Digital Processor Reviews

I let the Allnic D-5000 DHT warm up for an hour before passing any signal, to give the tubes' operating conditions a chance to stabilize. (When the unit is first turned on, the needle on the front-panel meter slowly rises until it rests just to the right of the indicator on the scale.) I examined the processor's electrical performance with my Audio Precision SYS2722 system (see www.ap.com and the January 2008 "As We See It"), mainly using the TosLink optical input, though I repeated some of the tests using a USB connection from my MacBook Pro Laptop running on battery power. Apple's USB prober utility identified the processor as the "D-5000" from "ALLNIC SOUND" with the serial-number string "USB," and confirmed that the USB input operated in the isochronous asynchronous mode. The AudioMIDI utility revealed that the D-5000's USB input would accept 32-bit data with sample rates ranging from 44.1 to 384kHz, including all intermediate rates. The TosLink input operated with data having sample rates up to 96kHz, while the electrical inputs operated up to 192kHz.

The maximum output level was the same from both the balanced and unbalanced outputs, at 3.43V, which is 4.7dB higher than the CD standard's 2V RMS. Upsampling incoming data to DSD dropped the maximum level by 4.8dB, and both sets of outputs inverted signal polarity. The output impedance was the same from both the balanced and unbalanced outputs, at 150 ohms across the audioband. Along with the fact that the output at the XLR jacks does appear to be a true balanced signal, this suggests that the output selection switch on the rear panel connects one end of the output transformer's secondary winding to ground when "Unbalanced" is selected.

The D-5000's time-symmetrical impulse response with 44.1kHz data (fig.1) was typical of a linear-phase digital reconstruction filter, and as you can see, a single positive sample at 0dBFS results in a negative-going impulse at the Allnic's output. This graph was captured with the D-5000 set to DSD upsampling; no upsampling and PCM upsampling had no effect, other than to increase the maximum level, as noted earlier. Fig.2 reveals the behavior in a manner suggested to me by JÅrgen Reis of MBL: The red and magenta traces are wideband spectral analyses of the processor's output as it decoded 44.1kHz-sampled data representing white noise at –4dBFS, while the blue and cyan traces are similar analyses as the processor decoded 44.1kHz data describing a full-scale tone at 19.1kHz. With white noise, the spectra in fig.2 shows the ultrasonic rolloff of the filter, which, while it hasn't reached full attenuation at half the sample rate (vertical green line), eliminates any trace of the aliasing product at 25kHz (44,100 minus 19,100). The right channel's ultrasonic noise floor (magenta trace) is not as low as the left channel's (red). With the 19.1kHz tone, the channels behave identically, with strong harmonics visible at 38.2, 57.3, 76.4, and 96.5kHz.

Fig.2 was taken with upsampling and DSD conversion turned off. With DSD conversion (fig.3), the audioband noise floor has dropped with the 19.1kHz tone, and the distortion harmonics are lower in level. However, the ultrasonic attenuation is reduced, with the result that there is now a strong tone at 25kHz, with other spectral spikes present at the mathematically related frequencies of 13.2 and 7.3kHz. Upsampling 44.1kHz data to 352kHz gives you the worst of both worlds: poor ultrasonic image rejection and high-level distortion harmonics (fig.4).

The traces in fig.5 show a more conventional presentation of the D-5000's frequency response, with data sampled at 44.kHz (green and gray traces), 96kHz (cyan, magenta), and 192kHz (blue, red). (I haven't shown the response with 384kHz data because it was identical to the 192kHz response.) The first thing to note about this graph is that the left channel (green, cyan, blue traces) is about 0.6dB higher in level than the right (gray, magenta, red). More significant, the low frequencies roll off prematurely, especially in the right channel, which reaches –3dB at 25Hz. Ripples can also be seen in the responses in the octave below each Nyquist frequency (half the sample rate), which correlate with dispersion problems in the time domain.

Channel separation (not shown) was good in the L–R direction, at >105dB in the upper midrange and treble, but 25–30dB worse in the other direction. The Allnic's noise floor was clean, with AC-supply components—60Hz at –89dBFS and 120Hz at –93dBFS—respectably low in level for a tubed design (fig.6).

However, with signal present, the D-5000 behaved very strangely. Fig.7 shows the result of my usual test for resolution: I send the processor first 16-bit data, then 24-bit data representing a dithered tone at –90dBFS, and examine how the noise floor drops with the increase in bit depth. The right channel's floor drops by up to 17dB in the treble when the bit depth increases from 16 (magenta trace) to 24 (red), which suggests ultimate resolution close to 19 bits, which is good, though the rise in noise below 1kHz obscures any increase in resolution. However, not only is there no corresponding drop in noise in the left channel (cyan, blue), the audio signal is obscured by a regular series of spurious tones! This is pathological behavior. Fig.7 was taken with S/PDIF data; the picture was the same with USB data, with only the right channel offering any increase in resolution. Upsampling to 352.8kHz PCM didn't change the behavior, and DSD conversion actually reduced the level of the 1kHz tone, strong components appearing at 3, 5, and 7kHz (fig.8).

Another test for resolution is to feed the subject device undithered 16-bit data representing a tone at exactly –90.31dBFS. The output should be a stepped, symmetrical waveform with three clearly defined DC voltage levels (see fig.8 here). The red trace in fig.9 shows the Allnic's right channel on this test. A 1kHz waveform is visible, though the symmetry is disturbed by low-frequency noise and the voltage levels are obscured by higher-frequency noise. But the left channel (blue trace, not plotted to scale) is contaminated by a high-level spurious tone at approximately half the signal frequency.

I had some problems using FFT analysis to examine the D-5000 DHT's distortion, as small changes in the amplitude and timing of the reconstructed signal led to spectral spreading around the primary signal frequency (visible in fig.10) when more than two FFT analyses were averaged, the 1kHz tone acquiring wide spectral "skirts." But you can see that, even into the benign 100k ohm analyzer load impedance, the D-5000 produces a regular series of distortion harmonics, with the highest in level, the second at 2kHz, reaching –42dB (0.8%) in the right channel (red trace) and –53dB (0.2%) in the left (blue). Dropping the level of the signal by 20dB reduced the level of the harmonics by almost another 10dB (fig.11), but now the spurious tones appear in the left channel, with a 500Hz tone visible at –80dBFS (0.01%). At low frequencies, the Allnic couldn't reproduce a full-scale tone without introducing massive distortion (fig.12), the second harmonic of a 50Hz tone reaching more than 10% in the right channel. Perhaps this is due to core saturation in the output transformers (though that would have produced odd-order harmonics).

I use an equal mix of 19 and 20kHz tones to test a product's rejection of intermodulation distortion. You can see that the 1kHz difference product with this signal reaches –50dB (0.3%, fig.13). With DSD conversion (fig.14), the difference product is joined by higher-order products, as well as ultrasonic image tones at 25.1 and 24.1kHz, as expected from fig.3. The same was true with PCM upsampling (fig.15).

Finally, the D-5000's jitter performance with 16-bit J-Test data was the same for all the inputs, including USB. No discrete sidebands can be seen (fig.16), and the noise floor obscures the high odd-order harmonics of the low-frequency, LSB-level squarewave, whose levels are indicated by the green line in this graph.

The Allnic D-5000 DHT has the dubious honor of being the worst-measuring digital component I have encountered, exceeding even the Lector Strumenti Digitube S-192 D/A processor that Art Dudley reviewed last June. I must admit that, unless something broke and affected the left channel's behavior while the review sample was in transit from AD's place to my lab in Brooklyn, I am baffled by this product's measured performance.—John Atkinson