Grace Design m902 Reference D/A headphone amplifier Measurements
Fed digital data with the volume control at its maximum setting of "99.5," the Grace m902 Reference had a maximum output level of 7.52V at 1kHz from its headphone output and 7.34V from its RCA outputs, both more than 11dB above the CD standard of 2V RMS. The polarity was correct from both sets of outputs; ie, the unit was non-inverting, and the source impedances were very low: 47.5 ohms from the RCAs, 1.5 ohms from the headphone jack (both figures include the series resistance of 6' of interconnect).
The volume control operated in accurate 0.5dB steps, with superb channel matching. As set up, the review sample offered different maximum gains for analog input signals at the unbalanced and balanced connections: 12dB and 0dB, respectively. The analog input impedance at 1kHz was to specification at 52k ohms unbalanced and 104k ohms balanced; neither input inverted polarity.
Measured with an analog input signal, the m902's frequency response from all its outputs extended well above the audioband, not reaching –3dB until above 200kHz (fig.1, top pair of traces). Fed 96kHz-sampled digital data, the output very slowly rolled off above 10kHz, reaching –0.75dB at 40kHz, implying that the digital reconstruction filter has perhaps been optimized for good time-domain performance (see Keith Howard's article on this subject in January, pp.57–65), before rolling off sharply above 42kHz. The response with CD data overlays this response exactly. Midband channel separation via the line outputs (with the crossfeed circuit off) was superb, at better than 110dB in both directions, but decreased to a still good 90dB from the headphone jack, perhaps due to the shared ground between channels.
Fig.1 Grace m902, frequency response into 100k ohms via (from top to bottom at 50kHz): analog input at 1V, digital input at –12dBFS (right channel dashed, 1dB/vertical div.).
Fig.2 shows the spectra of the Grace's headphone output signal, produced by sweeping a 1/3-octave bandpass filter from 20kHz to 20Hz while the unit was fed dithered 16-bit data representing a 1kHz tone at –90dBFS (top traces), 24-bit data representing the same signal (middle), and dithered 24-bit data representing a 1kHz tone at –120dBFS (bottom). The increase in bit depth drops the noise floor in the upper midrange and treble by almost 15dB, suggesting ultimate resolution approaching 19 bits, which is excellent. The tone at –120dBFS can clearly be differentiated from the noise floor, and there is no sign of harmonic- or power-supply-related spuriae. Increasing the measurement bandwidth to 200kHz and feeding the m902 16-bit digital data representing a –1LSB DC offset gave a smooth noise floor below 20kHz (not shown), but a rising level of noise above the audioband due to the noiseshaping used to get high resolution from the DAC chip. However, as this noise still lay at just –70dBFS at 200kHz, it will not be a problem for downstream components.
Fig.2 Grace m902, 1/3-octave spectrum with noise and spuriae, of (from top to bottom): dithered 1kHz tone at –90dBFS, 16-bit and 24-bit data, dithered 1kHz tone at –120dBFS, 24-bit data (right channel dashed).
With the Grace m902's low levels of audioband noise, it comes as no surprise to find that with dithered 16-bit data, any linearity error is merely the contribution of the dither noise (fig.3, top trace below –100dBFS). Increasing the word length to 24 bits results in negligible linearity error to below –120dBFS (fig.3, lower trace). The m902's reproduction of an undithered tone at –90.31dBS is thus essentially perfect, both for 16-bit data (fig.4) and 24-bit data (fig.5).
Fig.3 Grace m902, left-channel departure from linearity (from top to bottom below –100dBFS), 16-bit and 24-bit data (2dB/vertical div.).
Fig.4 Grace m902, waveform of undithered 1kHz sinewave at –90.31dBFS, 16-bit data.
Fig.5 Grace m902, waveform of undithered 1kHz sinewave at –90.31dBFS, 24-bit data.
Measured at its headphone output, the Grace offered very low levels of harmonic distortion with a full-scale, 24-bit sinewave (fig.6), though the right channel at 0.0016% THD was higher than the left at 0.0007% (both figures the true sum of the harmonics). This was due to the subjectively innocuous second harmonic being higher in the right channel, though at –96.5dB, this is still irrelevant to the listening experience, I would have thought. The noise floor in this graph is actually that of the analyzer rather than the Grace, but the fact that it rises slightly at low frequencies is not characteristic of the analyzer.
Fig.6 Grace m902, spectrum of 1kHz sinewave at 0dBFS into 8k ohms, DC–10kHz, 24-bit data (linear frequency scale).
I repeated the measurement with a 24-bit, 1kHz tone at –90dBFS. The resultant spectrum is shown in fig.7. The distortion harmonics are buried in the noise, but the Grace does appear to have a rising noise floor below 3kHz or so. This can also be seen in the spectrum resulting from the m902 decoding data representing an equal mix of 19kHz and 20kHz tones, the waveform peaking at 0dBFS (fig.8). The 1kHz difference component into this load lies at a very low –102dB (0.0008%); even into a low 600-ohm load, it rises only slightly, to –99dB, which is superb performance.
Fig.7 Grace m902, spectrum of 1kHz sinewave at –90dBFS into 8k ohms, DC–10kHz, 24-bit data (linear frequency scale).
Fig.8 Grace m902, HF intermodulation spectrum, 19+20kHz at 0dBFS peak into 1k ohm, DC–24kHz (linear frequency scale).
I assess a component's rejection of word-clock jitter by driving it with data representing a high-level tone at exactly one quarter the sample rate, combined with a low-frequency squarewave at the LSB level, this again at an exact integer fraction of the sample rate. Any artifacts other than these components that show up in the product's analog output will thus be due not to quantization but to problems in the D/A circuitry, including any susceptibility to jitter. I use an analyzer marketed by Miller Audio Research to examine the product's noise floor for such artifacts. This looks for and identifies symmetrical sidebands around the spectral line that represents the high-level tone.
The result of this test for the Grace m902, fed 44.1kHz-sampled, 24-bit data from my PC via a TosLink connection from an RME soundcard, is shown in fig.9. The jitter level was 405 picoseconds peak–peak, which is low, but not as low as the best components I have measured. The three highest-level sideband pairs lie at ±15.6Hz (purple "1" markers), ±23Hz (purple "2"), and ±813.5Hz (purple "6").
Fig.9 Grace m902, high-resolution jitter spectrum of analog output signal (11.025kHz at –6dBFS sampled at 44.1kHz with LSB toggled at 229Hz), 24-bit data from PC via TosLink connection. Center frequency of trace, 11.025kHz; frequency range, ±3.5kHz. (Grayed-out trace is with 16-bit data.)
The grayed-out trace in fig.9 shows what happened when I fed the m902 with 16-bit data, again via a TosLink connection. The jitter level has increased slightly, to 420ps, mainly due to the appearance of data-related sidebands (circled in red). These are slightly higher than the residual level of the test signal, but otherwise the spectrum has not changed. Changing the data source to CD played on a PS Lambda transport connected to the Grace with an electrical S/PDIF link dropped the jitter level to an excellent 295ps, mainly due to a reduction in level of the three pairs of sidebands noted earlier. This suggests that the m902 is somewhat susceptible to the datalink used. I also wondered if the slight spectral spreading of the 11.025kHz tone apparent in this graph correlated with the low-frequency rise in the noise floor seen in the distortion graphs.
Wes Phillips reported that the m902 sounded "murky" when he connected it to his Macintosh G5 via USB. I therefore used the Miller analyzer to measure the jitter driving the Grace with USB data from my lab's PC. The result was the highest jitter level I have encountered, at 28.6 nanoseconds—almost 100 times greater than via the well-behaved S/PDIF connection. However, when fed a continuous, constant-level sinewave via USB, the m902's output level kept changing slightly. It appeared that this was due to a high level of hum; no matter how I arranged the m902's grounding or lack thereof, I couldn't eliminate what must have been a ground loop between the computer and the Grace.
The only way I could get hum-free audio from the Grace via its USB input was from my battery-powered PowerBook laptop. Fig.10 shows the jitter spectrum taken under these circumstances. The jitter level was lower than before, though still high, at 1.3 nanoseconds, and the spectrum is dominated by strong sidebands at ±23Hz (purple "1"), ±34.7Hz (purple "2"), ±230Hz (red "4"), and ±690Hz (red "9"). Note that there is also some spectral spreading of the central tone, due to the presence of low-frequency random-noise jitter. The correlation between jitter levels and spectra is unclear. However, I am not surprised that WP found the Grace's USB input to sound "murky." I must point out, however, that I have generally found much worse measured performance from USB-connected D/A converters.
Fig.10 Grace m902, high-resolution jitter spectrum of analog output signal (11.025kHz at –6dBFS sampled at 44.1kHz with LSB toggled at 229Hz), 16-bit data from Apple PowerBook via USB connection. Center frequency of trace, 11.025kHz; frequency range, ±3.5kHz.
The idiosyncratic nature of our sample's USB input aside, the Grace Design m902 Reference offers excellent measured performance.—John Atkinson