iFi foe fum

[mansr]

A while ago, iFi released firmware version 5.30C for most of the their DAC models. The big news in this update is what they call the GTO filter. This filter is, according to iFi, so good that they produced a 9-page PDF espousing its virtues with much technobabble and several outright falsehoods. They even "feel" this filter is so fantastic that they decided to replace all the previously available filter choices with this one. In other words, the little filter selection switch on the DAC now does nothing at all.

 

In order to get a closer look at this magical filter, I connected a logic analyser to the inputs on the DSD1793 DAC chip. With this new firmware, the chip is put in filter bypass mode, the XMOS now upsampling everything to 352.8/384 kHz. Only when the input already has one of these rates is it left untouched.

 

While preparing to capture the GTO filter impulse response, I noticed that non-zero data was being sent to the DAC chip even with no input over USB. As it turns out, the firmware is truncating the samples to 18 bits with shaped dither. The figure below shows the spectrum of this dither along with TPDF dither at 16 and 22 bits for comparison.

 

 

As we can see from the graph, the dynamic range is equivalent to 22 bits for frequencies below about 24 kHz. The noise level then rises, at 96 kHz reaching a level corresponding to 16 bits, then remains flat. This noise spectrum is the same whether the input is 48 kHz, 96 kHz, or 192 kHz. With 384 kHz input, no dither is applied. The analogue noise level of the DSD1793 falls somewhere between the high and low levels seen here. This means the added dither noise somewhat reduces the dynamic range at high frequencies. For some reason, no mention of this is made by iFi.

 

The presence of the dither noise interferes somewhat with capturing the impulse response of the GTO filter, but by averaging a large number of responses, the effect of it can be minimised. Starting with a 48 kHz input, we obtain this impulse response:

 

 

And the frequency response:

 

 

The filter is close to minimum phase, but it does not meet the strict definition as a few of its zeros are outside the complex unit circle, not that this is important. More interestingly, it has a slow roll-off reaching only -2 dB by 24 kHz and -90 dB at 48 kHz. It then stays (mostly) below -80 dB until the vicinity of 150 kHz where it shoots up to -40 dB before falling off again towards the end of the spectrum. Besides an overall unimpressive performance, any high-frequency content in the 48 kHz original will see poor attenuation of its images, both immediately above 24 kHz as well as around 168 kHz. This can lead to problems with intermodulation distortion.

 

To check for IMD resulting from the poor upsampling filter, we use a single 22 kHz tone at -3 dBFS. First, the spectrum on the analogue output when using a 384 kHz sample rate, bypassing the GTO filter:

 

 

Not the prettiest spectrum, but at least none of the distortion products are audible. Next we send the tone using a 48 kHz sample rate.

 

 

This may not be the ugliest spectrum I've seen, but it is definitely in the running. These distortion products are plainly audible as a nasty, harsh whine.

 

On closer inspection, it turns out that the filter is clipping when presented with this input, and this contributes to the distortion. Lowering the amplitude to -4 dBFS avoids the clipping. Below, the resulting spectrum of a digital capture of the filter output.

 

 

All the images are as expected from the filter frequency response. The shaped dither noise is also clearly visible. Additionally, there are some smaller spikes at 2 kHz intervals below the primary one. These are probably due to rounding errors in the filter implementation. The spectrum on the analogue output is now much cleaner:

 

 

Cleaner, but still far from clean. Compared to the digital signal, harmonic and intermodulation distortion, rather than images, dominate the picture. This includes a readily audible component at 4 kHz.

 

Moving up to 96 kHz, the impulse response looks like this:

 

 

And thus the frequency response:

 

 

All the characteristics remain. Minimum phase like, slow roll-off, poor stopband attenuation, and an even bigger hump towards the end.

 

Finally, at 192 kHz, the impulse response has shrunk to a mere 5 samples:

 

 

The frequency response is equally simple-looking:

 

 

That's barely even a filter at all.

 

Reaching a verdict on this update is easy. GTFO: Get This Filter Out of my DAC.

[mansr]

16 minutes ago, Miska said:

The filter doesn't seem so far from the MQA filters. And it was introduced at the same time with MQA support. So I'm wondering how much there is actually common ground between the two...

They say it "was developed by iFi according to our specifications in conjunction with the MQA team."

[mansr]

13 hours ago, Em2016 said:

And they knew some of you would wonder this so they added: 

 

"We must make clear that GTO is not directly related to filter types used by MQA, it is not “MQA through the backdoor”

How is it then that the 96 kHz and 192 kHz versions are exactly the same as MQA filters found in the Bluesound firmware? I don't immediately see a direct match for the 48 kHz version, but the MQA decoding wouldn't normally need such a filter either.

[mansr]

12 minutes ago, Em2016 said:

The good thing about the current iFi DACs is you have options to use other non-MQA related filters, so not too big a deal.

By using the older firmware, yes. Will this work on their next DAC, whatever it turns out to be? I'm doubtful.

[mansr]

It turns out the horrible distortion from the 22 kHz tone is partly due to the filter clipping. I have update the OP with some more info.

 

In case anyone wonders, decent filter implementations do not clip from a pure tone at -3 dBFS.

[mansr]

7 hours ago, Jud said:

Or you can upsample non-MQA material with Audirvana+, HQPlayer, or some other program

One problem. The DSD1793 works better with 192 kHz input than 384 kHz. This is because the 192 kHz input is upsampled 8x to 1536 kHz, greatly reducing the demands of the analogue filter (external to the chip). With the weak 2nd order filter of the Nano, this makes a real difference.

[mansr]

9 minutes ago, Jud said:

Is upsampling to, for example, DSD256 what you would describe as putting greater demands on the analog filter?

I was talking about PCM only. DSD is a different kettle of fish with its own issues.

[mansr]

7 minutes ago, Em2016 said:

And what happens with 384kHz input? Up-sampled to what rate?

 

Or it's not up-sampled to a higher rate?

384 kHz input is only possible in filter bypass mode, so that isn't upsampled at all.

[mansr]

6 minutes ago, Em2016 said:

Can you share what these issues are, with feeding DSD256 to your iFi DAC?

DSD64 has enough modulator noise at low frequencies that it can cause audible distortion artefacts. DSD256 is reasonable.

[mansr]

3 minutes ago, Shadders said:

I tell you what, i was looking at the iFi Dac literature, and their plots look remarkably like the Stereophile ones.

 

I reckon Stereophile borrowed them from the iFi literature. When iFi finds out, i think they ain't gonna be happy. I mean, copying someone elses drawings, is an utter disgrace. It just goes to show, you can't trust some people.

iFi already admitted to pilfering the plots from Stereophile. Now let's leave that topic aside, however amusing it may be.

[mansr]

2 hours ago, crenca said:

Do you or anyone else know if the analogue filter in the Nano is the same as the one in the micro iDAC2 (non black label version(s) - and for that matter do the black lable version have a different analog filter)?  I have a Nano I travel with, but I use the micro iDAC2 in my HP rig when I am not using some other DAC...

The Nano has a simple, passive 2nd order RC filter between the DAC and the headphone amp. The iDAC2 has a more elaborate active filter incorporating an OPA1654 opamp. I'm not sure of the exact topology.

[mansr]

4 minutes ago, crenca said:

Thanks.  Do you know enough about the topology to say if either the passive filter in the Nano or the OPA1654 in the iDAC2 is in the signal path when using the RCA outs?

The low-pass filter is required with the DSD1793 chip, so yes, it is used with both RCA and headphone outputs.

[mansr]

I played some CD quality music and looked for clipping in the data sent to the DAC. Here's an example, original in blue, GTO-filtered in red:

 

[mansr]

25 minutes ago, Em2016 said:

As a baseline for comparison, can you share the same measurements with the DSD1793 doing the digital filtering , for this same DAC?

The best I can do in that regard is recording the analogue output. I did that a while back: