[SOLVED] Simulation method for high resolution ADCs

[je01911]

Hi all,

I am designing the high resolution SAR ADC upper 14 bit.

However it can not be simulated using spectre because of very long simulation time.

Is there the accurate and fast method to simulate the high resolution ADC?

Thanks

 

[kemiyun]

What kind of simulations you are looking for?

If it is INL DNL: People generally split the capacitances in SAR ADCs. So once you simulate the inl dnl performance of the one side of the split cap array, you do not have to run those codes again. Assume that you split your cap array in half, you can follow a sequence like 0.......127, 128, 256, 384, 512 .... This will reduce the number of codes you need to simulate significantly. And I've tried, it works without a problem up to 16 bit arrays. I mean you can get nearly the same inl dnl without running all the codes. As an example, INL DNL simulations of 16 bit converter took around 45 minutes or something with this method. But the worst part is, if you design it correctly, there should be minimum amound of static error, so this is not performance validation, it just tells you that you designed it correctly. So I guess you need to run monte carlo to actually see how you did which again takes much longer.

If it is SNR or spectrum related: There is not much you can do. But you can switch to Spectre APS, and force it to use many cores or open many threads. Also, this might not be something you want, you can increase the frequency to fs/2. In this case you can just take a 32 or 64 point FFT, just to make things a little faster. But beware, this might give you deceiving results if your ADC doesn't startup in a short time (you might want to give some time for ADC, like 1 or 2 sample), and since you are taking less samples, their effect would be extreme. Sorry, there is not much I know that could help if sine wave is considered, because basically this should be the fastest thing spectre solve to my knowledge. In short, take less samples, use APS.

One last addition, just to clear things, in industry, generally ADCs with more than 16 bit resolution are referred as high resolution converters. I know it is not that important, but still I believe this is the notation.

I know I said one last addition, but here is another one: It can be simulated with spectre. I did , actually if you can find any industrial product with your specifications, it is most probably simulated using spectre or hspice, but I believe spectre is more popular.

 

[je01911]

Thank you, kemiyun.

Your reply is very helpful to me.

In exactly, I want to find to simulate the dynamic performance like SNR or spectrum related.

I have several question.

1. What is the mean of 'switch to Spectre APS'.

I already use the multi-threads by chaging the simulator option.

Is there different things between multi-thread in simulator option and Spectre APS ?

Or do you know the documentation about Spectre APS?

2. the relationship between short points and input frequency.

You said that 'increase to fs/2 (it means nyquist rate) and just take 32 or 64 point'.

I also has experience using short point like 32 or 64 to simulate higher resolution ADC fastly.

However, i don't know which relationship between input frequency and the number of points.

And please explain the 'this might give you deceiving results if your ADC doesn't startup in a short time' in detail.

Thanks

 

[kemiyun]

1-Spectre APS is the high performance extension of Spectre. It really increases speed of simulation, without reducing precision. If available to you, use it. You can select it under the high performance simulation tab in test editor or ADE L. Documentation is already included, cadence guys write good documentation really.

2- That could allow you to force the ADC to the edge. I mean for lower frequencies you will have better results generally. Just push it to the end, so you know your limits. Also if you want to take FFT, longer than 1 period, this is the way to go. More number of periods at low frequency means increased number of points. Just take 32 or 64 points FFT. Simple yet effective. BUT: Be careful when it comes to selecting the points (this is not that important for ADCs, but extremely important for DACs), as you would like your simulator to actually simulate the point you are sampling. You wouldn't want an artificial point that is in between two simulation points.

If you don't wait for internal signals to settle, I mean if you don't allow for example your references to settle, the first samples will definitely make an error. This error can lead to deceiving results. But it is nice that whatever you do it is very hard to get a better result when doing FFT. It is easy to mess the spectrum up, but it is hard to get a better result than the real result, making FFT a unique function among many

 

[je01911]

Thank you kemiyun!