This started out as three inexpensive `semi-pro' 8-channel converters. Some of the home recordists among you will probably recognise them from the photos. They were bought a few years ago, when good sounding, reasonably priced converters weren't readily available like they are now.
Unfortunately, like many such units they had a couple of annoying performance bottlenecks. The clock oscillators were on the PCI cards in the computer, but the converters were in the 1U rackmount breakout box. A 2.8m cable separated them, giving plenty of opportunity for jitter due to HF losses and EMI pickup in the cable.
Jitter on these units manifests itself mainly as a `splashy' and poorly defined high frequency range. Listening tests involved putting the sound through the AD-DA stages six times to exaggerate the signal degradation. This made any differences due to modification clearly audible. Spectrum analyser tests were also done, with a single pass through the DAC and ADC on one channel.
Shortening the 2.8m host cables to 1m gave an obvious improvement in sound quality. The cables are standard 25-pin D-sub IEEE 1284 type.
A further improvement was to remove the clock crystals from all three PCI cards and hard wire in a couple of clock modules from Tent Labs (one for 44.1/88.2 and one for 48/96 kHz). These very high quality clock oscillators are designed by Guido Tent of the Grimm Audio design team. They were wired to all three cards in parallel, to synchronise all the cards to the same internal clocks. This removed the need for external synchronisation.
Another weak point was the voltage regulation on the power supplies. The +/- 15V op amp supply regulators were unable to regulate properly because they were being fed too low an input voltage. Initially this was dealt with by increasing the value of the boost capacitors in the half-wave voltage doubler. The regulator ICs were changed for low dropout types and mounted on a heavy heatsink, to give a clean and ripple-free supply to the op amps. The 5 volt regulators supplying the converters and digital ICs also required attention, as they run very hot in their standard positions on the PCB. They were moved to the black heatsink strip in the top of the unit, where they run much cooler. The final power supply modification was to build an external DC PSU feeding the internal regulators with a cleaner input voltage.
Analogue circuitry received some attention too. The drive capability of the output stages was compromised by the cheap design of the -10/+4dBu switches. The switches were removed on outputs and inputs, which were then hard wired for +4dBu. The 1 Kohm resistors in series with the outputs were replaced with 330 ohms (lower than that would probably be inadvisable due to the limited heat dissipation capability of the SMT op amps). Electrolytic capacitors in the signal path were upgraded, and the polyester caps across the ADC modulator inputs were replaced with C0G MLC SMT caps. Protection diodes were added to all analogue outputs to prevent possible damage from +48V phantom power on mixing desk inputs.
The 8-channel units were mounted in a purpose-built rack case with cooling fans (which can be turned off where the noise is a problem) and a custom power supply. A TRS patch panel was added for inputs and outputs, and a passive mixer to combine the hardware monitoring outputs on each unit.
Test Results
The screenshots in the links below were made with jaaa, the JACK and ALSA Audio Analyser. A 3kHz sine wave was generated in jaaa, fed to the DAC and straight back into the ADC via a short balanced patch cable on one channel. Sample rate for the tests was 48kHz. Buffers were set to maximum to minimise jitter cancellation in the DAC and ADC.
Unmodified unit with standard 2.8m host cable. Jitter is clearly visible. The skirts at the base of the 3kHz signal spike start around -90dB.
Unmodified unit with 1m host cable. Jitter is reduced substantially. Intermodulation products at 50Hz intervals due to power supply ripple are no longer masked by the jitter and are clearly visible.
Modified unit with DC PSU and 0.7m host cable. Much better. There are still low level intermodulation products from residual 100Hz PSU ripple, but all below -120dB. Note that the noise visible at around -145dBFS is the noise density when the signal is analysed in 1.46 Hz chunks. Integrated over the audio frequency range this gives an actual SNR of about 103dB (unweighted) for the output and input stages combined.
The same but with a -20dBFS test signal Most of the jitter and IM artifacts are now below the noise floor. This is a clear demonstration of why conservative recording levels are often a good idea.
