Review and Measurements of Emotiva XPS-1 Phono Amp

[wynpalmer]

I just read Amir's description of his test methodology, and if he uses 32k bins/48k samples as he describes then this has the same bin width as the 128k bins/192k samples that was used above, so other than a possible error associated with the window "shape" the noise floor results should be comparable.
The test results above use a rectangular window, corrected for overlap. The AP default seems to be the AP equi-ripple design.
For the purposes of the tests above it seems that the two should provide very similar results. Please correct me if this is wrong.
If this is right then the above results should be essentially equivalent to Amir's test results.
Using a scope to "calibrate" the ADC gives 0dBFS at -0.2dBV (c.+/-0.1dB), so the output level is actually at -3dBv for the first plot with the signal at 1kHz.

 

[SIY]

When I've compared noise measurements with AP Equi-ripple with rectangular (the right way to measure noise), there's typically about a 2dB difference, with the rectangular being higher noise and likely more accurate. If memory serves, AP has an app note about this somewhere on their site.

 

[wynpalmer]

When I've compared noise measurements with AP Equi-ripple with rectangular (the right way to measure noise), there's typically about a 2dB difference, with the rectangular being higher noise and likely more accurate. If memory serves, AP has an app note about this somewhere on their site.

Thank you. Very much appreciated. I've elaborated the test results a bit to explain a bit of the methodology employed.
So, this implies that the comparable results for Amir's testing should be c. 2dB more optimistic (and in error) than the one that I posted.
Interesting.

 

[DonH56]

Theoretically you can account for windowing but it gets trickier when noise is involved (need to know the noise spectral density -- its "shape"). The IEEE and other standards choose specific frequencies to avoid windowing.

 

[SIY]

Thank you. Very much appreciated. I've elaborated the test results a bit to explain a bit of the methodology employed.
So, this implies that the comparable results for Amir's testing should be c. 2dB more optimistic (and in error) than the one that I posted.
Interesting.

Here's a handy chart showing the correction factors. According to AP, the actual number for rectangular vs Equi-ripple is a bit higher, 2.63dB.

 

[wynpalmer]

Theoretically you can account for windowing but it gets trickier when noise is involved (need to know the noise spectral density -- its "shape"). The IEEE and other standards choose specific frequencies to avoid windowing.

I converted the noise output to NSD using the REW option and compared to the simulated results. It was remarkably close.
The input device is either an AD797A or an LT1115.
I've largely abandoned the LT1115 as it seems to exhibit substantial NSD part to part variances, which resulted in the 20Hz noise being up to 7dB higher than simulated and 2.4dB higher at 1kHz. I've seen no such variations with the AD797, with approximately a dozen of both parts being tested.
The measurement shown was with AD797s as the input devices.

 

[wynpalmer]

Here's a handy chart showing the correction factors. According to AP, the actual number for rectangular vs Equi-ripple is a bit higher, 2.63dB.

So, the equi-ripple is 2.63dB too optimistic?

 

[SIY]

So, the equi-ripple is 2.63dB too optimistic?

Yes, assuming a random (or nearly random) noise distribution. Of course, if noise is dominated by, say, a 60Hz spike, that number goes out the window, so to speak.

 

[wynpalmer]

Yes, assuming a random (or nearly random) noise distribution. Of course, if noise is dominated by, say, a 60Hz spike, that number goes out the window, so to speak.

Understood. Fortunately in this instance that is not the case, which suggests that the preamp seemingly does quite well.
Does it also imply that Amir's SINAD results for these amps is optimistic by 2.63dB in cases where stochastic processes dominate?
Thank you once again.

 

[SIY]

Does it also imply that Amir's SINAD results for these amps is optimistic by 2.63dB in cases where stochastic processes dominate?

Assuming he used AP-Equiripple, yes. If he used a different window, AP's handy chart will be of use.

 

[wynpalmer]

Perhaps it's time to get the straight scoop from Amir.

 

[DonH56]

I converted the noise output to NSD using the REW option and compared to the simulated results. It was remarkably close.
The input device is either an AD797A or an LT1115.
I've largely abandoned the LT1115 as it seems to exhibit substantial NSD part to part variances, which resulted in the 20Hz noise being up to 7dB higher than simulated and 2.4dB higher at 1kHz. I've seen no such variations with the AD797, with approximately a dozen of both parts being tested.
The measurement shown was with AD797s as the input devices.

Good to know. I am not much of an REW user (use it when I set up, don't tend to touch it after that).

I was thinking of the windowing and NSD itself; there is also the impact of FFT "gain" due to bin widths and resolution (number of points and sampling frequency), potential for aliasing the noise (likely not a problem for the AP since it should have a good anti-aliasing filter), and so forth. I think @SIY has that covered plus he provided that nice chart for window weighting factors. Most DSP texts will have a similar chart and I think it is also in the IEEE Standard (have not looked for a while).

 

[wynpalmer]

Good to know. I am not much of an REW user (use it when I set up, don't tend to touch it after that).

I was thinking of the windowing and NSD itself; there is also the impact of FFT "gain" due to bin widths and resolution (number of points and sampling frequency), potential for aliasing the noise (likely not a problem for the AP since it should have a good anti-aliasing filter), and so forth. I think @SIY has that covered plus he provided that nice chart for window weighting factors. Most DSP texts will have a similar chart and I think it is also in the IEEE Standard (have not looked for a while).

@SIY has provided sage advice. It appears that the bin width is the same for my tests as for Amir's, based on his published test conditions, and the windowing issues can be resolved once we verify what filter Amir uses. If it is the AP default then his S/N numbers are optimistic by 2.63dB whereas my noise levels are spot on.
I've PM'd Amir to ask about his test conditions.

 

[wynpalmer]

Good to know. I am not much of an REW user (use it when I set up, don't tend to touch it after that).

I was thinking of the windowing and NSD itself; there is also the impact of FFT "gain" due to bin widths and resolution (number of points and sampling frequency), potential for aliasing the noise (likely not a problem for the AP since it should have a good anti-aliasing filter), and so forth. I think @SIY has that covered plus he provided that nice chart for window weighting factors. Most DSP texts will have a similar chart and I think it is also in the IEEE Standard (have not looked for a while).

I use the RME ADI-2 PRO FS ADC at 192k samples/sec. It also has an excellent anti-aliasing filter.
It's not an APX555 for sure, but it's a pretty decent test vehicle for short change, especially if you're using it as a part of your audio system as I am, so that it basically comes for free.
REW is an impressive piece of free software. I can't seem to make RIAA measurements work, although theoretically it should be the same as in the APX, so clearly my understanding in setting up the test and comparison files is lacking.

 

[wynpalmer]

I heard from Amir. As this design has no commercial value it's of limited interest to him, but he generously offered to measure it if the membership would like it, and he suggested creating a thread with the results.
Do you think that doing this has some value?
He also outlined his test conditions. He uses 32k bins, 44.1k samples/sec and the AP equi-ripple window.
I use 128k bins, 192k samples/sec and a rectangular window.
As a result I measure 2.63dB more noise than he does due to the window- or, as was suggested before he actually measures 2.63dB too little noise, so his SINADs are indeed 2.63dB too high for the noise limited cases.
In addition, I believe that my slightly smaller bin width (1.2097Hz vs. 1.3458Hz) means that relatively speaking I measure too little noise/bin by 0.46dB so my noise spectra are overall too high by 2.17dB relative to his other measurements. This gives a fixed correction factor for my plots relative to his- that is my noise levels need to be "shifted" downwards by c.2.2dB, making the results look even a bit better.

 

[scott wurcer]

Here's a handy chart showing the correction factors. According to AP, the actual number for rectangular vs Equi-ripple is a bit higher, 2.63dB.

I prefer to export the data to an external tool and use, for instance, 96k unwindowed FFT's for 96k sample rate. The bins are exactly 1Hz wide and signal and noise are both accurate on the same axis with no ambiguities. If you are generating and measuring off the same clock everything is synchronous so no windowing is required.

 

[wynpalmer]

I prefer to export the data to an external tool and use, for instance, 96k unwindowed FFT's for 96k sample rate. The bins are exactly 1Hz wide and signal and noise are both accurate on the same axis with no ambiguities. If you are generating and measuring off the same clock everything is synchronous so no windowing is required.

Hi Scott. That's certainly less ambiguous than the translation that I'm having to go through to get equivalency between say, the Emotiva design measurements and the AD797 based design.
I have to scale for gain (the AD797 preamp has c.6.5dB higher gain) and for the windowing and bin size differences (Amir's reduces to a 2.2dB lowering of the RME measured result) and the level correction, as I did not report it in dBC but in dBFS (-2.9dB)- this produces an overall correction factor of +5.8dB I believe, I then have to correct for the Emotiva output level- -4dB,, so an identical measurement in dBFS to the dBrA would actually be 1.8db "better" than the Emotiva result.
Assuming I got all the corrections etc. correct of course.
So, reading from the graph, the emotiva reads c.-107dB at 1kHz, the AD797 reads c.-108dB, which implies that the AD797 design is c. 2dB better.
I could easily have this wrong as it's just off the top of my head, so any corrections would be appreciated.

 

[amirm]

He also outlined his test conditions. He uses 32k bins, 44.1k samples/sec and the AP equi-ripple window.
I use 128k bins, 192k samples/sec and a rectangular window.
As a result I measure 2.63dB more noise than he does due to the window- or, as was suggested before he actually measures 2.63dB too little noise, so his SINADs are indeed 2.63dB too high for the noise limited cases.

I am not following what you guys are discussing but SNR/Dynamic Range measurements do NOT use FFT so window type doesn't matter. Ditto for SINAD which uses a dedicated meter and not FFT.

If you are looking at the noise floor and trying to compute the actual noise floor, yes you need to use the AP technote to compensate for the window type. I am usually not concerned with 2 or 3 dB differences so don't worry about that in my reviews. If you are designing gear and want to know the actual noise floor, then yes, you need to incorporate that.

 

[wynpalmer]

I am not following what you guys are discussing but SNR/Dynamic Range measurements do NOT use FFT so window type doesn't matter. Ditto for SINAD which uses a dedicated meter and not FFT.

If you are looking at the noise floor and trying to compute the actual noise floor, yes you need to use the AP technote to compensate for the window type. I am usually not concerned with 2 or 3 dB differences so don't worry about that in my reviews. If you are designing gear and want to know the actual noise floor, then yes, you need to incorporate that.

Thanks. I'm attempting to correlate the two measurement systems (we can't all justify having an APX555) as well as getting an idea just how well the preamp does relative to the Emotiva which uses a similar input device, so the noise floor does matter, whereas SINAD is less important and less obviously extracted and actually one of the reasons that I'd like the part to be tested.
Incidentally this was not my idea. It was suggested by someone who is building the DIY amp and he follows the evaluations here.

 

[scott wurcer]

I could easily have this wrong as it's just off the top of my head, so any corrections would be appreciated.

All the correction factors make my head spin. Both pre-amps seem to have a noise floor that conforms to the RIAA curve so I would simply compute the RTI noise at 1kHz in nV/rt-Hz and be done with it.