Digital Processor Reviews

I measured the DirectStream MK2 first using my Audio Precision SYS2722 system, then with the magazine's higher-performance APx500. The review sample was fitted with the v2.3.5 FPGA firmware, but after I experienced some anomalous measured results, I installed the v2.3.3 firmware package, which was available as a download on the DirectStream MK2's information webpage. I then installed the v2.3.6 firmware when it was available and repeated the testing.

The DirectStream MK2's optical input locked to sample rates up to 96kHz, the coaxial and AES3 inputs locked to datastreams with all sample rates up to 192kHz. Both Roon and Apple's USB Prober app identified the USB-connected D/A processor as "PS Audio Extended Resolution USB" from "PS Audio," with the serial number string "Version 3.02 11/30/2022." USB Prober revealed that the USB port's "Primary Play Interface" operated in the optimal isochronous asynchronous mode. Apple's AudioMIDI utility also identified the PS Audio as "Primary Play Interface" and indicated that the DirectStream accepted 16-, 24-, and 32-bit integer data via USB sampled at all rates from 44.1kHz to 705.6kHz.

With the volume control set to its maximum, a 1kHz digital signal at 0dBFS resulted in an output level of 3.72V from the balanced outputs, which is 2.3dB higher than that of the original DirectStream. The maximum output level was 3.7V from the single-ended outputs when these were set to "Balanced," which was how the review sample had been set up. (In this mode, the unbalanced output is taken from the hot and cold transformer connections.) Using the Menu to set the single-ended outputs to "RCA Unbalanced" reduced their maximum output by 6dB, as expected, to 1.86V. The volume control operated in accurate 0.5dB steps, and the DirectStream MK2 preserved absolute polarity, ie, was noninverting, from both output types with the Phase switch set to "Norm." The balanced output impedance is specified as <200 ohms; I measured a still-low 400 ohms at 20Hz and 1kHz, dropping to 369 ohms at 20kHz. The unbalanced output impedance was lower, at 175.5 ohms, in the bass and midrange and 166 ohms at the top of the audioband. (The specified unbalanced impedance is <100 ohms.)

The PS Audio's impulse response with 44.1kHz data (fig.1) indicates that the reconstruction filter is a conventional linear-phase type, with time-symmetrical ringing on either side of the single sample at 0dBFS. As DSD data can be converted to analog with a simple low-pass filter, this impulse response is due to the resampling filter earlier in the processing chain. (In the video presentation mentioned in footnote 2 in the review text, Ted Smith says that the FPGA he uses allows him to implement a long digital filter with a large number of coefficients.) With 44.1kHz-sampled white noise (fig.2, red and magenta traces), the DirectStream MK2's response rolled off sharply above 20kHz, reaching full stopband suppression at half the sample rate. The level of the ultrasonic noisefloor starts to rise above 40kHz, which will be due to the PS Audio converting the PCM data to DSD. There are no aliased images of a full-scale tone at 19.1kHz (blue and cyan traces), though some distortion harmonics are visible, with the third lying at –80dB (0.01%).

When I examined the PS Audio's digital frequency response with AES3 data at 44.1, 96, and 192kHz, the response at all three rates dropped off sharply just below half of each sample rate (fig.3). The balanced output is down by just 0.5dB at 20kHz; the unbalanced outputs behaved identically, unless set to "Balanced," when the output was –1.9dB at 20kHz. Channel separation (not shown) ranged from 70dB at the top of the audioband to 80dB at low frequencies.

When I examined the level of the PS Audio's noisefloor, I ran into anomalous results. The upsampling to DSD results in a relatively high level of ultrasonic noise—using the TosLink optical input, I measured 210mV of noise with a center frequency of 380kHz present in the balanced outputs. By contrast, the original DirectStream DAC fitted with the "Windom" FPGA firmware had just 11mV of ultrasonic noise with a center frequency of 283kHz in its balanced outputs (footnote 1). Both Audio Precision analyzers can apply a "brickwall" low-pass filter with a cutoff frequency of 20kHz or 40kHz, but with this filter, I wasn't able to get consistent results with the MK2 until I updated the FPGA firmware.

Fig.4, therefore, shows the DirectStream's low-frequency noisefloor after the change in FPGA firmware as it drove a full-scale 1kHz tone from the balanced outputs, with the volume control set to its maximum of "100." The level of the random noise is about 30dB higher than I was expecting, and if I lowered the volume by 6dB the noisefloor rose by the same 6dB (footnote 2). This graph was taken with the SYS2722; repeating the spectral analysis with the APx500 gave a similar result, even with the brickwall filter operating. The level of the ultrasonic noise wasn't affected when I used the Menu command to lift the output ground or that of the XLR shells.

The presence of this noise affected my measurement of the PS Audio's linearity and resulted in a rising error below –80dBFS (fig.5). However, when I looked at the spectrum of the DirectStream's output while it decoded dithered 24-bit data representing a 1kHz tone at –90dBFS, the tone was reproduced at the correct level (fig.6). The apparent positive linearity error at low levels must therefore be due to the noise.

With undithered 16-bit data representing a tone at exactly –90.31dBFS (fig.7), the three DC voltage levels described by the data were obscured by high-frequency noise.

As suggested by fig.2, the third harmonic was predominant but slightly higher in level at low frequencies, at –70dB or 0.03% (fig.8), than it was in the midrange and treble. The intermodulation products with an equal mix of 19 and 20kHz tones with a peak level of 0dBFS mostly lay around –90dB (0.003%, fig.9), though a series of products spaced at 1kHz intervals can be seen.

The high level of random noise meant that I couldn't meaningfully examine the PS Audio's rejection of word-clock jitter. For example, fig.10 was taken with 16-bit AES3 J-Test data and the Audio Precision brickwall filter set to a 40kHz bandpass. Nevertheless, all the odd-order harmonics of the LSB-level, low-frequency squarewave lay beneath the noisefloor. Peculiarly, the level of random noise in the right channel (red trace) is lower than in the left channel (blue). The levels of the noise became equal if I bypassed the low-pass filter but were about 15dB higher than they were in fig.10.

In most ways—the behavior of the reconstruction filter, the frequency response and channel separation, the sufficiently low levels of distortion—the PS Audio's measured performance is respectable. But in others—the apparent linearity error and the rejection of jitter—the results were dominated by the presence of noise in the DirectStream MK2's output, mostly ultrasonic but in-band, too. This is concerning because amplifiers with a wide ultrasonic bandwidth but poor compensation will suffer more in-band modulation noise, which may impact subjective performance (footnote 3).—John Atkinson

Footnote 2: In other words, the absolute level of the noise was not affected by the volume-control setting.

Footnote 3: The main concern, specifically, is intermodulation of ultrasonic noise or signal-correlated products, into the audible range, which could be the reason for the in-band noise in the DirectStream MK2. Signal-related IM products, if they occur at audible levels, are likely to be more objectionable than noise.—Jim Austin