Schiit Audio Bifrost D/A processor Measurements
Sidebar 3: Measurements
I examined the Schiit Bifrost's electrical performance with Stereophile's loan sample of the top-of-the-line Audio Precision SYS2722 system (see www.ap.com and the January 2008 "As We See It"). To test the USB input, I played files with Bias Peak Pro running on my MacBook Pro, using the AudioMIDI utility to make sure that the data output from the USB port had the correct sample rate and bit depth. I performed a complete set of measurements on both samples and have noted in the text where these differed.
The coaxial input of both samples locked to datastreams with sample rates of up to 192kHz (including 88.2 and 176.4kHz)as did, to my surprise, the TosLink input of the first sample (serial no. A041462). (A TosLink connection is not formally specified to work above 96kHz, due to its limited bandwidth.) The second sample's TosLink input (SN A042654) would operate only up to 96kHz. The USB input of the first sample operated at all sample rates up to 192kHz but not 176.4kHz; the second sample worked correctly with this sample rate. Apple's USB Prober utility indicated that the USB input of both samples operated in the optimal asynchronous isochronous mode. US Prober identified the Bifrost as the "Speaker-Schiit USB Interface," with the manufacturer string given as "CMEDIA" for the first sample, "Schiit" for the second. One peculiarity was that AudioMIDI indicated that the setting of both Bifrosts' USB connection defaulted to 16-bit integer when each was first connected to the Mac. The Bifrost's owner should make sure to use a programeg, Amarra, Audirvana, or Pure Musicthat sets the bit depth correctly for each file played when using the USB connection.
Both samples offered a maximum output level of 2.17V at 1kHz, 0.7dB higher than the CD standard's 2V, and both preserved absolute polarity (ie, were non-inverting). The two samples differed when it came to output impedance. While the second sample's impedance, with the Uber Analog board, was a low 77 ohms at all frequencies, the first sample's impedance varied from 74.5 ohms at 20kHz to a still-low 78.5 ohms at 1kHz, then to a very high 2930 ohms at 20Hz. This increase at low frequencies is presumably due to the use of a coupling capacitor in series with the output. The maintenance of the second sample's low output impedance in the bass might well correlate with JI's finding it to sound "tighter" overall.
My now-standard examination of the behavior of a DAC's reconstruction filter is to play first a 19.1kHz tone at 0dBFS, then high-level white noise, both sampled at 44.1kHz, while performing a wideband FFT analysis on the DAC's output, a test suggested to me by MBL's Jürgen Reis. The result for the first sample of the Bifrost is shown in fig.1: the white-noise spectrum (blue and magenta traces) rolls off sharply above 22kHz, with a slight rising trend visible above 60kHz. The 19.1kHz tone (cyan and red) has second and third harmonics visible at 94dB (0.002%) and 79dB (0.011%), while the first aliasing product, at 25kHz, is suppressed by 87dB (0.005%). This suggests the use of a conventional digital filter. However, there are many other products visible, though it's fair to note that these all lie at very low levels. The second sample (fig.2) behaved similarly in this respect, but now with the second harmonic the highest in level, at 96dB (0.0015%), and the third harmonic buried in the noise floor.
Both samples offered the same frequency responses at 44.1, 96, and 192kHz (fig.3, green and gray, cyan and magenta, and blue and red traces, respectively). A smooth overall rolloff above the audioband is broken in each case by a sharp drop just below half the sample rate. The channel separation of both samples was superb, at >120dB below 1kHz, though the second sample's separation suffered from more of a decrease in the top octaves, reaching 97dB RL and 104dB LR at 20kHz (not shown). Analysis of the Bifrost's noise floor while it decoded a full-scale tone at 1kHz (fig.4) uncovered some spuriae at 60Hz and its harmonics, though these are all very low in level.
Whether fed S/PDIF data or USB data, the Bifrost offered almost 19 bits of resolution, which is superb performance at the price. The cyan and magenta traces in fig.5, for example, show an FFT analysis of the processor's output while it decoded 16-bit data representing a dithered tone at 90dBFS. Increasing the bit depth to 24 gave the blue and red traces in this graph; the noise floor has dropped by an impressive 15dB. The graph was taken with S/PDIF data; unusually, via USB the 24-bit tone had a small amount of odd-order distortion visible (fig.6). Even so, the Bifrost's reproduction of an undithered 16-bit tone at exactly 90.31dBFS (fig.7) was essentially perfect, with a symmetrical waveform and the three DC voltage levels clearly defined. With undithered 24-bit data, the result was an excellent sinewave (fig.8).
As suggested by fig.1, the two samples differed in the amount and type of distortion they produced. Fig.9 shows a spectral analysis of the first Bifrost's output while it played a full-scale 1kHz tone into 100k ohms. The primary harmonics visible are the second and third, each at 96dB (0.0015%), and some higher harmonics are visible, albeit at a much lower level. While the second harmonic lies at the same level in the spectrum of sample 2's output (fig.10), the third harmonic has dropped to 110dB (0.0003%). The second sample also produced lower levels of higher-order intermodulation distortion, even into the punishing 600 ohm load (fig.11, first sample; fig.12, second sample).
Both samples offered anomalous behavior when decoding the Miller and Dunn J-Test data via S/PDIF. Fig.13, for example, shows the spectrum of the first sample's output while it decoded 16-bit J-Test data. Not only is the primary pair of sidebands at ±229Hz accentuated, and some off-order harmonics of the low-frequency, LSB-level squarewave suppressed, but there is a peculiar rise in the noise floor to either side of the spectral spike that represents the 11.025kHz tone. By contrast, fed the same 16-bit data via USB (fig.14), the noise floor now lies at the correct level and the odd-order harmonics lie at almost the correct level, though the sidebands at ±229Hz are suppressed. The same behavior is evident with the second sample, with fig.15 the 16-bit S/PDIF spectrum and fig.16 the 16-bit USB spectrum. However, with the USB data, the odd-order harmonics vary too much in level, the sidebands at ±229Hz are accentuated, and another pair of sidebands makes an appearance at ±120Hz. But in all four jitter graphs the 11.025kHz tone is sharply defined, with no trace of the spectral spreading often seen around the base of the spike.
Overall, both samples of the Bifrost measured well, especially given the affordable price. And the second sample's Uber Analog board offers lower distortion and a usefully lower output impedance at low frequencies, justifying the cost of the upgrade. However, given that Jon Iverson used both samples of the Bifrost's S/PDIF input with his Sooloos server, I can't help but wonder if the slight "fuzziness" he noted was related to this input's idiosyncratic behavior when tested for jitter rejection.John Atkinson