FFT Spectrum Analysis Reviewed

[Performa01]

FFT appears to be a strongly demanded feature in DSOs and there is an ongoing discussion how it compares to a ‘real’ spectrum analyzer and particularly whether it can be useful with just 8 bits resolution in a standard DSO.

Traditionally, there are two camps of totally opposite opinions:

• 8 bits might be low resolution and have a poor dynamic range in itself, but this can be vastly improved by certain resolution enhancement techniques. In particular, the FFT in itself acts as multitude of band pass filters (number of ‘bins’) which leads to a dramatic noise reduction for each individual bin. In short, we just need a high number of FFT sample points in order to achieve a decent dynamic range. The catchword here is ‘process gain’.
  
• 8 bits are low resolution and the dynamic range of even an ideal 8-bit A/D converter is only 48dB. This is just the laws of physics and we cannot get anywhere, no matter what mathematical tricks are used for signal processing. In short: crap in -> crap out.
  

Even though I’ve always held the view that the latter position would be the much safer bet, it is still obvious that the truth would be somewhere in between these two extreme positions.

People advocating the ‘process gain’ as the solution for the limitations of an 8-bit system sometimes even pull out some textbook article to back their claims. The catch is – as always – that it is just theory. As we all know, in theory there is no difference between theory and practice, but in practice there is…

Of course, the theory is right if we have an ideal system, which cannot exist in real world, plus there’s even more requirements…

If we actually had an ideal 8-bit ADC, it should be obvious that a signal with a peak-peak amplitude of less than 1 LSB just cannot produce any change in the output code of the ADC. Consequently, any signal lower than some 48dBfs (decibel below full scale) would just be suppressed. That means, we can well have a very high SNR (signal to noise ratio) due to ‘process gain’ but still the usable dynamic range is limited to 48dB.

To overcome this, we need to add some dither – usually white noise – and we will inevitably have to use averaging to hopefully get some meaningful readings for the spectral components below -48dBfs whose readings are now very unstable since they are riding on top of the dither.

And here the troubles start already. Who actually adds dither? It is not so easy either. For a 50 ohms input, we can just use an off the shelf resistive power combiner and a suitable wideband noise source, but what if we want to preserve one of the very few advantages that the general purpose DSO can have over a genuine SA, namely the high impedance input? We’d need to construct a power combiner with two 1Mohms ports for the scope and the probe and one 50 ohms port for the noise source. In any case this means 6dB attenuation, thus increasing the noise figure and lowering the sensitivity by this same figure.

On the topic of noise figure, the amplitude of the dither should be reasonably accurate. If it is too low, the dynamic range will not be extended by much – if at all. If it is too high, it will decrease the SNR, but then again, this is something where a narrow ‘bin bandwidth’, i.e. a high number of FFT points actually helps.

If the DSO front end and/or the signal source are very noisy, we might get the dither for free. But this is rather ill-defined. The scope frontend might contribute enough noise at the high vertical sensitivity settings such as 5mV/div or lower, but is not guaranteed to do so at gain settings where the input amplifier works at x1 gain. The noise from the signal source itself is an external factor, hence totally variable and mostly unknown. On top of that, the noise of a wideband input amplifier with equalization networks might be anything but white.

This is not the end of the theoretical consideration though. In reality, any fast 8-bit ADC, will exhibit an INL (integral nonlinearity) of at least ±0.5 LSB. What if just the least significant bit, e.g. the transition from 0 to 1 on the ADC output happens to be off by half a LSB? The answer should be obvious – the amplitude accuracy for signals < -48dBfs would be off by ±6dB.

Well, one might think we don’t need high accuracy for levels that low, but I cannot think of many applications where a 1:4 uncertainty could be tolerated.

For an FFT analyzer, the ADC does a comparable job as the RF mixer in an SA. And the highest concern for such a mixer is linearity, as this will affect the IPi3 (third order input intermodulation intercept point) and subsequently the intermodulation-free dynamic range. Consequently, the ADC will produce a dense (numerical) fog of distortion and intermodulation products at its output, which the FFT algorithm will show, thus obscuring the real low level signals.

Of course, we might be lucky and the particular ADC in our scope might have a relatively low INL for the number range used for close to zero signals, but this is nothing we can rely on. Furthermore, there will always be some nonlinearity, no matter how much some particular ADC exceeds its worst case specifications. So one ADC might produce 10dB lower intermodulation products than some other, but we’ll still never be able to get the intermodulation-free dynamic range as calculated in the textbook.

Now for some practical evaluation…

[Performa01]

8-bit FFT
Now we are going to explore what really can be expected from the FFT in an 8-bit DSO. For all the following examples, full scale is ±2V, which is equivalent to +16dBm with 50 ohms termination at the scope input. All measurements use a ~1.95MHz FFT bandwidth with 2Mpoints, so we have 1M bins each of them being 1.863Hz wide. Blackman Harris window is used throughout and the actual -60dB bandwidth is about 7 times the bin width, resulting in a frequency resolution of some 13Hz.

Let’s start with a 990kHz signal at +10dBm without averaging (FFT8_2M_+10dBm)



This looks impressive at a first glance. Even without averaging, the noise level appears quite low, but there are lots of spurs and for the reasons already discussed, we cannot know which of them result from the imperfections in the signal itself and which are just due to the shortcomings of the ADC. Even though it would be tempting to relate the more prominent spectral lines to the signal, we just don’t know for sure at levels below -50dBfs.

The strongest spur is -48.62dBm and since full scale would be +16dBm it looks like we have a SFDR (spurious free dynamic range) of 64.62dB – some 15dB better than what we’d normally expect from 8 bits.

The automatic measurement of the SFDR already hints something different though. The average value reads only 50.76dBc, which would still be 56.76dB total, hence 7dB better than the 8 bit dynamic. But the min/max readings show that the spurious signals change a lot and if we take the minimum of 41.76dBc we get a total SFDR of 47.76dB. But we still don’t know if this SFDR is related to the signal or the measurement system.

Now let’s lower the signal level to -10dBm and use averaging to get a clearer picture. The signal related spurs should drop accordingly, whereas the ADC noise should remain largely unaffected (FFT8_2M_-10dBm_avg)



The spurs from the signal generator have clearly decreased together with the signal level and are now all the more indistinguishable from the ADC noise. If we look at the SFDR measurement, the uncertainty has decreased to less than 3dB thanks to averaging and the total SFDR is now some impressive 69dB.

Now let’s find out the lowest level we can get decently accurate results by lowering the signal level step by step. With a signal level of -30dBm we are at -46dBfs what is close to the limit of the 8 bit dynamic range. No averaging this time, but the screenshot has been taken at a moment where the most prominent spurs were close to a maximum (FFT8_2M_-30dBm_max)



The signal level can still be measured quite accurately, whereas the spurious signals have decreased a lot. This could lead to the false impression that they were related to the signal, but in fact they are just part of the intermodulation ‘interference fog’ caused by the ADC nonlinearity. Since every possible ADC output has its own individual INL error, this creates a multitude of distortion and intermodulation products. Now that pretty much only the LSB of the ADC is used, the number of spurs decreases accordingly and even the SFDR reading of the automatic measurement is quite stable at some 12dBc. The total SFDR is therefore some 58.3dB, which is not consistent with the last measurement at a signal level of -10dBm. Let’s investigate that a bit further…

What does it look like with a signal level of -40dBm? (FFT8_2M_-40dBm_max)



Now there are only a couple of spurs left, but at the same time the signal level has become quite inaccurate. Automatic measurement shows an average of -36.16dBm, which is about 4dB high. Only the minimum reading is close to the true value and there is 6.7dB difference between minimum and maximum.

Let’s see if we can get better results by adding some white noise as dither. Due to the INL error of the LSB discussed earlier, a peak to peak noise level of 1 LSB isn’t enough to get accurate readings, hence a much higher level of noise of some 120mVpp had to be applied (FFT8_2M_-40dBm_Noise-18dBm)



Now the level display is almost spot-on, but we also start seeing a lot of spurious signals, as was to be expected. After all, we are now able to accurately measure a signal at -56dBfs and the signal is still almost 14dB above noise/spurs, resulting in a total SFDR of nearly 70dB.

The dither needs not be white noise, we could try a triangle wave as well. For optimum accuracy, a peak to peak level of 58mV was required and the frequency was chosen at 30Hz (FFT8_2M_-40dBm_Triangle_58mVpp)



This throws off the peak level detection of course – the automatic measurement now sees the dither instead of the weak test signal. Other than that, the signal displays correctly and the SFDR is just a little bit worse than it was with the white noise dithering.

Now let’s lower the signal level by another 10dB, thus applying -50dBm while still using the 30Hz triangle dither level of 58mVpp (FFT8_2M_-50dBm_Triangle_58mVpp)



We still get an fairly accurate level display and there’s only one spurious signal left. So we can actually quite accurately measure a signal at -66dBfs and the signal is still more than 18dB above the spur, resulting in a total SFDR of healthy 84dB. The noise floor is down at <-80dBm, so we could claim a first order dynamic range of more than 96dB – that’s crazy performance from an 8 bit system, is there no difference between textbook theory and practice after all?

Well, hopefully it could be demonstrated that the ‘process gain’ – while it has a real effect – isn’t something we can use to make an 8 bit system perform like a higher resolution one. The high number of FFT sample points reduces the bandwidth of the individual bins and lowers the quantization noise, but it cannot prevent the nonlinearity of the ADC from generating lots of unwanted mixing products that immediately rise as soon as the signal level increases and will get particularly bad as soon as the real dynamic range of the 8 bits is entered. Just a look at the screenshots for higher signal levels show that very clearly and the confusing number of spurious signals isn’t exactly my idea of accurate and reliable measurement.

Here’s a demonstration of the results for the -50dBm signal without dither (FFT8_2M_-50dBm)



The noise level is down to -88dBm and there are only two spurs left, so the picture doesn’t look too bad at first glance, but then the signal level is 6dB low and not stable – the signal variation is more than 8dB.

Finally, if averaging is used for the very same situation as above, we get the following picture(FFT8_2M_-50dBm_avg)



While the noise level is even lower at <-90dBm and the signals appear much more stable on the screen, the situation hasn’t significantly changed otherwise.

[Performa01]

16-bit FFT
From the previous demonstration, it should be clear by now that an 8-bit system has its tight limitations, no matter what the number of FFT sample points is. While the resolution enhancement can indeed work as long as the right amount of dither is added, which would just be 1 LSB pk-pk if we had an ideal ADC with zero INL, the linearity error of any real ADC will not only require a much higher dither level in order to get accurate results, but also produce a high number of spurious signals that are not random noise, hence cannot be reduced by resolution enhancement and filtering. Only a better ADC (equivalent to a better mixer in a conventional SA) can provide a true improvement in terms of accuracy and usable dynamic range.

To prove this, we now look at the FFT of a scope that has a true 16bit ADC. When I say ‘true 16bit ADC’ I actually mean it that way. No other ‘16bit solution’, achieved by some dubious tricks, be it hardware (combining several lower resolution ADCs) let alone software (resolution enhancement by averaging/filtering) can replace the real deal, because none of these tricks will improve the linearity. Particularly in spectrum analysis, linearity is paramount.

There is no point in having some virtual ADC that might have 16 bits of resolution, but only 8 bits accuracy/linearity. Just take my word for that.


For all the following examples, full scale is ±2V, which is equivalent to +16dBm with 50 ohms termination at the scope input. Many measurements use just 4kpoints, as there’s no need for a ridiculously high number of FFT points just for the sake of noise reduction. Blackman Harris window is used throughout, where the actual -60dB bandwidth is about 7 times the bin width.

Let’s start with the same 990kHz signal as before at +10dBm with averaging. This makes sense here, as there are no spurious signals due to ADC nonlinearity to be expected. FFT length is 16kpoints, so we have 8192 bins each of them being 203.5Hz wide. The -60dB resolution bandwidth is therefore some 1.4kHz (FFT16_16k_+10dBm_avg)



Now this is finally a measurement that lets us actually judge the signal quality. Even with just 16k points, the noise level is more than 100dB below full scale. The specified SFDR for this scope is >96dB, so we can confidently assume that any spurs down to -80dBm belong solely to the signal. Only below that level, the puzzle whether it’s the signal or the analyzer starts.

There aren’t many spurious signals and the strongest one is -59.51dBm, whereas on the 8bit FFT it was -48.62dBm despite the 2M points analysis. But then, it cannot be surprising that we were not able to seriously measure a signal level below 1 LSB of the 8bit ADC back then.
 
The automatic measurements of both the signal level and the (signal related) SFDR are rock solid now, the latter being stable to 0.01dB for 1000 captures!

Here’s another example, this time a different signal – 2MHz at +10dBm coming from an analog synthesizer, where we can clearly see the much stronger phase noise compared with the DDS generator used for all the other measurements. This time, Analysis bandwidth is 5MHz and FFT length is only 4k, so we have 2048 bins each of them being 2.441kHz wide. The -60dB resolution bandwidth is therefore some 17kHz. Note that this is also at a different vertical gain, resulting in ±1V and therefore +10dBm full scale (FFT16_4k_+10dBm_avg)



Because of the higher sensitivity, we can trust everything down to -86dBm to be signal related. Even with just 4k FFT, we get a detailed picture, low noise level and stable measurements.

I stick with the analog synthesizer signal, as this allows me to reach very low levels easily. Now we look at it at a level of -60dBm. FFT length is 4k again (FFT16_4k_-60dBm_avg)



With only 4k FFT, the noise level is below -98dBm, so we get a first order dynamic range of 108dB. There are no spurs and the only other signal we see is the 2nd harmonics from the test signal. The automatic measurements are still very stable. Particularly, the signal level is accurate to 0.07dB and has only 0.06dB variation – and that level is only 174µV right now!

Well, let’s go down to -90dBm now (FFT16_4k_-90dBm_avg)



Now the signal level is -100dB below full scale, so we’re finally leaving the dynamic range of the 16bit converter. We still get an accurate amplitude measurement, even though there is a variation of 3dB now. Why does it work here, when it failed on the 8bit scope? Well, it must have something to do with the dithering, as even in an exceptional low noise scope like this there will be still enough for a 16bit ADC, where the LSB is just about 30.5µV for this particular gain setting.

After all, we are measuring a ~7µVrms signal at a full scale of 2Vpp (700mVrms) here, a task where we’d otherwise need at least a 5 digit DMM …

Going to the extremes, we now look at -110dBm with 2M points (FFT16_2M_-110dBm_avg)



For better visibility, I’ve zoomed to the amplitude range below -93dBm. Noise floor is ridiculously low at about -124dBm now, but at levels that low, the nonlinearities of even a 16bit converter start producing some spurious signals again. Note that there is now a 2nd harmonic only a couple dB lower than the signal itself, that is clearly generated within the ADC.

The signal level is still fairly accurately displayed at -110dBm, so a first order dynamic range of up to 120db can be usable under certain conditions. At a signal level of just 2.2µV, automatic measurements are no longer useful, as they now pick mains hum as the strongest signal at a level of ~4.4µV. Well, I don’t know if it is the shielding of the scope frontend or – more likely – just the cabling in my highly mains-contaminated home lab. I forgot to choose a double shielded cable and just used a high quality RG58 for my tests.

[Performa01]

2 Tone Comparison
Up to now, we’ve just examined the first order dynamic range and the test scenarios didn’t really reflect any practical application. No one would look at a single signal far below full scale of the instrument, rather increase the sensitivity to make good use of the dynamic range available. In case of the first order dynamic range this is actually full scale or a couple dB below the 1dB compression point on a real SA.

Much more relevant is the analysis of small signals in presence of a large one. This applies to measurements of signal harmonics (of course sub- and non-harmonics as well), modulation spectra and phase noise. So if there’s always at least one strong signal, we might get away without an additional dither for the 8-bit system maybe?

Let’s start with 8 bits and two signals at 990kHz, +10dBm and 1010kHz, -40dBm. Full scale is +16dBm and 2M points are used again. FFT bandwidth is 1.953MHz. The required dynamic range is 56dB, thus exceeds what would normally be possible with just 8 bits (FFT8_2M_+10dBm_-40dBm)



Well, this doesn’t look too bad and is actually quite usable. For some unknown reason, automatic measurement of the peak level has a strong variation which isn’t real – I probably forgot to reset the statistics after setting up the signals.

Now decrease the level of the 2nd signal to -50dBm (FFT8_2M_+10dBm_-50dBm)



The result is still usable, so we have just proved that we can work within a 66dB first order dynamic range. However this result cannot be generalized, as it heavily depends on the individual properties of the ADC. This particular one seems to be quite good in terms of INL near zero, but there’s no guarantee that all 8bit scopes will behave the same in this regard – not even all scopes of the same type from the same manufacturer.

Now let’s take it to the extremes and lower the 2nd signal level once more to -60dBm (FFT8_2M_+10dBm_-60dBm)



Well, the good news is that the measurement of the 2nd signal level is still accurate, but it now submerges in the jungle of spurious signals. Not a problem as long we know exactly what signals to measure, but we won’t be able to identify unexpected spectral components like non-harmonics when analyzing a signal. So as much as I would love to conclude that we can get a usable dynamic range of 76dB from just 8 bits, it is not really true. While it basically works, it is no fun for sure to hunt for your wanted signal in the noise, even though that particular noise has a discrete spectrum that leaves slots for the test signal to fit into.

In comparison, the 16bit FFT has no troubles with a signal level of -70dBm. At an FFT length of 2M points we need no averaging to get a noise floor close to -100dBm and we can be confident that all spurs down to at least -80dBm are real, i.e. belonging to the signal itself. In the following screenshot, we can even see the phase noise of the 990kHz signal quite clearly (FFT16_2M_+10dBm_-70dBm)



Now the question might arise, how does that compare to a real SA? Well, as much I’d love to demonstrate that on a R&S FSEA30, right now I only have a cheap USB spectrum analyzer (Signal Hound SA44). While amazingly useful, it still has lots of limitations but nevertheless is good for a very descriptive demonstration.

So here it goes. One of the limitations is that the span must not exceed 1MHz so to get at least 200Hz RBW and this still takes a sweep time of nearly 4 seconds. Nothing like 210ms for a 5MHz sweep with 30Hz RBW as I did on the 16bit scope…

So I used a slightly different setting for various reasons. The signals are now at 990kHz and 1.5MHz and the levels are +10dBm and -80dBm, a 90dB difference! (SA_RBW200Hz_+10dBm_-80dBm_avg8)



Now this is very revealing. We can clearly see that a traditional superheterodyne receiver with an IF bandwidth of 500kHz is ‘swept’ (probably more like ‘stepped’) and an FFT processing is done to provide the narrower resolution bandwidths. The dynamic range of the superhet is quite good and clearly exceeds 100dB at 200Hz RBW. But the FFT shows a signal dependant noise floor, which makes the signal source look much worse than it is. From the +10dB 16bit FFT of the DDS signal we know that the phase noise doesn’t start any higher than -90dBm and looks completely different to what we see here. One of the reasons for the high noise floor might be the flat top window, which has less side lobe suppression compared to Blackman Harris. Apart from that, I strongly suspect that the ADC might only be 14 bits.

Of course, if we lower both the span and the RBW, FFT noise will drop too, but for wider spans, this SA clearly cannot compete with the 16bit FFT.

With a Span of 50kHz and a RBW of 25Hz (13Hz would be possible too), we can resemble the original setup, i.e. make the 2nd signal 1010kHz again. The noise floor is now low enough to pass the test with -80dBm for the 2nd signal (SA_RBW25Hz_+10dBm_-80dBm_avg8)




The two-tone intermodulation tests can be found here:
https://www.eevblog.com/forum/testgear/fft-spectrum-analysis-reviewed/msg921837/#msg921837

[uncle_bob]

Hi

Nicely done. There really is no "free lunch" in this case.

There are other studies comparing 16 bit ADC based systems to (older) "Real SA's". In most cases, the modern 16 bit ADC wins out. Of course, that only works up to a limited frequency point due to sample rates. Digitizing a 22 GHz spectrum at 44 Gs/s and 16 bit samples is not quite possible (yet).

Bob

[Performa01]

Hi Bob,

thank you!

I am not at all surprised about that studies. After all, most 'real' SAs only have a SFDR of some 70dB. That's also why I particularly mentioned the R&S FSEA 30 - which I will get eventually, once I've rebuilt my home lab to get a larger bench and deeper shelves, that is

The real deal would be the professional 24 bit audio interfaces, but then their frequency coverage is just too limited for anything other than audio...

[Kleinstein]

There are spectrum analyzers that combine classical heterodyne and FFT: so the heterodyne is used to bring the signal down to a range where the higher resolution ADC works well and than use FFT to get the high resolution / small RBW and fast "scan". To get the real dynamics and less spurious they need the higher resolution ADCs, though they might get rid of some of the spurious signals using more than one IFs.

[Performa01]

There are spectrum analyzers that combine classical heterodyne and FFT: so the heterodyne is used to bring the signal down to a range where the higher resolution ADC works well and than use FFT to get the high resolution / small RBW and fast "scan". To get the real dynamics and less spurious they need the higher resolution ADCs, though they might get rid of some of the spurious signals using more than one IFs.

Well, that's exactly what the SA44 does. But as I wrote, I'm not entirely sure if that device actually utilizes a 16 bit ADC - it might be just 14 bits. Nevertheless, the performance isn't bad at all, especially considering its pricepoint.
The third order dynamic range is only about 65db, but then again, many of the old boat anchors haven't been any better either - and this doesn't even seem to have changed a lot...

[uncle_bob]

There are spectrum analyzers that combine classical heterodyne and FFT: so the heterodyne is used to bring the signal down to a range where the higher resolution ADC works well and than use FFT to get the high resolution / small RBW and fast "scan". To get the real dynamics and less spurious they need the higher resolution ADCs, though they might get rid of some of the spurious signals using more than one IFs.

Well, that's exactly what the SA44 does. But as I wrote, I'm not entirely sure if that device actually utilizes a 16 bit ADC - it might be just 14 bits. Nevertheless, the performance isn't bad at all, especially considering its pricepoint.
The third order dynamic range is only about 65db, but then again, many of the old boat anchors haven't been any better either - and this doesn't even seem to have changed a lot...

Hi

The issue is still that analog mixers (plus amps plus filters) are limited in their IP3 performance. That has not changed a whole lot over the last decade or two. A relatively cheap modern 16 bit ADC does better on IP3 than the mixer chain you can afford to put in a piece of test gear. That assumes the test gear costs less a pretty nice new car. Of course if the sky is the limit, there are some pretty crazy ADC's out there ....

Bob

[hendorog]

There are spectrum analyzers that combine classical heterodyne and FFT: so the heterodyne is used to bring the signal down to a range where the higher resolution ADC works well and than use FFT to get the high resolution / small RBW and fast "scan". To get the real dynamics and less spurious they need the higher resolution ADCs, though they might get rid of some of the spurious signals using more than one IFs.

Well, that's exactly what the SA44 does. But as I wrote, I'm not entirely sure if that device actually utilizes a 16 bit ADC - it might be just 14 bits. Nevertheless, the performance isn't bad at all, especially considering its pricepoint.
The third order dynamic range is only about 65db, but then again, many of the old boat anchors haven't been any better either - and this doesn't even seem to have changed a lot...

FWIW, the new software Spike doesn't support anything less than 6.5kHz RBW below 16MHz on my SA124B.
I don't know why that is, but it made me think that perhaps you are in a frequency/span where the SA is particularly weak?

Also I get much better sweep rates with Spike, and just about everything else works much better so its well worth the upgrade - if you can get off XP.

[egonotto]

Hi

16-bit FFT

To prove this, we now look at the FFT of a scope that has a true 16bit ADC. When I say ‘true 16bit ADC’ I actually mean it that way. No other ‘16bit solution’, achieved by some dubious tricks, be it hardware (combining several lower resolution ADCs) let alone software (resolution enhancement by averaging/filtering) can replace the real deal, because none of these tricks will improve the linearity. Particularly in spectrum analysis, linearity is paramount.

Please can you say which PicoScope oscilloscope you use.
Do you think that the PicoScope 5243A has a true 12 Bit till 16 Bit ADC.

Thanks

egonotto

[ADT123]

Hi

16-bit FFT

To prove this, we now look at the FFT of a scope that has a true 16bit ADC. When I say ‘true 16bit ADC’ I actually mean it that way. No other ‘16bit solution’, achieved by some dubious tricks, be it hardware (combining several lower resolution ADCs) let alone software (resolution enhancement by averaging/filtering) can replace the real deal, because none of these tricks will improve the linearity. Particularly in spectrum analysis, linearity is paramount.

Please can you say which PicoScope oscilloscope you use.
Do you think that the PicoScope 5243A has a true 12 Bit till 16 Bit ADC.

Thanks

egonotto

I just did some captures with a couple of PicoScopes and also a LeCroy WaveAce 102 in all cases the input was a 1.8V pk-pk 20kHz low distortion sine wave.  SFDR figures:

LeCroy WaveAce 102    50dB (approx)
PicoScope 5244B          68dB (8 bit mode)
PicoScope 5244B          78dB (16 bit mode)
PicoScope 4262            93dB (16 bit)

One thing I found was that at 16 bits the quality of the BNC cable matters.  Swapping a cheap (foil wrapped internally) cable for a nice braided one gave a couple of dB extra.  Suspect the 5244B would have got 80dB.

The PicoScope 5244B does change its resolution in hardware rather than software trickery (see
https://www.picotech.com/library/oscilloscopes/flexible-resolution-oscilloscope ) but does not get as good a result as the 4262.  I suspect this is because the PicoScope 5000 is a 1GS/s, 200MHz scope whilst the PicoScope 4262 is a low frequency 5MHz scope optimized for high dynamic range.

Will post screenshots in a monent

[ADT123]

Screenshots from the last post.  One tip that can help when using scopes that offer large numbers of FFT points is to use the maximum number of points and a higher bandwidth than you want to view.  Next zoom in on the area of interest.  In these examples the signals I want to see are below 1MHz but by doing an FFT to 5MHz the higher frequency noise is not folded back into the area of interest so lowering the noise floor.

Lecroy - about 50dB


PicoScope 5244B in 8 bit mode.  68dB


PicoScope 5244B in 16 bit mode.  78dB


PicoScope 4262 16 bit.  93dB

[Performa01]

There are spectrum analyzers that combine classical heterodyne and FFT: so the heterodyne is used to bring the signal down to a range where the higher resolution ADC works well and than use FFT to get the high resolution / small RBW and fast "scan". To get the real dynamics and less spurious they need the higher resolution ADCs, though they might get rid of some of the spurious signals using more than one IFs.

Well, that's exactly what the SA44 does. But as I wrote, I'm not entirely sure if that device actually utilizes a 16 bit ADC - it might be just 14 bits. Nevertheless, the performance isn't bad at all, especially considering its pricepoint.
The third order dynamic range is only about 65db, but then again, many of the old boat anchors haven't been any better either - and this doesn't even seem to have changed a lot...

FWIW, the new software Spike doesn't support anything less than 6.5kHz RBW below 16MHz on my SA124B.
I don't know why that is, but it made me think that perhaps you are in a frequency/span where the SA is particularly weak?

Also I get much better sweep rates with Spike, and just about everything else works much better so its well worth the upgrade - if you can get off XP.

Well, I don't think the SA124 can be compared to the SA44, as the latter explicitly goes down to 1Hz. Also, my experience with this neat little device did not hint on any particular weakness in the lower frequency regions, quite the contrary. If anything, noise floor gets notably worse above 3GHz.

A RBW of 6.5kHz sounds really odd, particularly on the lower frequencies. Maybe the high bandwidth SA124 actually has some limitations at its lower bandwidth limit...

And yes, for various reasons that would be off topic to explain here, I'll stick with Win XP on my lab computer, so the latest Software isn't an option for me. Maybe I'll take my company notebook to my lab some day, to check how much of an improvement the new Spike software actually is for the SA44...

[Performa01]

Please can you say which PicoScope oscilloscope you use.
Do you think that the PicoScope 5243A has a true 12 Bit till 16 Bit ADC.

Hi,

it is a 4262, since this is the only true 16bit realtime scope I'm aware of.

The 5000-series 'variable resolution' scopes are certainly an improvement, but it involves combining several lower resolution ADC cores. While this gives much better results than any 8-bit DSO, it is still not as good as a true 16 bit ADC. Just look at the dynamic performance specifications published by Pico Technology:

Scope	crosstalk	THD	SFDR	Noise
4262	50000:1	-95dB	96dB	8.5µVrms
5000	400:1	-70dB	70dB	70µVrms

The maximum bandwidth is limited to some 31MHz for the 5000 series in 16bit mode, whereas it is just 5MHz on the 4262. But all the other specs are much better on the 4262. For me, it is just the right tool for low frequency signal analysis at a very attractive price.

[bozidarms]

Rigol DS4012(4052)   ------------    FFT about 70dB (without trace averaging)

[hendorog]

FWIW, the new software Spike doesn't support anything less than 6.5kHz RBW below 16MHz on my SA124B.
I don't know why that is, but it made me think that perhaps you are in a frequency/span where the SA is particularly weak?

Also I get much better sweep rates with Spike, and just about everything else works much better so its well worth the upgrade - if you can get off XP.

Well, I don't think the SA124 can be compared to the SA44, as the latter explicitly goes down to 1Hz. Also, my experience with this neat little device did not hint on any particular weakness in the lower frequency regions, quite the contrary. If anything, noise floor gets notably worse above 3GHz.

A RBW of 6.5kHz sounds really odd, particularly on the lower frequencies. Maybe the high bandwidth SA124 actually has some limitations at its lower bandwidth limit...

And yes, for various reasons that would be off topic to explain here, I'll stick with Win XP on my lab computer, so the latest Software isn't an option for me. Maybe I'll take my company notebook to my lab some day, to check how much of an improvement the new Spike software actually is for the SA44...

Yes I understand being stuck on XP - had the same problem but eventually bit the bullet. Now I'm stuck on Windows 7...

The SA44 also has this limitation - this is from the SA44 manual:
"For spans larger than 99MHz or sweeps that start below 16MHz, the RBW can be set between 6.5 kHz and 250kHz"

"All RBWs are available for spans 200kHz or less."

My thought was that the old software may not be indicating this limitation correctly.
If you are lucky the limitation may not be present in the older software, but that would be odd - why add a new limitation?

[G0HZU]

FFT appears to be a strongly demanded feature in DSOs and there is an ongoing discussion how it compares to a ‘real’ spectrum analyzer and particularly whether it can be useful with just 8 bits resolution in a standard DSO.

I have several decent (but ~30yr old) RF spectrum analysers here covering up to 22GHz but I still use the FFT feature on my little Tek TDS2012 from time to time for certain tasks. However, since I bought an Agilent E4406A vector signal analyser I tend to use the FFT feature on the TDS2012 less.

The TDS2012 can do some things that the big old spectrum analysers can't such as look at very narrow bandwidth RF signals that require sub 10Hz or sub 1 Hz RBW.

A typical test signal would be an on/off keyed signal at 70MHz that is being keyed at a 1Hz rate. The TDS2012 can display the spectrum of this with RBW <1Hz and so can the E4406A but none of the other analysers can do it 'properly'.

A typical ham/CB use for the FFT feature in the TDS2012 would be to look at narrowband AM/FM/SSB signals up to 145MHz using either test tones or human speech. The displayed range is only 50-60dB but this is usually enough to see if a transmitter is basically OK or if it has an issue with excessive bandwidth. It can refresh the display quickly in this mode and show speech peaks etc. Not quite real time but close enough and MUCH quicker than a typical swept spectrum analyser.

[hendorog]

Rigol DS4012

With trace averaging set to max (8192) noise floor is down below -100dB

[G0HZU]

But the (Signalhound) FFT shows a signal dependant noise floor, which makes the signal source look much worse than it is. From the +10dB 16bit FFT of the DDS signal we know that the phase noise doesn’t start any higher than -90dBm and looks completely different to what we see here. One of the reasons for the high noise floor might be the flat top window, which has less side lobe suppression compared to Blackman Harris. Apart from that, I strongly suspect that the ADC might only be 14 bits.

I've not used a Signalhound analyser but I think that at least some of the noise on the Signalhound plots may be the noise profile of the PLL used for its local oscillator. Try measuring the Signalhound up at 3 or 4GHz with a similar span and this noise level will be much higher because the 1st LO will be much noisier up at these frequencies. I'd expect it to be in the ballpark of 30dB higher. But that's just a guess based on how I think the Signalhound 44 LO system is designed

[Performa01]

Intermodulation Tests
We still have just examined the first order dynamic range so far. But there are also applications, where we want to measure spectral components with tiny amplitude in presence of more than just one strong signal, e.g. when observing modulation spectra but also – and primarily – when testing mixers, amplifiers, attenuators and the like for their intermodulation performance. After all we want to measure the performance of the device under test – and not of the analyzer. This is where the third order dynamic range becomes important, and this will always be worse than the first order one.

We start with the 8 bit FFT, feeding two tones at 990kHz and 1010kHz and a level of -13dBm. The scope is in the ±200mV full scale range, which is equivalent to -4dBm. FFT length is 2M points and FFT bandwidth has been chosen to ~4.5MHz. The picture is zoomed so we can clearly see what’s going on near the test signals (FFT8_2M_Two_-13dBm_avg_zoom)



Once again, this looks surprisingly good. Both 3rd order intermodulation products at 970kHz and 1030kHz are at -85dBm and it looks like we have no troubles measuring them, even though they are at -81dBfs, hence far outside the 8bit dynamic range.

Looking at the whole picture, we can see that there are indeed stronger spurs at odd frequencies, not related to the test signals or their harmonics and associated intermodulation products (FFT8_2M_Two_-13dBm_avg)



Nevertheless, this test provided a 3rd order dynamic range of 72dB, thus competing with many ‘real’ spectrum analyzers.

Now for the 16bit FFT. Same settings as before, except for the FFT bandwidth now being exactly 5MHz. Here is the zoomed picture again (FFT16_2M_Two_-13dBm_avg_zoom)



There is little guessing now. The intermodulation product at 970kHz is much stronger at -107.7dBm and also slightly outside the genuine dynamic range of the ADC. There are only few spurs, some of them a little stronger at -105dBm. The whole picture reveals nothing new, just the harmonics of the test signals together with their associated intermodulation products. All in all we get an intermodulation-free dynamic range of 94.6dB and only very few ‘real’ spectrum analyzers can beat that.

Should anyone wonder what this two-tone test signal looks like in the time domain, here it is (FFT16_2M_Two_-13dBm_time)



Pretty much a 100% amplitude modulated waveform. A real (analog) amplitude modulator has to be very good and tweaked a lot in order to produce a DSB (double sideband) signal like this – in fact it is next to impossible. A little bit of the carrier will always be left. But the reverse direction works beautifully. Just mix the two sidebands together and there is no delicate carrier suppression anymore we have to worry about, bingo!


Finally, let’s have a look at the ‘real’ SA. These devices are much more sensitive than any DSO, but conversely can’t handle high signal levels. Even though -13dBm is not a problem for the SA as far as the first order dynamic range is concerned, however this level is too high for the 3rd order dynamic range test. Usually, -20dBm is used to determine the third order intermodulation intercept point and that’s exactly what we want to do now.

Very similar setup as before, two tones at 990kHz and 1010kHz, both at -20dBm, but due to the restrictions of the SA the scan is limited to the range from 800kHz to 1.2MHz and RBW (resolution bandwidth) has to be much higher at 100Hz (SA_RBW100Hz_Span400kHz_Two_-20dBm_avg16)



Now this is gorgeous. We get a 3rd order dynamic range of 91dB right from the start!

This would be equivalent to an IPi3 of +20dBm, which is 10dB higher than e.g. on a Rigol DSA 815.

But this is not the entire truth. This looks like the FFT results we’ve seen before, so we have to repeat the test with the two tones further apart in order to force the SA44 to do a real scan. So the tones are now at 1.2MHz and 1.7MHz and the span has been increased to 2MHz. By this, the RBW has increased to 400Hz as well (SA_RBW400Hz_Span2MHz_Two_-20dBm_avg16 )



Now we’re talking! The 3rd order dynamic range is only 60dB now, but the intermodulation products are still well above the noise floor, so that is not the end yet. Let’s calculate the IPi3 first – it is the familiar +10dBm! As the noise floor is about -105dBm, we could lower the signal amplitudes by 10dB and the intermodulation products would submerge in the noise at -110dBm. But we could just use the internal attenuator instead (SA_RBW400Hz_Span2MHz_Two_-30dBm_avg16)



Bingo! The signal levels appear unchanged, but interestingly the noise floor has dropped too despite activating a 10dB internal attenuation. i.e. the NF (noise figure) seems to have even improved – quite the opposite to what I would have expected. My only explanation is that the displayed noise isn’t real noise from the frontend, but just noise from the ADC. If anyone has a better explanation, I would like to hear (read) it.

Either way, the intermodulation products have dropped by 20dB as expected and thanks to the fact, that the noise floor has not increased, the third order dynamic range is now respectable 80dB.

With the internal 10dB attenuator, the IPi3 is +20dBm of course, which wouldn’t mean a thing were it not for the not elevated noise level, which actually gives us a rather decent intermodulation-free dynamic range.

So in the end I have to apologize to the SA44, which I have previously accused for only having 65dB 3rd order dynamic range. In fact, it is much better than that – at least at lower frequencies.

[Performa01]

One tip that can help when using scopes that offer large numbers of FFT points is to use the maximum number of points and a higher bandwidth than you want to view.  Next zoom in on the area of interest.  In these examples the signals I want to see are below 1MHz but by doing an FFT to 5MHz the higher frequency noise is not folded back into the area of interest so lowering the noise floor.

That's some excellent suggestion!
That's also a reason why it's nice to be able to use a high number of sample points for the FFT, so we can avoid aliasing and still preserve a low RBW.

PicoScope 4262 16 bit.  93dB

Well, your signal source appears to be very clean indeed - but are you really sure it is that good? You are measuring the 3rd harmonic of the test signal, which in theory could be generated by the ADC nonlinearity, but this is unlikely, since it is still within the genuine dynamic range of the 16-bit ADC and Pico Technologies specify <-95dB THD for this scope. So I would rather suspect it's coming from the signal source...

EDIT: Btw, I've just noticed, you have the internal signal generator on. I've found that this also produces some spurs, even when the siggen output is not connected. So for high sensitivity measurements it should be turned off.

[Performa01]

I've not used a Signalhound analyser but I think that at least some of the noise on the Signalhound plots may be the noise profile of the PLL used for its local oscillator. Try measuring the Signalhound up at 3 or 4GHz with a similar span and this noise level will be much higher because the 1st LO will be much noisier up at these frequencies. I'd expect it to be in the ballpark of 30dB higher. But that's just a guess based on how I think the Signalhound 44 LO system is designed

Well, here's one for 1GHz (SA44_RBW200Hz_Span1MHz_1GHz_-10dBm_avg8)



Right at the moment I do not have any decent signal generator for 3GHz, so I have to stick with something of unknown qualities. But if the signal happens to be ok, then your guess would be about right (SA44_RBW200Hz_Span1MHz_3GHz_-10dBm_avg8)

[egonotto]

Hi,


thanks for the answers.

Regards
egonotto

[hendorog]

So to determine the source of the noise in the last post, is it simply a case of adding external attenuators and comparing?
i.e. If the level of noise doesn't drop then it's coming from the SA. If the level of the noise does drop by the same amount as the attenuation then it's coming from the signal source.

To test that theory, here is the SH tracking gen connected to my SA124B first with no external attenuators, and then with 40dB of external attenuators.
The plan is to determine if the noise is coming from the tracking generator or the spectrum analyser.

First image is with no external attenuators and reference set to 0dBm. It shows noise close to the signal starting at around -60dBm.
Second image is with added 40dB external attenuators, and reference decreased by 40dB. This shows noise close to the signal at around -100dBm

Therefore the noise is coming from the TG and not the SA - is this correct?

Third image shows just the noise floor - I just moved the TG off frequency to 3.5 GHz.