SAR ADC oversampling 1

[zappedagain]

I have a 400 KHz analog channel that I am digitizing with a 16-bit Linear Technolgoies LTC2202 that can sample up to 10 MSPS and has a full power bandwidth of 380 MHz (it is designed for undersampling). Will I get a better SNR if I sample at 10 MSPS and average the ten samples than if I sample at 1 MSPS?

This seems like a no-brainer, because I can sample for 1000nS or take the average of ten 100nS samples; either way I'm looking at the same noise*time product so there is no difference. Or is that the wrong way to look at it?

What makes me wonder is that on a single sample this A/D only has a SNR of about 80 dB in my frequency range; that's 10000:1 (about 13-bits) instead of a full 65536:1 for 16-bits. I've seen that by averaging I can easily push this part past 16-bit performance.

Z

 

[IRstuff]

I think you are misreading the sampling rate to be the inverse of the sample time, which is not the case. Generally, such scopes have extremely tiny sample times, i.e., on the order of a nanoseconds, regardless of programmed sample rate.

However, many oversampling oscilloscopes use a pseudorandom sampling time to minimize aliasing, so averaging of the samples may actually increase the apparent noise, because you're averaging a non-stationary signal level. You can get away with this if you're sampling a slow or DC signal.

TTFN

FAQ731-376
Chinese prisoner wins Nobel Peace Prize

 

[VEBill]

Oversampling is supposed to gain one bit of SNR for each 4:1 increase in sampling rate.

 

[2dye4]

Assuming the nyquist criterion is satisfied the ability to gain extra bits is dependent on the quality of the noise present on the signal.

If no noise then every conversion produces the exact same binary result and averaging does nothing.

If the signal has white noise with a standard deviation of roughly 4 samples then averaging will improve the SNR.

This is very course approximation. To get the real story you need to study the variance of the average of the quantitized noise.

 

[IRstuff]

Again, just remember that averaging requires a certain level of steady-stateness in the input signal. Otherwise, you'd wind up averaging the actual changes in the input signal.

TTFN

FAQ731-376
Chinese prisoner wins Nobel Peace Prize

 

[benta]

Also, oversampling allows you to use a much simpler anti-aliasing filter with better phase and impulse response.

Benta.

 

[VEBill]

"If no noise then every conversion produces the exact same binary result and averaging does nothing."

Hmmm...

Doesn't such a situation allow one to ever-more-accurately determine exactly how close to zero the noise actually is? In other words, oversampling would still add additional significant bits, but they'd be significant zeroes in the absense of measurable noise.

?

 

[IRstuff]

Let's say the signal V is pegged at 0.4 of a step and the digital count is D. Then, if every result is D, you gain zero knowledge into what V actually is. A million samples later, you still only have D, and no indication as to where V really is.

Conversely, with an rms noise level of at least 1/12th of a step, you can statistically compare the times it crosses D vs. D+1 to determine where V is.

TTFN

FAQ731-376
Chinese prisoner wins Nobel Peace Prize

 

[2dye4]

""Also, oversampling allows you to use a much simpler anti-aliasing filter with better phase and impulse response.""

Careful here. No software process can overcome weakness of the anti alias filter.


As for the noise i believe it may be possible to add noise from a simple circuit to the desired signal thus enabling count averaging and SNR improvement.

 

[benta]

"Careful here. No software process can overcome weakness of the anti alias filter."

Who's talking about software? If he's sampling a 400 kHz signal at 1 Msps, the AA filter needs to be a h*ll of a lot steeper than if sampling at 10 Msps, where the stop band only needs to be at 5 MHz.

Benta.

 

[zappedagain]

Thanks all. These are interesting discussions. I'm sampling a transimpedance amplifier with a very high gain, so I have lots of white noise for my floor.

I believe this is starting to cross over into the sigma-delta A/D realm. I worked on a project a very long time ago where we sampled a comparator (1-bit A/D) at 12 MSPS. The resolution we could get in our KHz bandwidth was quite impressive; I'll have to dig up those notes.

Z