Some measurements...

[nettings]

Hi Matt,

first of all, congrats for pulling this off - I still have no clue how you are piping this many channels through a single I²S bus!
I ordered two RCA Octos from Amazon UK and I'm now measuring them. For "consumer" sound, they are pretty ok, but for professional use, there appear to be number of limitations.

First, my measuring setup:

• Octo card running on an RPi 3 Jessie with a jack2 sound server and an eight-channel zita network-to-jack bridge. It resamples and is thus not bit-transparent, but measures perfectly.
• RCA outs via an 8way RCA->TS jack loom into an RME Micstasy (input impedance 5.6kOhm) at 10dB gain
• measuring PC with RME HDSPm running jack2
• jaaa as signal generator and analyser

Findings:

• Output level: I'm sending -3dbFS white noise in and measure between -15.6 and -16 dBFS at the Mictasy's MADI output, Micstasy has FS at +21dBu, the analog gain is at 10dB. That means the level is reasonably close to the nominal -10dBu for consumer line interfaces, although not terribly well matched with a maximum deviation of 0.3dB between channels. This result is consistent between both Octos I've seen.
• Output 7 is about 11dB down compared to the other channels, and this is again consistent across both Octos, suggesting a design issue.
• The left and right channels are consistently swapped, i.e. the white RCA is right or odd, and the red is left or even.
• The frequency response is blameless pretty much from DC to Fs/2.
• The noise floor is quite high, -85dB at 47Hz.
• The noise spectrum has clear harmonics at multples of 375 Hz, suggesting a power issue.
• I'm seeing noise peaks whenever the RPi goes to work, such as by issuing an "ls" or "ps aux" command, again suggesting power issues, strong interference from the RPi circuits, or maybe timing issues on the I²S bus introducing spurious LSBs?
• Distortion and crosstalk seem reasonable, I'm currently preparing more detailed measurements, and will look at the performance of the line inputs as time permits.

So, in short, the Octo is quite an achievement in terms of channel count, but ultimately delivers only consumer-grade audio quality, to the point where I would discourage its use for recording and FX, unless using a low-fi approach. The SNR does not give you sufficient headroom for a professional result, and I would be reluctant to use it in a quiet museum environment for a media installation...

The HifiBerry XLR outperforms the Octo RCA by orders of magnitude, alas at a lower channel count. I couldn't get my hands on the balanced version of the Octo for a fairer comparison, but I doubt its noise performance will be much better, given the load-dependent spikes I'm seeing.

Looking at the PCB, some components seem to be shifted a bit - if this is hand-soldered, OMG respect how is it even possible? But if it's factory made, whatever fab is doing it should really ratchet up their quality control (and seeing the design company's domain name misspelled on the PCB doesn't exactly increase my confidence...)

It's certainly an impressive project, and I would be prepared to back an updated version (at a higher cost if necessary) if the bugs mentioned above get shaken out and the noise performance is improved by at least 10dB.
I would also welcome any attempt to trade the pretty much useless 96kHz option for more dynamic range. I didn't look at the digital data on the I²S bus (wouldn't know how), but the noise suggests that the Octo uses a word length of 16-bit or less, and that's not optimal.
Also, I wonder if having the outputs go to DC is such a wise thing to do... then again, I'm not an electronics person, but a sound engineer.

Regards, Jörn

[nettings]

Here is a recording of output channels 5-8 playing digital silence. The noise peaks correspond to me running "ps aux" on the command line via an SSH session. You will notice that the noise level of channel 7 is also attenuated, not just the signal. Channels 1-4 look identical to 5,6 and 8.
The waveform display is logarithmic to make the low signal visible (we're peaking at -55dB), and the time range of the entire window is around 3.7 seconds.
The same noise peaks occur randomly as the RPi is performing various system tasks.

Screenshot_2017-09-23_15-49-14.png (154.96 KiB) Viewed 6116 times

[nettings]

Having read about an Octo user who is using his in an analog synth context, I do understand the benefits of going down to DC... it's just that in everyday music use, DC coupling tends to cause quite a few problems if you're unaware of it, but hey, I take back my remark now that I've heard of a usecase.

[nettings]

So I'm trying to quantify the noise floor by doing the following:
I play an 1kHz sine wave through all eight Octo outs at -1.0dBFS into the eight Micstasy inputs. When I set them to +18dB gain, I get just about -1.0dbFS of peak value out of the AD converters. With this gain, I'm starting jnoisemeter with DC and 20kHz filter, otherwise flat.
The resulting noise figures are around -73.5dBFS for all channels except 7, which is down at -85.
That is certainly ok for music enjoyment (given that high-end studio tape machines struggle to reach this noise level), but not quite in the comfort zone of contemporary digital equipment.
Of course the figues look a lot better when using the A curve (-85dB), but that is just about useless for noise, even though many manufacturers like to use it for exactly that reason

[nettings]

This is the noise spectrum, with the Micstasy gain at +18dB, so that a full-scale output produces a full-scale input to within maybe +/-2dB of error.
I wonder what causes this... Could it be fixed with a more stable power supply? I'm using a five-way USB charger capable of delivering 2.5A on each output, with only one output used, so my PSU should have ample headroom.

Noise Floor.png (49.03 KiB) Viewed 6110 times

[nettings]

For comparison, here is the HifiBerry DAC+ Pro XLR, measured through a different preamp (Direct Out Andiamo), but with comparable gain structure:

HifiBerry-DAC+ left channel, silence + zita-resampler.png (111.33 KiB) Viewed 6109 times

Still a different league, but if the Octo could lose its weird bumps below 4k, it would be closer...

[nettings]

Total harmonic distortion of an 1kHz sinewave at -3.8 dBFS:

Distortion spectrum.png (49.93 KiB) Viewed 6109 times

A rough sum of the harmonics yields a THD value of 0.00018. Not impressive, but certainly not enough to be worrisome given that the signal will eventually hit a speaker Mostly odd harmonics, though, so don't expect a tube-ish effect.

[flatmax]

Thanks for this information - and introducing me to jaaa, nice tool.

I though it would be a good idea to look at the minimum possible noise. For that reason, I have done a pin to pin connection, where I remove all analogue buffers and simply connect header pins together. The test looks like so for channel 1 :

Octo pin to pin test for channel 1

pin2pin.jpg (219.32 KiB) Viewed 6044 times

I used jaaa with the ALSA driver :

Code: Select all

jaaa -A

What I find is that the noise floor is sitting on -130 dB with about 10 dB up from the HifiBerry, with a small bump at around 8 kHZ, probably due to the cable. It looks like so :

Jaaa pin to pin for the octoc, -130 dB noise floor.

jaaa.pin2pin.png (36.92 KiB) Viewed 6044 times

I also see that there is almost a -100 dB difference between the 1 kHz tone and its 2nd harmonic :

Jaaa pin to pin for the octoc, 2nd harmonic is -100 dB down.

jaaa.pin2pin.tone.png (47.57 KiB) Viewed 6044 times

This - given a direct connection between devices, such as an analogue synth directly from the Octo header is some great low noise control capability.

Given that you can create your own breakouts, this is probably the lower limit you will reach.

Matt

[nettings]

Hi Matt, thanks for your detailed reply, and sorry for the long delay in coming back to this. I'm looking forward to revisiting the Octo now that other projects have finished, and I will see if I can reproduce your measurements. It would be very good news indeed if the signal degradation I'm seeing is just a consequence of the low-budget RCA breakout!