling/filtering technique to get another 2 bits out? > > Myron Mychal > DSP Consultant
First, as Mike wrote, try it and measure the results. Second, proper dithering before the converter can help. A requirement for dithering to yield the theoretical improvement --sqrt(n), where n is the oversampling ratio -- is that the precision and stability of the decision levels of your 12-bit converter meet the requirements of a 14-bit converter. Some do, some don't. Some improvement -- in resolution, if not precision -- should be possible, and more samples help. Can you oversample by 32? Will you settle for a bit and a half of improvement? Jerry -- Engineering is the art of making what you want from things you can get. &#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;&#2013266095;
> First, as Mike wrote, try it and measure the results. Second, proper
Thank you all for the suggestions. I'll try this with some slowly varying waveforms as suggested.
> dithering before the converter can help. A requirement for dithering to > yield the theoretical improvement --sqrt(n), where n is the oversampling > ratio -- is that the precision and stability of the decision levels of > your 12-bit converter meet the requirements of a 14-bit converter.
I don't think I'll have the resources available to do dithering before conversion, unfortunately.
> do, some don't. Some improvement -- in resolution, if not precision -- > should be possible, and more samples help. Can you oversample by 32?
Probably not. I'm sampling over multiple channels and I believe that an OSR of 32 wouldn't give me enough time to do the processing I need to do, though I'll take a closer look at it.
> Will you settle for a bit and a half of improvement?
Yes. In fact, just 1 bit would more than suffice, but I was trying to be conservative. Once again, to everyone, thanks for all of the input.