NEW Keysight HD3

[maxwell3e10]

You can make scope display always look like an eye diagram or change the definition of waveform update rate just to prove your point. But under standard usage as shown in Dave's video, the waveform update rate can't exceed horizontal span. "It's called math" to quote this video at 3m 42 sec:

[Someone]

You can make scope display always look like an eye diagram or change the definition of waveform update rate just to prove your point. But under standard usage as shown in Dave's video, the waveform update rate can't exceed horizontal span. "It's called math" to quote this video:

Math is garbage in garbage out. There is nothing linking horizontal sweep time to update rate in the general case. As I showed earlier with a pictorial representation:
https://www.eevblog.com/forum/testgear/new-keysight-infiniivision-hd3-oscilloscope/msg5634313/#msg5634313
There is nothing special about the trigger than happens to fall after the re-arm time of a specific scope, it is equally as valid a trigger as those possible triggers before and after it. Ideally a scope would have zero re-arm time and zero blind time and that is not impossible.

[maxwell3e10]

Ideally a scope would have zero re-arm time and zero blind time and that is not impossible.

Yes and that is exactly when waveform update rate is equal to 1/(horizontal span time) !

[Someone]

Ideally a scope would have zero re-arm time and zero blind time and that is not impossible.

Yes and that is exactly when waveform update rate is equal to 1/(horizontal span time) !

So we now have you agreeing that there could be a scope with zero dead/blind time? (the second example in this image)

It is not a significant leap to then consider the aggregate of the last 4 "capture systems". Nothing there is impossible, with a short (approaching zero) rearm time that aggregation could occur within a single instrument hanging off a single ADC/front end. Bits that need to be wider/faster were elaborated by David Hess:
https://www.eevblog.com/forum/testgear/new-keysight-infiniivision-hd3-oscilloscope/msg5635307/#msg5635307
Trigger rearm time would only need to be shorter than the rate of prospective/possible/reality trigger conditions arriving to capture every single edge.

[EEVblog]

Running the HD3 vas MXO4 FFT test again:

[egonotto]

Hello,

the peak at 50 MHz with the HD3 is about -108 dB, but with the MXO4 -100 dB. I don't understand that. I also don't understand the 15 dB difference in the noise floor.

Best regards
egonotto

[maxwell3e10]

In conventional scope operation only one trigger is allowed in the horizontal time span. So signals triggered on periods 1, 2 and 3 cannot be combined. Otherwise one would get something akin to a messy eye diagram for general (not perfectly periodic) signals. If you want to argue that one can make an eye diagram with faster update rate, I will grant you that, but that is not what we are talking about for basic waveform update testing.  Signal triggered on period 4 can be combined with 1, giving zero dead time.

[hpw]

Running the HD3 vas MXO4 FFT test again:

OK, as on HD3 looks like a 128k FFT, but as I read the HD3 user manual, as I understand it is an 64k FFT with decimation or boxcar detection algo's.

Nice would be a long CSV sample data save, to get the rtHz figures as we have done on SDS2000 HD ... SDS7000 12 bit's

[tautech]

In conventional scope operation only one trigger is allowed in the horizontal time span. So signals triggered on periods 1, 2 and 3 cannot be combined. Otherwise one would get something akin to a messy eye diagram for general (not perfectly periodic) signals. If you want to argue that one can make an eye diagram with faster update rate, I will grant you that, but that is not what we are talking about for basic waveform update testing.  Signal triggered on period 4 can be combined with 1, giving zero dead time.

Pray tell what is a conventional scope ?
If you mean a DSO that statement is wildly untrue.

[Someone]

the peak at 50 MHz with the HD3 is about -108 dB, but with the MXO4 -100 dB. I don't understand that. I also don't understand the 15 dB difference in the noise floor.

Details matter, if the memory depth of the FFTs were different while the shorter length was decimated and/or poorly anti-aliased then the noise floor can shift upwards.

[egonotto]

Hello,

thank you, how did you create the picture?


Best regards
egonotto

[Neganur]

Did you read it ?

1.5% of Full scale. For 500µV/div and 1mV/div a 2mV/div is used. Which means 6%  and 3% respectively.

Ofc I read it. Let's stick with DC gain accuracy, and I will ignore that the Siglent has just 200 MHz BW.

Because of magnification, for HD3 it is +/-6% for the 500uV/div and 3% for the 1mV/div range, and also +/-1.5% at 2mV and 5mV/div.
Then, Siglent specs it at +/-1% for 5mV/div and beyond, while the HD3 spec is still +/-1.5%

So,
better than HD3 at less than 2mV/div, (2x and 4x, no?)
equal at 2m/div, and
1.5x better than the HD3 at 5mV/div and beyond, (not generally "[...] better vertical DC accuracy then their new wonder.. By factor of 3X....")

the thing I disagreed with, was the blanket statement.

Personally, I don't take the "typical" spec seriously. Keysight does not list a typical spec. The datasheet does not mention, is it maximum or what is it, so it is difficult to compare to what Siglent found is a typical spec. I assume it is a maximum (worst case) spec. I don't mind that you compare it to the typical spec but it makes me wonder, what do you think a typical spec value is like for the HD3?

[Someone]

thank you, how did you create the picture?

Numerical simulation/analysis with python numpy, same data set used for both the traces to show the effect of simply decimating. I do not know if that is what the R&S instrument does but it is one possible explanation or part of it.

[maxwell3e10]

In conventional scope operation only one trigger is allowed in the horizontal time span. So signals triggered on periods 1, 2 and 3 cannot be combined. Otherwise one would get something akin to a messy eye diagram for general (not perfectly periodic) signals. If you want to argue that one can make an eye diagram with faster update rate, I will grant you that, but that is not what we are talking about for basic waveform update testing.  Signal triggered on period 4 can be combined with 1, giving zero dead time.

Pray tell what is a conventional scope ?
If you mean a DSO that statement is wildly untrue.

OK, maybe be you can educate me then, give me a link to a youtube video where more than one trigger is used, other than aformentioned eye diagrams.

[tautech]

In conventional scope operation only one trigger is allowed in the horizontal time span. So signals triggered on periods 1, 2 and 3 cannot be combined. Otherwise one would get something akin to a messy eye diagram for general (not perfectly periodic) signals. If you want to argue that one can make an eye diagram with faster update rate, I will grant you that, but that is not what we are talking about for basic waveform update testing.  Signal triggered on period 4 can be combined with 1, giving zero dead time.

Pray tell what is a conventional scope ?
If you mean a DSO that statement is wildly untrue.

OK, maybe be you can educate me then, give me a link to a youtube video where more than one trigger is used, other than aformentioned eye diagrams.

Easy to test for yourself and plainly see how a trigger will rearm within a screen width.
Just about any protocol packet will retrigger so to not provide stable triggering therefore Holdoff is required for it not to retrigger within a packet.

[maxwell3e10]

Another puzzling thing about FFT is why does the noise spectrum improve at all in HD mode. At 2mV/div scale both scopes are not really limited by ADC resolution, but by the front end noise (in fact HD3 is only using 10-11 bits). HD3 has twice lower front-end noise, so reducing MXO RBW to 2.5 kHz should give similar SNR as 10kHz RBW for HD3.

Now turning on HD limits the bandwidth, it is effectively a boxcar average. That will reduce RMS noise, but it should not change the noise spectral density at 50 MHz. Unless one does subsequent waveform average (which HD is not supposed to do), how can the front end noise spectral density be reduced by filtering high frequency out?

[hpw]

Another puzzling thing about FFT is why does the noise spectrum improve at all in HD mode. At 2mV/div scale both scopes are not really limited by ADC resolution, but by the front end noise (in fact HD3 is only using 10-11 bits). HD3 has twice lower front-end noise, so reducing MXO RBW to 2.5 kHz should give similar SNR as 10kHz RBW for HD3.

Now turning on HD limits the bandwidth, it is effectively a boxcar average. That will reduce RMS noise, but it should not change the noise spectral density at 50 MHz. Unless one does subsequent waveform average (which HD is not supposed to do), how can the front end noise spectral density be reduced by filtering high frequency out?

In addition as from HD304MSO_en.pdf on 2uV:

With the all-new architecture of the HD3 (including a custom 14-bit ADC and low noise front end), you can
analyze all signals in your design with high vertical accuracy. This ensures you make the most accurate
measurements possible and have the most realistic view into the signals inside your device.
You can achieve even greater accuracy (5x better) with up to 16 bits of resolution using the built-in
bandwidth filters. Need to use the full bandwidth to 1 GHz? You will still get extremely high accuracy at
the full bandwidth, with the ability to zoom to 500 uV/div.

[2N3055]

Did you read it ?

1.5% of Full scale. For 500µV/div and 1mV/div a 2mV/div is used. Which means 6%  and 3% respectively.

Ofc I read it. Let's stick with DC gain accuracy, and I will ignore that the Siglent has just 200 MHz BW.

Because of magnification, for HD3 it is +/-6% for the 500uV/div and 3% for the 1mV/div range, and also +/-1.5% at 2mV and 5mV/div.
Then, Siglent specs it at +/-1% for 5mV/div and beyond, while the HD3 spec is still +/-1.5%

So,
better than HD3 at less than 2mV/div, (2x and 4x, no?)
equal at 2m/div, and
1.5x better than the HD3 at 5mV/div and beyond, (not generally "[...] better vertical DC accuracy then their new wonder.. By factor of 3X....")

the thing I disagreed with, was the blanket statement.

Personally, I don't take the "typical" spec seriously. Keysight does not list a typical spec. The datasheet does not mention, is it maximum or what is it, so it is difficult to compare to what Siglent found is a typical spec. I assume it is a maximum (worst case) spec. I don't mind that you compare it to the typical spec but it makes me wonder, what do you think a typical spec value is like for the HD3?

It is 0.5% not 1% so yeah 3X.. with 2mV/div being the same, a sole exception.
And offset accuracy is also better...
Fact that it is 200MHz (really 300MHz) design has nothing to do with DC accuracy. Siglent's 4GHz 7000A also has better DC accuracy, but I wanted to show that Keysight blanket statement didn't hold, even compared to 600€ scope... So if it is not true, it is a lie. Or alternative, people in Keysight became incompetent. Which didn't happen.

And also completely inaccurate in concept, contradicting their own previous whitepaper, that I used as a reference. They are confabulating resolution with accuracy.
You can have 8 bit scope that is accurate to 1ppm, and 16bit one that has 10% accuracy.
And them being Keysight, i refuse to accept they don't know better. So it is deliberate misleading.. They hired used car salesman in marketing.

At least they removed "10x less noise" from new publications... Does their blanket statement they have half the noise of competing designs bother you? Because it was already shown that is also not completely true, i.e. it holds only for certain ranges. And that compared to scopes that cost third the money.

I know you like Keysight and respect them. I do to. That is why you should also be triggered by blanket statements from Keysight that are also very wrong, not at me, pointing the wrongs out. Don't shoot the messenger.

Caveat emptor. There is no scope police, judge or court. There is no commision making sure manufacturers don't get to creative with promises. It is our responsibility, as customers, to keep them in check and slap them on the wrist when they start being "too creative".

Alternative is we lose trust in them. Slowly but surely...  One little lie at a time..
I think Keysight prefers and deserves customer who are open and truthful to them.
To help them stay great and remain great asset to us.
Or we go elsewhere, where they don't insult our intelligence, and try to sell us cat in a bag.

[EEVblog]

At least they removed "10x less noise" from new publications... Does their blanket statement they have half the noise of competing designs bother you? Because it was already shown that is also not completely true, i.e. it holds only for certain ranges. And that compared to scopes that cost third the money.

Nerds arguing over marketing claims is always 

[barrds]

Just noticed two things:
- the HD3 is in 1MOhm (not in 50 Ohm like the MXO4), so it should lower the noise floor if it were in 50 Ohm

- the HD3 is in dBV and the MXO4 in dBm, so the -110 dBm in the R&S would be -123 dBV

[Someone]

In conventional scope operation only one trigger is allowed in the horizontal time span. So signals triggered on periods 1, 2 and 3 cannot be combined. Otherwise one would get something akin to a messy eye diagram for general (not perfectly periodic) signals. If you want to argue that one can make an eye diagram with faster update rate, I will grant you that, but that is not what we are talking about for basic waveform update testing.  Signal triggered on period 4 can be combined with 1, giving zero dead time.

Pray tell what is a conventional scope ?
If you mean a DSO that statement is wildly untrue.

OK, maybe be you can educate me then, give me a link to a youtube video where more than one trigger is used, other than aformentioned eye diagrams.

Why do you think an eye diagram is significantly different to other displays? At your request I showed a diagram with approximately the same signal as was used in the video for the testing (3 cycles of a square wave). Referring back to this diagram:
https://www.eevblog.com/forum/testgear/new-keysight-infiniivision-hd3-oscilloscope/msg5635709/#msg5635709
Why is it more correct to show any particular sequence of edges (possible/available triggers) in the middle of the screen than any others? Why is showing every single possible triggering edge/event in the middle of the screen less desirable or wrong in any way?

[2N3055]

At least they removed "10x less noise" from new publications... Does their blanket statement they have half the noise of competing designs bother you? Because it was already shown that is also not completely true, i.e. it holds only for certain ranges. And that compared to scopes that cost third the money.

Nerds arguing over marketing claims is always 

I know, I know...

[egonotto]

thank you, how did you create the picture?

Numerical simulation/analysis with python numpy, same data set used for both the traces to show the effect of simply decimating. I do not know if that is what the R&S instrument does but it is one possible explanation or part of it.

Hello,

can you please make the program available.

Best regards
egonotto

[Someone]

thank you, how did you create the picture?

Numerical simulation/analysis with python numpy, same data set used for both the traces to show the effect of simply decimating. I do not know if that is what the R&S instrument does but it is one possible explanation or part of it.

can you please make the program available.

Code: [Select]

import numpy as np
import matplotlib.pyplot as plt
from matplotlib.ticker import EngFormatter

formatterF = EngFormatter(places=1, unit="Sa/s", sep="")
formatterS = EngFormatter(places=0, unit="Hz", sep="")
fig, ax = plt.subplots(1,1,figsize=(6, 4),dpi=600)

fs = 3.2e9
nfft = int(320e3)
t = np.arange(nfft)/fs
noise = (3e-3)*np.random.rand(nfft) + (0.6e-4)*np.sin(2*np.pi*50e6*t)
w = np.hanning(nfft)
w_fft = np.fft.rfft(noise * w)
freqs = np.fft.rfftfreq(w.size,d=1.0/fs)
ax.plot(freqs, 20*np.log10(np.abs(w_fft)/nfft),lw=1.0,label=formatterF(fs))

frac = 12
noise_ss = noise[0::frac]
nfft=len(noise_ss)
w = np.hanning(nfft)
w_fft = np.fft.rfft(noise_ss * w)
freqs = np.fft.rfftfreq(w.size,d=frac/(fs))
ax.plot(freqs, 20*np.log10(np.abs(w_fft)/nfft),zorder=-99,lw=1.0,label=formatterF(fs/frac))

ax.set_xscale("log")
ax.set_xlabel('Frequency')
ax.set_ylabel('Amplitude')
ax.set_ylim([-120,-90])
ax.set_xlim([1e6,1e9])
ax.xaxis.set_major_formatter(formatterS)
ax.legend(title='both ΔF 10kHz')
ax.title.set_text("Noise redistribution when decimating")Where ADC noise is uncorrelated, a longer FFT length even when the bin/delta/RBW stays the same shows a lower noise floor. Dave had an excuse in the rush to make the first video, but has "failed" again in this second attempt to compare the noise and ADC differences.

Either same length FFT at full sample rate, or same length at same sample rate with same hi-res channel bandwidth. Those would be fair comparisons.

[EEVblog]

Just noticed two things:
- the HD3 is in 1MOhm (not in 50 Ohm like the MXO4), so it should lower the noise floor if it were in 50 Ohm
- the HD3 is in dBV and the MXO4 in dBm, so the -110 dBm in the R&S would be -123 dBV

Damn, I'll make the excuse that I was drunk.