RME Digi96/8 Pro computer soundcard Measurements part 2
The original RME's analog outputs had very high jitter—3974ps (3.97ns)— leading to an analog noise floor that on average was around 12dB higher than that of the outboard Musical Fidelity processor. However, there was a big improvement when I looked at the jitter in the second sample's analog output (fig.4). The absolute level of the jitter dropped from 3.97 nanoseconds to a superbly low 136.5 picoseconds, with almost no data-related components (red numeric markers) visible in this graph. Though a number of random noise spikes are still apparent—a PC is a very difficult environment for analog circuitry—the new RME's analog noise floor was at least 12dB lower than the old sample's.
Fig.4 RME Digi96/8 Pro, 44.1kHz sampling, high-resolution jitter spectrum of analog output signal (11.025kHz at -6dBFS with LSB toggled at 229Hz). Center frequency of trace, 11.025kHz; frequency range, ±3.5kHz.
Turning to more traditional measurements of the analog output (using PCM WAV files played by CoolEdit 2000 as the test signals), the maximum level was very slightly higher than the old's, at 2.495V vs 2.45V, the pushbutton dropping this to 620.5mV. The source impedance was lower than the original's, at 65 ohms across the audioband, and the output did not invert absolute phase.
Channel separation (not shown) was better than 100dB L-R and 108dB R-L at 1kHz, though this decreased above that frequency due to capacitive coupling between the channels, reaching 71dB and 81dB, respectively, at 20kHz. The frequency response at -12dBFS (fig.5) was identical with both normal and pre-emphasized data, with a very slight rise apparent at the top of the band.
I next performed a wideband spectral analysis of both cards' analog outputs while they decoded 16-bit data representing "digital black." The top traces in fig.6 show the old Digi96/8 Pro, the bottom traces the new sample. The noise floor has dropped by up to 20dB and power-supply components are conspicuous by their absence in the new card's traces, and while the new sample's noise floor rises above the audioband, this is not nearly to the same degree as the old sample's.
I then performed a spectral analysis on the current sample's analog output while it decoded data representing a dithered 1kHz sinewave at -90dBFS (fig.7). The word length was both 16 bits (top pair of traces) and 24 bits (lower traces). The increase in word length gives a 12dB improvement in the level of the noise floor in the treble, but less in the midrange and none in the bass. More significant, there is a large peak at 2kHz apparent with the 16-bit data but not with the 24-bit data.
The production of second-harmonic distortion in this kind of test suggests that the DAC chip has some missing codes, so I fed the card with data representing an undithered 1kHz tone at -90.31dBFS. This consists of just three levels, as can be seen when reproduced by a good D/A converter (see the measurements of the Mark Levinson No.30.6). But when this signal is reproduced by the Digi96/8 Pro's analog output (fig.8), only two levels can be seen. While this kind of behavior was common a decade ago, it is very much more unusual these days, and, to my surprise, was not evident when I tested the card's linearity error with dithered data (fig.9). (The missing code results in energy at twice the signal frequency at this very low level, removing energy from the fundamental and thus should produce a negative amplitude error in the linearity plot.) I will further test this behavior when I test a version of the Digi96/8 Pro with analog inputs, the Digi96/8 PAD. However, I did check the WAV files to satisfy myself that there hadn't been some kind of problem. There hadn't; both dithered and undithered waveforms appeared to be correct.
Fig.5 RME Digi96/8 Pro, D/A frequency response at -12dBFS with 44.1kHz sampling (right channel dashed, 2dB/vertical div.).
Fig.6 RME Digi96/8 Pro, D/A 1/3-octave spectrum of "digital black" with noise and spuriae, 16-bit data, original sample (top) and second sample (bottom). (Right channel dashed.)