Dragonfly Black teardown

[mansr]

The Audioquest Dragonfly Black is a small USB-stick DAC intended for portable use with laptops and phones. It also functions as an MQA "renderer," an ability that makes it, along with the Red variant, unique in this form factor, whatever that's worth. The supported audio formats are 24-bit stereo at 44.1, 48, 88.2, or 96 kHz sample rate. A volume control is provided over the USB interface. There are no buttons on the device itself. A single LED indicates through its colour the current sample rate.

 

The aluminium case gives a solid feel. Indeed, non-destructively opening it proves impossible. Once the end cap is forcibly removed, however, the circuit board inside slides out easily:

 

Besides a smattering of passive components, we find the interesting parts:

• ES9010K2M DAC
• TPA6130A2 headphone amplifier with volume control
• PIC32MX270F256B microcontroller
• 24 MHz crystal

The DAC chip is at the low end of the ESS catalogue. The headphone amp is, incidentally, the same as in the iFi Nano.

 

Basic characteristics

We start by looking at the impulse response:

As promised, we have a minimum phase filter with an impulse response extending a little over a millisecond at 48 kHz. Other sample rates have the same response scaled in time accordingly.

 

A full-scale square wave looks like this:

There is considerable clipping of the peaks of the filter. Reducing the amplitude by 3 dB avoids the clipping:

Real music should hopefully never run into this situation. If it did, the output would be distorted.

 

Clocking

The Dragonfly has a single 24 MHz clock driving both the ES9010 and the PIC32MX. As this is not a multiple of the bit rate for any of the sample rates, the bit clock for the I2S interface must be synthesised in the microcontroller. This is the result:

That is not pretty. The period time appears to be alternating between two values.

 

Zooming in for a closer look, we see that the difference is about 10.4 ns, which is close to the period of a 96 MHz clock. Most likely the 24 MHz clock input is multiplied by 4 and fractional divider used to approximate the target rate. The PIC32MX does indeed support such a configuration.

 

For a 48 kHz sample rate, the supported ratios in the divider allow an exact result aside from the period to period rounding errors. At 44.1 kHz, we are less lucky. The closest we can get with the available precision is 44102 Hz. It's not exact, but it's pretty close. However, looking at the LR clock, this is not what we find:

That's a much larger error than expected. It does, however, correspond exactly with a divisor fractional part of zero. A quick look at the bit clock confirms this:

There are no double edges, like those seen with 48 kHz sample rate, here.

 

Over an audio sample interval, 64 bit clock periods, the bit clock jitter looks like this (there is no appreciable difference between 44.1 kHz and 48 kHz base rates):

Although the single-period errors cancel out over this interval, there is still considerable jitter, probably from the PLL clock multiplier. As we can see from the histogram, the bulk of the clock edges fall within a 300 ps window with outliers as far as 1 ns to either side. Note that this is near the resolution limit of the oscilloscope.

 

DAC chip

At the heart of the Dragonfly is the ES9010 DAC chip. The ESS website has a helpful, though no doubt very simplified, block diagram.

 

An unusual feature of the ESS DAC chips is the ASRC, allowing the I2S input rate to be independent of the master clock, in this case the 24 MHz clock shared with the microcontroller. This also serves to reduce the effects of jitter in the input signal, something that should be useful given what we saw above.

 

The oversampling filter has a few built-in options and can also be programmed with custom coefficients. For regular PCM playback, the Dragonfly uses the built-in minimum phase filter. The volume control feature is not used.

 

Amplifier

The output of the DAC chip feeds the headphone amplifier. This part also provides a 64-step volume control ranging from -59.5 dB to +4.0 dB. When the Dragonfly is plugged in, it initially sets the volume to level 43, or -4.5 dB. A charge pump operating at approximately 400 kHz provides the negative voltage rail allowing the output to swing in both directions.

 

In the next post we will examine the Dragonfly audio performance using some test signals. Stay tuned.

[mansr]

1 hour ago, SJK said:

I wonder if in some cases the Dragonfly product line has become obsolete?  I have the latest Red and Black - neither would play on my work laptop without frequent crashes or nothing but static.  Mind you, we're still using Windows 7 (!!!!!), so that may be the issue. 

Sounds like a driver issue. I have no problems at all with the standard usb-audio driver in Linux.

 

1 hour ago, SJK said:

But, with my work laptop and the latest sound drivers, it's supposedly capable of 24/192 and sounds great either with headphones or my cheapy USB powered speakers.  For the few times I was able to compare when either of the Dragonfly's were working I couldn't say they had a better sound.

It supports up to 24/96. As for sound quality, it's not terrible, but neither is it spectacular. There are probably cheaper devices that perform just as well.

[mansr]

The wait is over. It is time to measure the analogue performance of the Dragonfly.

 

Test setup

The recording interface used for these tests is a Tascam UH-7000. Both it and the Dragonfly are powered from a linear lab supply with floating outputs. They USB connections are to separate laptops running on battery power. There is thus no possibility of ground loops, and and noisy USB power is not a concern. Disconnecting the USB ground is unfortunately not possible with the Dragonfly.

 

Silence

Recording the DAC while it plays back digital silence reveals the noise level as well as any spurious tones such as USB packet noise. The first graph shows the spectrum of the Dragonfly playing 48 kHz silence at the maximum volume setting (+4 dB) as well as the power-on default (-4.5 dB). Also shown is the ADC noise with the input shorted.

 

The noise levels at both volume settings, while different, are above the self-noise of the ADC, though not by much. Above 60 kHz, the ADC modulator noise starts dominating, making it impossible to discern anything about the signal from the Dragonfly. At lower frequencies, we see a smooth noise floor, interrupted only by a few small spurs near 20 kHz. To examine the lowest, yet most important, frequencies we need to zoom in a bit:

The spectrum is clean with only a hint of power line hum (expected since nothing is directly connected to the AC mains). As the Dragonfly is a full-speed USB device, packet noise, if present, would have been at 1 kHz.

 

Volume control

Since the default volume setting is reasonably loud with typical music, this will be used for the remainder of the measurements. Comparing different levels could be interesting, but the combinations are just too many.

 

White noise

Primarily, a white noise test show the frequency response of the DAC. If the interpolation filter is prone to overloading, this is also revealed.

 

In this spectrum plot of 48 kHz white noise, we see from the poor attenuation above the cut-off frequency that the filter has overloaded. With the input at -6 dBFS, the filter is working properly.

 

Here is the transition band in detail. The roll-off starts slowly soon after 22 kHz, reaching around 8 dB at the Nyquist frequency, 24 kHz. Final attenuation is only achieved at 26 kHz.

 

The graphs for 96 kHz sample rate look exactly the same, only stretched in frequency. This is no surprise since the same filter is used for all input rates.

 

1 kHz

Pure tones test for harmonic distortion. 1 kHz at full scale is a good place to start.

 

That's a fairly large amount of harmonic distortion, significantly higher than for example the iFi nano iDSD, which to be fair is a much larger and more expensive device.

 

The same graph expanded to the full recorded frequency range shows odd-order harmonics extending as far as the ADC is able to capture.

 

To see how the softest sounds are handled, we test a tone at -120 dBFS.

The 1 kHz tone is clearly visible, although it is lower in amplitude than the power line hum. This dynamic range ought to be sufficient for most music.

 

Finally, we compare a 1 kHz rendered at 44.1 kHz and 48 kHz sample rates.

The frequency error in the 44.1 kHz clock is readily apparent here. Specifically, it is 400 ppm too fast, in agreement with the earlier scope measurements. For reference, S/PDIF specifies two level of timing accuracy, high at 50 ppm and normal at 1000 ppm. Surprisingly, no golden-eared reviewers have picked up on this discrepancy.

 

10 kHz

Distortion often gets worse as the frequency goes up, so we check the behaviour at 10 kHz as well.

The peak levels of the harmonics are similar to those seen at 1 kHz. However, they drop off much more slowly with even the 9th harmonic only a little lower than the second and third. That said, these levels are not likely to cause any trouble.

 

Intermodulation distortion

To test for intermodulation distortion, we use pairs of tones spaced 1 kHz apart.

 

At 10 kHz, the IMD test tones produce a number of different distortion products. We see a 1 kHz difference tone, multiple side tones around the main pair, various sums with their own side tones, and some harmonics of other components.

 

With the test tones at 20 kHz, a similar pattern of distortion products results, and they are now stronger than at 10 kHz.

 

Doubling the frequency once more to 40 kHz, the pattern persists. The 1 kHz difference tone is now 30 dB higher than with the 9+10 kHz test tones. At this level it is readily audible while the true signal is anything but, being far outside the range of human hearing. This is the reason Monty cautions that high sample rates may do more harm than good.

 

Jitter

Many a problem, whether real or imagined, is attributed to jitter. When the DAC clock strays from a constant rate, the result is a smearing of frequencies in the reproduced audio. In a spectrum plot of a single tone, this typically shows up as a skirting around what should be a perfect vertical line. A good test signal is a square wave at ¼ of the sample rate as this is actually a perfect representation of a sine wave without any quantisation errors.

 

This is the result of such a test on the Dragonfly. The spreading starts some 100 dB below the peak, which is is perfectly acceptable. There isn't much else to say.

 

That's all for now, but we're not done yet. Watch this space.

[mansr]

6 hours ago, PeterSt said:

Was it impossible to calibrate the dB scale ? Add ~ 45dB myself to everything doesn't make more readable. 

Or did I miss something else ?

Not impossible, but it would have been more work. Where would you have me put the 0 dB level?

[mansr]

53 minutes ago, PeterSt said:

Assumed that you don't show Volts (like dBV) but digital level ... at dBFS - 0. At least this is normally done so and thus everybody can "read" it. This proper reading then includes the (also normal) -3dBFS for a test signal. Thus, Y-axis has a -0dBFS marker and a normal test signal will play at -3dBFS. And if you don't use -3dBFS for test signal but -60dBFS (also quite normal) then everybody can see what you're testing.

Etc.

The appearance of the graphs also depends on the FFT size and windowing function.

[mansr]

Just now, Miska said:

It is actually good to use the Miller-Dunn Jtest signal for testing jitter, because it has LSB modulation.

I thought that was meant to tease out problems in S/PDIF clock recovery.

 

Just now, Miska said:

It tells if there are capacitive leaks between I2S lines (data line leaking to clock and thus varying the switch-over point in the clock waveform).

Given the amount of jitter already present on the clock here I doubt it matters much. The ESS DAC also doesn't use this clock directly. Anyhow, I'll run the test.

[mansr]

13 minutes ago, Miska said:

What Peter said about calibrating the 0 dB level is true. First play 0 dBFS 1 kHz tone and then calibrate the 0 dBr point of the graphs to that voltage... This way level of the distortion components are easy to read...

Shifting the graphs vertically won't change the relative levels.

[mansr]

Here's the J-test at 48 kHz:

[mansr]

4 hours ago, Miska said:

How does it look at 44.1k?

The same. I see no reason why it would be different.

 

4 hours ago, Miska said:

For some reason the noise level is quite high, normalized it would be about -100 dB so it hides most of the detail.

Maybe your recording equipment is more expensive than mine.

[mansr]

16 hours ago, Miska said:

P.S. Nowadays I use different scale for the Jtest24 (44.1k), now I use 8k - 14k frequency range and -170 - -90 dBr level range. This seems to make things quite nicely visible.

What FFT parameters do you use?

[mansr]

2 hours ago, Miska said:

For Jtest, sampling rate 48 kHz, FFT size 256k, 8 averages.

 

All other measurements are always at 192k sampling rate.

Why the different sample rates? Same FFT size?

[mansr]

I turned up the gain on the ADC a bit and redid the recordings. At this level, some of its internal noise starts showing up, but it also reveals a little more low-level detail from the Dragonfly, which is by far the noisier of the devices.  The plots below are all normalised so a 1 kHz tone at 0 dBFS has its peak at 0 on the graph.

 

1 kHz

 

 

 

10 kHz

 

20 kHz

 

9 + 10 kHz

 

19 + 20 kHz

 

J-test

[mansr]

49 minutes ago, Ralf11 said:

Is there a bottom line summary yet?

Feel free to draw your own conclusions.

[mansr]

22 minutes ago, GUTB said:

Question, why aren't you using an Audio Precision analyzer?

Because I don't have one.

[mansr]

3 hours ago, semente said:

Did you publish measurements for the iFi?

No, but I could.

[mansr]

And now for something a bit different, MQA. After some delays, a firmware update to the Dragonfly Black and Red gave them the ability to function as MQA "renderer." As previously discovered, MQA rendering consists of two parts: upsampling and shaped dither. Since the PIC32 microcontroller in the Dragonfly has very limited computing resources, it seemed unlikely that it would be able to handle the upsampling. As the ESS DAC chip has programmable filters, utilising these would be a reasonable alternative. Playing some MQA files, the sample rate sent to the DAC chip is indeed unchanged at 96 kHz.

 

The ESS chip has a two-stage programmable FIR filter. The first stage is a fully programmable 128-tap filter. The second stage is a 28-tap symmetrical filter (only 14 values are programmed), meaning this stage is necessarily linear phase. The filter coefficients are programmed over I2C, so recovering them is trivial.

These graphs show the impulse responses of the first filter stage. They all have the by now familiar basic shapes.

 

The second stage filters are all simple triangle shapes of the same width but varying in height.

 

Combining the two stages results in these impulse responses.

 

 

This, finally, is the frequency response of the combined filters. Of course, the headphone amplifier has rather limited bandwidth, the actual output spectrum looks a little different.

 

Here we see the spectrum up to 400 kHz with a white noise input using MQA filter 10. At low frequencies, the response is recognisable from the previous plot, then the limitations of the amplifier take over.

 

Then there is the shaped dither. The ESS chip can't be made to perform this function, so it has to be done by the microcontroller. This is feasible as it requires far less CPU power than the resampling.

 

This graph shows the recorded spectrum playing back silence using filter 3 and the each of the dither shaping options. Although the stream format allows four values for the shaping, two of them give the same output. Note the symmetry around 48 kHz. The dither is indeed applied before upsampling. Contrast this with the Bluesound which applies dither after upsampling.

 

The digital output of the Bluesound renderer when upsampling to 192 kHz, while matching the shape from the Dragonfly, is stretched by a factor 2 and doesn't exhibit any symmetry. This is because here the dither is never subjected to the leaky MQA filters, and the shaping is able to concentrate the dither energy towards the highest frequencies as intended.

 

In summary, MQA rendering on the Dragonfly works exactly as I theorised it would and Gordon Rankin vehemently denied.

[mansr]

49 minutes ago, semente said:

There's nothing to be surprised about. People like what they like and their assessment is easily influenced by magazine reviews and hype...

Moreover, while some of the graphs aren't pretty, and there are better-performing DACs, the distortions seen here are unlikely to be at an audible level when playing music. Human hearing just isn't all that great.

[mansr]

1 hour ago, blownsi said:

Would love to see comparison measurements on the DF Red.  I have read numerous posts on other forums that the Black sounds horrible and the Red is great. 

I have no intention of buying a Red. Sorry.

[mansr]

1 hour ago, Vitor said:

What’s the alternative that the OP suggests as being cheaper and at least with the same performance as the black?

I don't recall making such a suggestion.

[mansr]

Just now, Vitor said:

There you go.

Note the word "probably." I haven't personally tested any, so I can't make a specific recommendation.

[mansr]

20 minutes ago, Vitor said:

If you haven’t heard any you can’t even remotely state that they’ll perform the same IMO. Cheaper or not.

All I said was that the same performance ought to be achievable for less.

[mansr]

1 hour ago, Vitor said:

I understand. However, your posts/tests are based in objective reports/measurements.

With that sentence you entered in a very subjective ground. Just sayin.

What's subjective about making a rough estimate of the cost of achieving a given performance level? Component prices are not secret.

[mansr]

Since the price issue seems contentious, let's look at what goes into the Dragonfly. Bulk pricing used where available.

• PIC32: $3.02
• ES9010: $6 (bulk price unknown)

• TPA6130A2: $0.48

• Crystal: < $1 (exact part unknown)

• USB connector: < $1 (exact part unknown)

• Headphone connector: < $1 (exact part unknown)

• Resistors: < $0.10 total

• Capacitors: < $0.10 total

• Other: < $1

That's less than $15 in parts. With bulk pricing on the DAC chip it is probably closer to $10. PCB assembly adds maybe $1. Adding a few dollars more for the case, final assembly, and packaging, we're still looking at around $15, $20 at most, in unit cost to manufacture. The engineering of a product like this should cost no more than $50k. Amortised over 10k units, that would be $5 apiece. This gives us a conservative estimate of $25 for the total production cost. Allowing a 100% dealer markup, we're still only at $50, whereas the Dragonfly sells for $100. Someone is making a tidy profit.

[mansr]

1 minute ago, vortecjr said:

Aren't you leaving some things out...  

Like what?

[mansr]

5 minutes ago, Miska said:

For hifi hardware, BOM is typically around 10% of the sales price.

And the Dragonfly is no exception, apparently.