Siglent SDG7000A 350, 500 MHz and 1 GHz AWG's

[mawyatt]

We were quite interested in this new high end AWG, and were scheduled to have it for awhile to demo. Siglent NA kindly offered to allow us to evaluate, and our plan was looking into the NPR (Noise Power Ratio) at various frequencies up to 1GHz.

When we found out the price, we couldn't afford this and notified Siglent we were not able to acquire and didn't want to take time away from potential other customers.

Would be appreciative of someone looking into the NPR at various frequencies, and 1GHz if possible. NPR has a long history with us, culminating in evaluation of a very high performance DACs awhile back.

Will discuss further if folks are interested.

Edit: BTW does anyone know what DACs Siglent is utilizing, and what type amps in the output channels?

Best,   

[gf]

But the most important reason is the limited slew rate of the generator output. It can deliver a clean sine wave up to 1 GHz at 3 Vpp amplitude max., but it cannot deliver 500 ps risetime pulses at that amplitude level. This was an oversight on my part and so I’ve repeated the old measurement with comparable parameters and got quite similar resilts, see first screenshot.
...
It’s the slew rate limit of the output amplifiers. 500 ps rise time is certainly a bit of a challenge for a linear power amplifier, don’t you think?

Sure, it's challenging, and obviously reducing the amplitude did help.
But how can it reproduce a 1GHz sine wave at 3Vpp if it cannot reproduce this pulse at 3Vpp?
Don't underestimate the slew rate required for a 1GHz sine wave either, it's 9.42 kV/µs at 3Vpp! And its 10-90% rise time is 295ps.
The maximum dV/dt of a 1GHz sine wave is IMO almost 2x higher than the maximum dV/dt of this pulse, if both have the same Vpp.
So regarding dV/dt, the 1GHz sine wave is the bigger challenge.

Anyway, I’ve repeated the pulse train experiment at 1 MHz repetition rate and low 600 mVpp amplitude and it actually yields better results, see 2nd screenshot.
We still don’t get 612.5 MHz, but -3 dB at 378 MHz is still significantly better than the first attempt at 3 Vpp amplitude.

Looks very plausible now

The "flat region" in the spectrum is of course still limited by the smallest pulse width and rise time you can set. A rectangular pulse with 1ns width has a sinc-shaped spectrum (with its first zero at 1GHz, which we can clearly see in your FFT plot), and the edge softening does some additional low pass filtering. While this is obviously the limit achievable in pulse mode, my guess is still that the generator per se can likely produce even narrower pulses in AWG or DDS mode (leading to an even wider flat region of the comb).

[Performa01]

Would be appreciative of someone looking into the NPR at various frequencies, and 1GHz if possible. NPR has a long history with us, culminating in evaluation of a very high performance DACs awhile back.

Will discuss further if folks are interested.

I would be glad to help out if time and circumstances permit, but I'm not sure what you want and how it should be done.

What do you want to check? Since there is no explicit DUT, is it the linearity of the output stage? Is the dual tone test with internal (numeric) combiner not indicative enough for that?

[Performa01]

But the most important reason is the limited slew rate of the generator output. It can deliver a clean sine wave up to 1 GHz at 3 Vpp amplitude max., but it cannot deliver 500 ps risetime pulses at that amplitude level. This was an oversight on my part and so I’ve repeated the old measurement with comparable parameters and got quite similar resilts, see first screenshot.
...
It’s the slew rate limit of the output amplifiers. 500 ps rise time is certainly a bit of a challenge for a linear power amplifier, don’t you think?

Sure, it's challenging, and obviously reducing the amplitude did help.
But how can it reproduce a 1GHz sine wave at 3Vpp if it cannot reproduce this pulse at 3Vpp?
Don't underestimate the slew rate required for a 1GHz sine wave either, it's 9.42 kV/µs at 3Vpp! And its 10-90% rise time is 295ps.
The maximum dV/dt of a 1GHz sine wave is IMO almost 2x higher than the maximum dV/dt of this pulse, if both have the same Vpp.
So regarding dV/dt, the 1GHz sine wave is the bigger challenge.

You are right of course that the 1 GHz sine wave has a faster “rise time” and that it requires an even higher slew rate than a perfect pulse with equivalent rise time on top of that. And the generator manages this quite well, as can be seen in the first screenshot.

SDS6204_Sine_10MHz_W1ns_RT500ps_3V

Ironically, we get a “pulse width” of less than 500 ps and “transition times” of <350 ps. So your thoughts are perfectly valid: why can’t we get a 1ns wide pulse with 500 ps at full amplitude?

I only can remind you that we are moving on unspecified territory here. The minimum specified rise time is 1 ns. The SDG7102A can do rise times down to 500 ps, but nothing is guaranteed in this region. This seems to indicate that we are indeed moving at the edge of technical feasibility and I’m inclined to believe that Siglent engineers would have been happy to make this instrument even more powerful and impressive – but they had to stick with the limits of practical implementations.

By the way, we need not go as low as 600 mVpp to get closer to the 500 ps rise time. I’ve done a couple of measurements on 2 ns wide pulses with 500 ps rise time at 1 MHz to illustrate this.

At 3 V amplitude, the pulse is rather ugly – just look at the undershoot at the end of the falling edge:

SDS6204_Pulse_1MHz_W2ns_RT500ps_3V

But at 2.9 V, an internal relay clicks and we already get a completely different picture:

SDS6204_Pulse_1MHz_W2ns_RT500ps_2.9V

Rise time is close to 500 ps again and it seems to be just an unlucky circumstance that
The full 3V cannot be delivered without deactivating that particular attenuator stage.

[mawyatt]

Would be appreciative of someone looking into the NPR at various frequencies, and 1GHz if possible. NPR has a long history with us, culminating in evaluation of a very high performance DACs awhile back.

Will discuss further if folks are interested.

I would be glad to help out if time and circumstances permit, but I'm not sure what you want and how it should be done.

What do you want to check? Since there is no explicit DUT, is it the linearity of the output stage? Is the dual tone test with internal (numeric) combiner not indicative enough for that?

The NPR test is one of the more demanded linearity tests, since it fills the entire bandwidth under consideration with an equal power density waveform. Consider that the BW is occupied by multiple frequency channels and one selective channel has no energy within. This creates a somewhat like "notch" in the otherwise uniform density across the band spectrum of channels.

Not sure if Siglent has access to a NPR creating routine, or builtin type waveform. A waveform covering say 10MHz to 1GHz with 100 or more channels per decade and one channel (or more) without any signal within.

With the channels full to the "brim", say to -40dBm or higher if they can be generated, then the one channel without any signal will show a lower level where the other channels "spill" into the "well". The NPR is the "well" signal level minus the "brim" level in dBm.

Of course this needs to be viewed on a very high quality SA if these results are good, so the SA doesn't distort the final resulting spectral display.

Long ago in early 80 we needed to create such a NPR test for the RT SA we were developing based upon the CZT. At that time very expensive specialized instruments created the NPR signal by creating a broadband noise signal, frequency converting to the band where custom high performance band reject crystal filters were located, then frequency converted to the band of interest. The band reject crystal filters would "notch" out the desired signal and produce the NPR for use as a input test signal.

Due to cost constraints, we developed an alternative method of generating the NPR signal. Very simplified we supplied a high speed and resolution DAC with a time domain file created by an IFT which displayed the desired NPR frequency domain waveform with the selective "notch", this produced a frequency band and resolution based upon the DAC sample rate and memory depth (overall cycle time). The DAC output was either used directly, bi-phase modulated, or frequency scaled to the desired frequency band of interest. This worked beautifully and allowed us to evaluate our RT SA way back with a NPR parameter.

So our plan if we were able get ahold of this new AWG, was to search for any NPR files available from anywhere (hoping Siglent already had such), then loading the files for the SDG7000 and take a quick look with our modest SSA3021X+, then begin looking for an acceptable SA (maybe a higher end Siglent SA) to evaluate the SDG7000. The overall goal was to acquire the SDG7000 knowing the NPR performance (couldn't afford a new SA at that time), and proceed forward on a specialized project. With the cost of the AWG, then later the SA, this placed the concept out of reach at that time.

Our interest in the DAC utilized (and channel components) was in reference to some work done awhile back where we were involved with a new ultra-high performance dual DAC called Griffin from KS in SiGe BiCMOS, earlier renditions of this DAC became benchmarks for HS and HR DAC performance. Our interest was in use with EW applications.

Anyway, taken up enough of your time and space here, as a much longer story and such if interested.

Best,

[gf]

But at 2.9 V, an internal relay clicks and we already get a completely different picture:

Aaaahhh!

[2N3055]

Would be appreciative of someone looking into the NPR at various frequencies, and 1GHz if possible. NPR has a long history with us, culminating in evaluation of a very high performance DACs awhile back.

Will discuss further if folks are interested.

I would be glad to help out if time and circumstances permit, but I'm not sure what you want and how it should be done.

What do you want to check? Since there is no explicit DUT, is it the linearity of the output stage? Is the dual tone test with internal (numeric) combiner not indicative enough for that?

The NPR test is one of the more demanded linearity tests, since it fills the entire bandwidth under consideration with an equal power density waveform. Consider that the BW is occupied by multiple frequency channels and one selective channel has no energy within. This creates a somewhat like "notch" in the otherwise uniform density across the band spectrum of channels.

Not sure if Siglent has access to a NPR creating routine, or builtin type waveform. A waveform covering say 10MHz to 1GHz with 100 or more channels per decade and one channel (or more) without any signal within.

With the channels full to the "brim", say to -40dBm or higher if they can be generated, then the one channel without any signal will show a lower level where the other channels "spill" into the "well". The NPR is the "well" signal level minus the "brim" level in dBm.

Of course this needs to be viewed on a very high quality SA if these results are good, so the SA doesn't distort the final resulting spectral display.

Long ago in early 80 we needed to create such a NPR test for the RT SA we were developing based upon the CZT. At that time very expensive specialized instruments created the NPR signal by creating a broadband noise signal, frequency converting to the band where custom high performance band reject crystal filters were located, then frequency converted to the band of interest. The band reject crystal filters would "notch" out the desired signal and produce the NPR for use as a input test signal.

Due to cost constraints, we developed an alternative method of generating the NPR signal. Very simplified we supplied a high speed and resolution DAC with a time domain file created by an IFT which displayed the desired NPR frequency domain waveform with the selective "notch", this produced a frequency band and resolution based upon the DAC sample rate and memory depth (overall cycle time). The DAC output was either used directly, bi-phase modulated, or frequency scaled to the desired frequency band of interest. This worked beautifully and allowed us to evaluate our RT SA way back with a NPR parameter.

So our plan if we were able get ahold of this new AWG, was to search for any NPR files available from anywhere (hoping Siglent already had such), then loading the files for the SDG7000 and take a quick look with our modest SSA3021X+, then begin looking for an acceptable SA (maybe a higher end Siglent SA) to evaluate the SDG7000. The overall goal was to acquire the SDG7000 knowing the NPR performance (couldn't afford a new SA at that time), and proceed forward on a specialized project. With the cost of the AWG, then later the SA, this placed the concept out of reach at that time.

Our interest in the DAC utilized (and channel components) was in reference to some work done awhile back where we were involved with a new ultra-high performance dual DAC called Griffin from KS in SiGe BiCMOS, earlier renditions of this DAC became benchmarks for HS and HR DAC performance. Our interest was in use with EW applications.

Anyway, taken up enough of your time and space here, as a much longer story and such if interested.

Best,

Mike,

NPR test strictly calls for using noise.
Would multi tone test (with a gap) with sufficient density of frequencies generated be sufficient approximation or a strict noise spectrum would be needed for whole bandwidth in question, excluding "quiet" channel ??

Regards,

Siniša

[mawyatt]

Mike,

NPR test strictly calls for using noise.
Would multi tone test (with a gap) with sufficient density of frequencies generated be sufficient approximation or a strict noise spectrum would be needed for whole bandwidth in question, excluding "quiet" channel ??

Regards,

Siniša

Good question, not sure.

Our experience has been with all channels except a few filled to brim with signal across the channel space.

Originally technique was from a proper broadband noise source with notched out channels, this was the specialized dedicated NPR equipement. This equipement was designed around the old telephone channel requirements with 4KHz channels stacked against each other in long frequency chains. However, we did this with a DAC, actually a pair of advanced ECL driven high resolution and speed ADI hybrids, and a bunch of static ECL memory.

When we were shown the Griffin DAC back around 2009, the NPR file was already created and "looked" like a noise source with filtered channel, so no details there.

Best,

[2N3055]

Mike,

NPR test strictly calls for using noise.
Would multi tone test (with a gap) with sufficient density of frequencies generated be sufficient approximation or a strict noise spectrum would be needed for whole bandwidth in question, excluding "quiet" channel ??

Regards,

Siniša

Good question, not sure.

Our experience has been with all channels except a few filled to brim with signal across the channel space.

Originally technique was from a proper broadband noise source with notched out channels, this was the specialized dedicated NPR equipement. This equipement was designed around the old telephone channel requirements with 4KHz channels stacked against each other in long frequency chains. However, we did this with a DAC, actually a pair of advanced ECL driven high resolution and speed ADI hybrids, and a bunch of static ECL memory.

When we were shown the Griffin DAC back around 2009, the NPR file was already created and "looked" like a noise source with filtered channel, so no details there.

Best,

I'm not sure either, that is why I'm asking. My speculation here is that if we created dense enough multitone with random phase for the tones it could be "good enough" approximation, at least for order of magnitude type of measurement.
Not an expert though(not even close to that  ), would need to investigate to see it that supposition of mine is correct...

Take care.

[Njk]

BTW if one is interested in measuring the linearity, I think it can be done through measurement of sub-carrier isolation in a test OFDM-like signal (using a criteria like BER, etc). Perhaps it'll be more simple to tweak the existing codec SW than to invent something unique.

https://en.wikipedia.org/wiki/Orthogonal_frequency-division_multiplexing

[gf]

When we were shown the Griffin DAC back around 2009, the NPR file was already created and "looked" like a noise source with filtered channel, so no details there.

I guess it shouldn't be that difficult to create such an arbitary waveform pattern oneself, e.g. in Matlab or Octave?
Create Gaussian noise, FFT, notch out positive and negative frequencies of the desired channel, IFFT.

I found this paper from Analog Devices on the NPR topic: https://www.analog.com/media/en/training-seminars/tutorials/MT-005.pdf

SDG7000A has 14-bit vertical resolution, and according to figures 2 and 4, even an ideal 14-bit system is limited to a maximum of 74dB NPR, and this maximum can only be achieved when the noise has the optimal level with respect to the ADC's full scale range. Any smaller or larger level reduces the achievable NPR.

So the important question is, how well the utilization of the full-scale range of the AWG's ADC can be controlled. Is the output amplifier gain adjusted (almost) continuously with a VGA? Or is its gain only adjustable in larger steps, while the fine adjustement is done digitally, by amplitude scaling of the waveform? The latter would of course reduce NPR, even if the waveform was created with optimal noise level.

Any distortion and amplifier noise reduces NPR additionally, of course.

My speculation here is that if we created dense enough multitone with random phase for the tones it could be "good enough" approximation, at least for order of magnitude type of measurement.

You need of course enough frequency points falling into the notch. The above mentioned paper asks for "at least 25 to 50 samples within the filter notch". 512M samples enable a resolution of 4.66 Hz at 2.5 GSa/s. I think that's not bad

IFFT of a spectrum with flat amplitude and random phase leads to a bell-shaped PDF in the time domain, too. I'm not sure if it Gaussian, but it looks similar. And after quantization in the time domain and FFT, the amplitude is again "noisy" in the frequency domain, and no longer flat. So I would directly use Gaussian noise in the time domain, as it should be.

[gf]

But at 2.9 V, an internal relay clicks and we already get a completely different picture:

Aaaahhh!

Seriously, I guess they leave a few percent headroom for possible overshoot from the DAC's 2.5 -> 5 GSa/s upsampling filter. The output of this filter still needs to fit within the full-scale range.

[Performa01]

Not sure if Siglent has access to a NPR creating routine, or builtin type waveform. A waveform covering say 10MHz to 1GHz with 100 or more channels per decade and one channel (or more) without any signal within.

I don’t think so, but you can have a look yourself. The inbuilt arbitrary waveforms are almost exactly the same as in an SDG6000X (and presumably SDG2000X).

In general, when you find a way to perform this test with your SDG6000X, I can certainly repeat it with the SDG7102A.

You seem to be interested not only in the linearity of the output (buffer) stage, but also the DAC itself. In this regard, the SDG6000X should be even better, as it is 16 bits as opposed to the 14 bits of the SDG7000A. On the other hand, we have seen that its output stage is a little on the weak side…

With the channels full to the "brim", say to -40dBm or higher if they can be generated, then the one channel without any signal will show a lower level where the other channels "spill" into the "well". The NPR is the "well" signal level minus the "brim" level in dBm.

Of course this needs to be viewed on a very high quality SA if these results are good, so the SA doesn't distort the final resulting spectral display.

This begs the question, what NPR ratio do you expect to see, or in other words, what NPR do you need for your tests?

The SDG7102A can deliver up to +13.5 dBm at 1 GHz. If we use noise instead of sine waves, then the total output signal has to be limited to 160 mVrms for a reasonably flat spectrum up to 1 GHz, see first screenshot:

 
SDS6204 Pro H12_Noise_1GHz_160mV

We get a similarly steep transition from passband to stopband if we limit the bandwidth to e.g. 100 MHz:

 
SDS6204 Pro H12_Noise_100MHz_334mV

As to be expected, the measurement statistics for the 1 GHz noise spectrum shows exceptional high variations. Of course, this would have been even more impressive if I had the statistics left running overnight…

The noise floor of this measurement is about -105 dBV, this would limit the measurable NPR to about 36 dB – even if we were able to apply a notch in the noise spectrum. With a “real” SA (one that is definitely lower noise than the SSA3021X), the situation is rather worse. So I guess this test is not well suited for wideband systems. Not only do we need the notch function, but we also have to apply a lower frequency limit, so to define a relatively narrow band (e.g. GSM band) where the measurement can deliver useful results.

Especially the oscilloscope demonstrates the problem very clearly, because other than the usual SA, its input sensitivity and full scale voltage is totally transparent and very well controllable, with low noise – at least for frequencies above some 10 MHz. Even though the noise signal level was just 160 mVrms, I had to use the 500 mV/div vertical gain setting to avoid overrange. Consequently, the noise floor cannot be better than around -105dBV. A higher generator level cannot solve this, because it requires a lower input sensitivity, hence also higher noise floor. The only way to get a better S/N ratio than the ~36 dB demonstrated above would be to reduce the bandwidth. I would think, half the bandwidth should result in 3 dB better S/N, because the signal levels can go up by 3 dB. For a total of 72 dB S/N, we’d need to halve the bandwidth 12 times, which would be about ~250 kHz. Not very useful unfortunately...

EDIT: corrected the expcted process gain to 3 dB.

[gf]

The only way to get a better S/N ratio than the ~36 dB demonstrated above would be to reduce the bandwidth. I would think, half the bandwidth should result in 6 dB better S/N, because the noise floor gets 3 dB lower whereas the signal level can go up by 3 dB at the same time. For a total of 72 dB S/N, we’d need to halve the bandwidth six times, which would be about 16 MHz

Unfortunately, processing gain is only 10 dB per decade, or 3dB per octave of BW reduction, i.e. in order to improve the scope's NPR by 36 dB (from 36 to 72), you need to reduce the bandwidth of the noise signal by a factor of 10^(36/10), or approx. 3981. 1GHZ/3981 is only ~251kHz.

[Performa01]

The only way to get a better S/N ratio than the ~36 dB demonstrated above would be to reduce the bandwidth. I would think, half the bandwidth should result in 6 dB better S/N, because the noise floor gets 3 dB lower whereas the signal level can go up by 3 dB at the same time. For a total of 72 dB S/N, we’d need to halve the bandwidth six times, which would be about 16 MHz

Unfortunately, processing gain is only 10 dB per decade, or 3dB per octave of BW reduction, i.e. in order to improve the scope's NPR by 36 dB (from 36 to 72), you need to reduce the bandwidth of the noise signal by a factor of 10^(36/10), or approx. 3981. 1GHZ/3981 is only ~251kHz.

Unfortunately you are right. Somehow I wanted to gain the 3 dB at both ends – noise floor and signal level, but this is not going to happen. Limiting the noise signal to 500 MHz bandwidth yields close to 39 dB SNR, see attached screenshot.

SDS6204 Pro H12_Noise_500MHz_160mV

The rather low SNR is not a coincidence. While the noise floor is always distributed over the full bandwidth, the usual signals consist of just a few individual frequencies on which the total signal energy is concentrated. Yet in this scenario, where the signal itself is white noise, the signal energy is spread over the full bandwidth as well, which leaves little energy per frequency bin.

If we reduce the signal bandwidth to one half, then the noise signal density can be increased by 3 dB per frequency bin in order to get the same total signal energy (the one that doesn't overdrive the acquisition system). The noise floor remains the same, of course.

[mawyatt]

@ gf,

Walt's paper is a good reference wrt NPR, thanks for referencing such

BTW we've seen over 74dB a decade ago, and progress was slated to exceed 80 with future generation systems. We had achieved ~66dB back in early 80s with our RT SA based CZT as DUT, and this was the result of the CZT RT SA and our custom AWG producing the NPR waveform, so sort of a RRS result.

@ Performa01

We wanted to resurrect an old concept from ~2009 which involves eliminating/reducing electronic fratricide (self jamming) for co-located jamming and communication systems, especially where the comm systems are operating within the jamming environment, but that's about all we can say. We don't have any specific requirements, but wanted to see how well the new 7000 behaves wrt NPR. The 6000 isn't an option since we need to go beyond 500MHz, altho we could apply Bi-Phase modulation, or look into Super Nyquist techniques to extend the upper frequency range.

The idea was to demonstrate the basic concept with OTS equipment with minimum additional custom hardware, then proceed with a custom chip development. If funding ever materializes, then we can more seriously revisit this concept and project.

Best,

BTW none of the images after #53 are displaying? Anyone else have this issue? Edit, now they are showing???

[Performa01]

Here’s yet another example, this time for just 1 MHz noise bandwidth. The SNR doesn’t quite reach the expected 66 dB, but not far off either.

SDS6204 Pro H12_Noise_1MHz_160mV

I’ve tried to optimize my 1 GHz test results by avoiding any attenuator in the input path and adapting the noise signal level to better utilize the ADCs dynamic range.

At 500 mV/div, there is already the first 20 dB attenuator stage active, so I stick to the lowest sensitivity (highest V/div) where no attenuator is needed, i.e. 100 mV/div. For this, the optimum noise signal level for a 1 GHz wide noise spectrum is 50 mVrms, which in turn results in a maximum voltage of 700 mVpp.

SDS6204 Pro H12_Noise_1GHz_50mV

We achieve a SNR of 40.47 dB after all. Now let’s analyze that:

The signal level cannot be much higher, because the maximum Vpp value in the statistics is already pretty close to the full screen voltage of 800 mVpp.

The noise floor can be said to be -120.8 dBV pretty accurately (we also get very similar readings without any input signal).

The RBW of this measurement is 4.45 kHz. The noise floor of -120.8 dBV equals 912 nV. This results in a noise density of 13.67 nV/√Hz. This is of course much higher than what would be achievable at higher sensitivities, like 10 mV/div and below, but it is still not too bad.

In order to achieve 66 dB SNR, we’d need the noise floor to be at least 25.5 dB lower. This means just 48 nV, resulting in a noise density of 0.72 nV/√Hz. At the same time the instrument is supposed to deal with a 700 mVpp wideband signal without intermodulation or – even worse – compression.

Instead of noise density, we could calculate the required (3rd order) dynamic range. The total input signal is 50 mVrms. We just assume that the first mixer of the SA can handle the high crest factor, i.e. the 700 mVpp, I have serious doubts though.

Divided by 48 nV this is 1041666 or 120 dB. Well, the very best (with regard to dynamic range) spectrum analyzers that I’m aware of, like the R&S FSEA30, can reach 115 dB. As usual, this requires the instrument to be operated at a certain sweet spot, 115 dB aren’t quite 120 and I’m doubtful if this exceptional high dynamic range still applies to wideband noise signals with high crest factor, but then again, I have no means to try it out.

[gf]

At 500 mV/div, there is already the first 20 dB attenuator stage active, so I stick to the lowest sensitivity (highest V/div) where no attenuator is needed, i.e. 100 mV/div. For this, the optimum noise signal level for a 1 GHz wide noise spectrum is 50 mVrms, which in turn results in a maximum voltage of 700 mVpp.
...
The signal level cannot be much higher, because the maximum Vpp value in the statistics is already pretty close to the full screen voltage of 800 mVpp.

According to figure 4 in the paper, the optimum Gaussian noise signal level for an 8-bit ADC with a full scale range of 800mV_pp is 800/2/3.92 or ~102mV_RMS. The optimum is a trade-off between clipping and quantization noise. They do accept some clipping in favor of lower quantization noise. Note that on average only one out of ~11000 samples from a Gaussian PDF exceeds +/- 3.92 sigma, i.e. only a small fraction of the samples is really clipped. If the theory from the paper also works in practice, then doubling the level from 50mV to 100mV should give you an extra 6 dB.

[Performa01]

According to figure 4 in the paper, the optimum Gaussian noise signal level for an 8-bit ADC with a full scale range of 800mV_pp is 800/2/3.92 or ~102mV_RMS. The optimum is a trade-off between clipping and quantization noise. They do accept some clipping in favor of lower quantization noise. Note that on average only one out of ~11000 samples from a Gaussian PDF exceeds +/- 3.92 sigma, i.e. only a small fraction of the samples is really clipped. If the theory from the paper also works in practice, then doubling the level from 50mV to 100mV should give you an extra 6 dB.

Well, I’ve tried that and it didn’t work out in practice.

SDS6204 Pro H12_Noise_1GHz_100mV

Yes, we gain some signal level, but only 5 dB on average instead of six, simply because of the signal loss due to clipping. All amplitude measurements show overload (e.g. >963.529 mVpp). The visible range is 800 mVpp and these measurements indicate that we’re actually using up the full ADC range – and beyond, which of course cannot be measured anymore.

The noise floor on the other hand has risen by about 2.5 dB because of the ADC overloading. All in all we get barely 3 dB improvement on the SNR – the marker list shows now 44.5 dB, which would be 4 dB more, but this is only because the level at 200 kHz happens to have improved by 6.5 dB, which is an outlier within the whole picture.

Your proposal was for an 8 bit ADC, whereas we have 12 bits here, yet the ill effects shown here have certainly nothing to do with ADC resolution or quantization noise. If we look at the signal in the time domain, we can also see that it is quite saturated.

[gf]

Your proposal was for an 8 bit ADC, whereas we have 12 bits here,

Sorry, I missed that. Then your noise profile is certainly dominated by other sources while quantization plays just a minor role. I was wondering anyway that you got >40dB. I had not expected so much with 8 bits, 50mV_RMS @500mV/div and 1GHz BW @5GSa/s.

Well, I’ve tried that and it didn’t work out in practice.
...
The noise floor on the other hand has risen by about 2.5 dB because of the ADC overloading. All in all we get barely 3 dB improvement on the SNR – the marker list shows now 44.5 dB, which would be 4 dB more, but this is only because the level at 200 kHz happens to have improved by 6.5 dB, which is an outlier within the whole picture.

Thanks for trying. The predicted +6dB have not materialized as your noise profile is different from the assumptions in the paper, but the NPR still did improve! If you vary the noise signal level, you should still be able to find a point where the NPR is at its maximum (so that any lower or higher level results in a lower NPR).

Yes, we gain some signal level, but only 5 dB on average instead of six, simply because of the signal loss due to clipping.

Hmm, if the noise signal is Gaussian, then clipping at +-4 sigma should not result in a 1dB loss of RMS level, but the expected order of magnitude is rather 0.001dB, i.e. virtually negligible. Either the PDF is not Gaussian, or something else strange is going on.

[mawyatt]

Instead of noise density, we could calculate the required (3rd order) dynamic range. The total input signal is 50 mVrms. We just assume that the first mixer of the SA can handle the high crest factor, i.e. the 700 mVpp, I have serious doubts though.

Divided by 48 nV this is 1041666 or 120 dB. Well, the very best (with regard to dynamic range) spectrum analyzers that I’m aware of, like the R&S FSEA30, can reach 115 dB. As usual, this requires the instrument to be operated at a certain sweet spot, 115 dB aren’t quite 120 and I’m doubtful if this exceptional high dynamic range still applies to wideband noise signals with high crest factor, but then again, I have no means to try it out.

The R&S maybe the best available SA wrt to dynamic at 115dB (haven't looked), maybe for general purpose commercial use. Way back in 2009 when we witnessed the Griffin DAC (recall better than 76dB NPR then), KS pulled out an experiment SA to demonstrate this level of NPR (their standard high end SA wasn't good enough as we saw), don't know if this ever became a general purpose commercial product tho.

There are specialized types of spectrum measurement devices, SA like if you will, that far exceed 120dB Dynamic Range. One might consider these as more like specialized wide-band tunable receivers than SA, but they still cover a relatively large frequency range while processing narrow band signals in the presence of extremely high interference (mostly as in adversarial jamming) over a relatively wide-bandwidth.

Our interest in the NPR measurement was related to allowing communication within our own intentional jamming signals as well as observing any "other" signals of interest (much more involved but not open for discussion). The core DAC NPR performance starts the whole process of signal generation and gets "enhanced" by additional post processing and frequency translation, however the DAC is not a simple straight forward n-bit DAC but has dynamic error correction and predistortion.

Thus our interest is in how good the new AWG performs and the DAC and signal chain utilized.

Best

[tautech]

New firmware for SDG7000A models.

Version： V1.1.1.31
85 MB
https://int.siglent.com/upload_file/zip/firmware/Signal_generator/SDG7000A_V1.1.1.31_EN.rar

Release notes.
Support frequency hopping function
Built in EasyWaveX supports multi pulse function
Add the function of erasing user waveform files with one click
Solve the problem of system unresponsiveness when loading incorrect IQ waveform files
Solve the problem of occasional loss of locking when using OCXO
Solve the problem of loading network disk waveform files getting stuck
Fixed the bug of false alarms for overvoltage

[Performa01]

Time Resolution

Inspired by the application presented in this thread:

https://www.eevblog.com/forum/testgear/keysightagilent-81160a-highest-time-resolution/

I thought I’ll check this with the SDG7102A. To cut a long story short, I was able to enter at least 12 relevant digits for the pulse period, e.g. 123456789.123 ns, which is equivalent to 123 milliseconds at one picosecond resolution. Well, there’s the question: is this still accurate?

If we are satisfied with a time interval of 123 µs and a time resolution of just 10 ps, then an SDS6204 H12 Pro oscilloscope is good enough to check the accuracy of the generator.

For this test, I’ve fed channel 4 with a 2 ns wide 600 mVpp pulse with 500 ps rise time, at a pulse period of exactly 123456,78 ns:
 

SDS6204_Pro_H12_Pulse_W2ns_RT500ps_P123456780ps_Zoom_Run

As can be seen, period measurement is off by just 3.4 picoseconds – I’d think that the main portion of this error can be attributed to the DSO.

Pulse width is spot-on and transition times are pretty close, even though the measurements use the data from the main window. The zoom window shows the pulse in detail. There is a little variation in the pulse shape; this is because of the fast 500 ps rise time whereas this generator is only specified down to 1 ns.

[huverson]

Yippie, welcome youngest member to the lab 

So I obviously haven´t been long time here on this forum and neither especially active - but - I have already been helped quite some and if I can give something back by for example some tests with the new device just get in contact!
Best
Martin