Topping E50 DAC Review and Measurements: I Can't Go for That

being a pathologist for bad gear isn't very fun. it's like figuring out what killed a 400 pound guy who's bound to explode like the fat man over nagasaki

how many good sounding records have 10 khz content that high though?

the topping could also have garbage from its switching power supply. they're not the type of manufacturer who will test them like apogee, motu, grace, dangerous etc

 

Are we missing an opportunity to call them Atomic NADS (tm)? I mean, right?

 

Yup, that still is the aliasing grass. 19kHz at 44.1kHz sampling rate has the aliasing bands at 25.1, 63.1, and 69.2 kHz:

Failing to knock them down with FIR filters made that 25kHz noise. I can't find ESS9068 datasheet, but your measurement of the filters are pretty inline with the ESS9069 datasheet. That's exactly where every DAC chip engineers were stuck at, doing a longer filter takes more computing power than what they can fit in the chip and budget. Also the stereotypical ESS9018+OPA1612 engineers don't bother upgrading the filter cuz it already looks good enough for 1kHz measurements. Tuning with ear would lead to completely handling all the digital stuff outside of the chip, where the engineers know what's going on.

I would agree in theory. -140dB shouldn't make any difference, but there are way too many shits in the digital path that weird things happens. Clipping protection takes away 10ish dB, and the music is playing at -20dBfs. Heck even OSing and redithering 16 bit files to 24 bit files makes them sound flatter than 32 bit files on my Gungnir Multibit, which is only a 18 bit DAC and does OS+NS again in itself. Trimming away the ultrasonics makes the headstage smaller too. My hearings just don't agree with that DACs don't need more than 20kHz and 96dB assumption.

That said, I could somewhat fixed my RME ADI-2 by doing OS in computer and adding pink and/or white noise, but it still hurts my ear, albeit a little more tolerable. There's definately some other shittiness not being measured out yet.

 

Just one thing, the E50 isn't one of their latest products; the ESS9068 was also the next ESS chip after ESS9038PRO and ESS9038Q2M, with the 9068 one including an on-board, always-enabled MQA decoding capability. yay.. the Chinese really did love MQA after all. It's a shame it's f'ing dead.

ESS latest chips are in fact ESS9039PRO and ESS9039Q2M. I read through the diyaudio threads of those and it's bonkers, the word salad that is thrown around there lol would not recommend. Anyway it seems the ESS9069 is the latest in the ESS906- line, probably containing some 'bugfixes' because that what ESS seems to do nowadays, less so actual innovation.

 

Back in the 9018 days Weiss were the ones asking for implementation advice on DIYA. ESS was legendary bad with their datasheets and specified an 80MHz master clock. The DIYA guys found out that 100MHz worked miles better. Since then I've had reservations about ESS.

As for the new chips - ES9068/69 is a 2-channel SOIC with ADC and other amenities on board. Not really their flagship. Their flagship still is the ES9038/39Pro with appropriate channel summing done. ES9039Q2M is their 2-channel flagship which is usually used in lower-tier offerings by the SINAD gang.

Generally, most ESS chips can be used in either voltage or current output mode. Voltage output gives the infamous Sabre IMD hump but is easier to implement. Current output is harder to get implement as DAC chips get hot and require heatsinking and the output stage design becomes less trivial. I remember the DIYA gang using 4U amp chassis to do their Sabre DACs with Pass-designed IV stages.

 

The only ESS DAs I’m 100% happy with I’ve tried are the newer Apogee Symphonies, ie Symphony mkII 2x6 SE, Symphony Desktop, and now there’s a new lower power Symphony 16x16 SE module. They cheaper ones have higher imd but still sound good and older Symphonies and Quartet had heat issues.

 

No. That’s total bullshit. Clipping protection takes away no real world headroom. Stop playing clipped music or lower gain digitally on your computer for it. There is no penalty with dithered audio directly outputted to the converter. Gungnir multibit has a noise floor of what? -90 or so line level +4dbu? That’s still lower than any tape recording nevertheless a multitrack tape recording, about that of a cd, and lower than the quietest mic in the world amplified by 50 db, and definitely lower than your hvac system.

32-bit floating point pcm has more rounding errors than 24-but fixed pcm. Good luck hearing that distortion.

SRC is inaudible now. See the Voxengo and Goodhertz ones. It’s inaudible. You cannot hear the ringing especially when going upwards. The only difference is your converter might suck ass at a different sample rate. There is no distortion beyond the rounding of 64-bit floating point operations -300 ish dbfs. We’re talking atomic or subatomatic particles here beneath the thermal noise of any resistor. you’re not hearing it and modern productions are full of constant src to run operations at their optimal sample rate.

16-bit to 24-bit does absolutely nothing except pad the file with unneeded headroom and add dither noise -140 something down.

You’re just feeding an 18-bit dac 24-bit audio and wondering why it might distorts? get a life, get a beer, it will make your music better instead of making it worse.

 

Sorry I wasn't being specific. I was talking about the clipping caused by FIR filter and noise shaping. The FIR filter can cause the maximum transient increase by 5-15ish dB, varing from implementation. Minimum phase ones needs a couple more dB. The DSD modulator's noise shaping also needs 6dB of headroom.

Again, sorry for omitting the details. I was doing OS and NS in my computer. 16bit file OS with 4096 taps 32bit FIR filter makes 60 bit files. If I don't dither it before truncating to 24 bit for my DAC, it will sound like ass and gives me fatigue. It's way more noticable with 1 million tap length. In theory 24 bit dithering shouldn't do anything to the noise floor of files already dithered at 16 bit, but the difference was still audible for me compared to redithered 32 bit files.

I would agree with you on this. SRC done right can be theoretically perfect and inaudible, especially in computers with great CPU and 64 bit precision. The least common multiple of 44100 and 48000 is only 7,056,000. However the in-chip SRC are not that great. The least common multiple of 44100 and 10MHz (the frequency of the oscillators they typically use for in-chip SRC) is 4,410,000,000. Good luck computing that in bit perfect SRC in real time, let alone the sucky computation power of the in-chip arithmetic units. It demands 1000x more computing power on a 1000x weaker computing chip.

That I absolutely agree! I wanted to write things about this last year around this time, but I never finished. Doing boring AB tests made me doubt my hearing, and hate this hobby when I got listening fatigue. I flip-flopped on my preferences quite a few times, and never really settle with any results. I still don't quite have a life, but a beer would definately be a yes!

 

I see the noise floor on Amir's FFT is a few dB lower than yours. Just out of interest - @purr1n what would your SINAD figure have been had you limited bandwidth to 20 kHz?

 

According to @GoldenOne ESS DACs can exhibit ultrasonic issues due to sloppy chip configuration. He reported 25kHz and 65kHz in the Weiss DAC204 review.

 

Fir filters do not lead to transient increase. That transient is already there and will be there in the 44.1 or 48 khz file after it is reconstructed. The higher sample rate just gives a better connect the dots image on your display or primitive pcm digital meters. FIR filters can reduce perceived transient impact if in the middle of the audible band though or even the upper treble.

Iir filters shift phase and will lead to transient increase in the real world. This is why your brick walled mastered music still has some perceived transients in the real world from dc filters and the crossovers in your speakers.

The only difference you are hearing is minute volume differences and you distorting your own signal. Why the hell are you doing that? 60-bit fixed files? What the f**k? Sub atomic particle bullishit. Why not just use a standard decent dither? You’re creating rounding distortion which is very hard if not impossible to hear that far down, adding a 24-bit dither to a 32-bit float file that can contain the 24-bit word, that’s being fed to a 24-bit asio or 32-bit float coreaudio file, that’s feeding into your dac and eventually being truncation to 18-bits. What the f**k? Why not just stick a 16-bit dither on the output from your media player and feed it to your converter like a normal person?

Your dac chip is doing all other sorts of garbage, mostly has additional dsp chips in it, and best operates at 44.1 and 48 khz even if it needs an external filter. There’s a ton of noise shaping and you have no idea how it f'ing works and they’re not going to tell you. Deal with it. Why do you care? Your choice is you use the TI, AD, Cirrus, or ESS chip solution or you do something else and f**k it up. And the best from those beat every single resistor based converter. Schiit spends a crap ton in build of materials in the Yggdrasil to equal what someone like Lynx does with 20 year old off the shelf chips and the Yggdrasil is worse in many ways. You cannot get 32 channels of Yggdrasil in one 19” rack unit you can literally fry an egg on while the thing still works and almost certainly sounds better than whatever the hell you’re doing.

 

Weiss is overrated now. It was state of the art 20 years ago but their processing has been surpassed by free plugins and standard multichannel converters above the prosumer level. It’s now licensed overpriced Softube plugins and hifi toys for rich guys to oogle at the white casework of because the same brand made an eq that was used on their favorite late 90s and early 2000s records. He also had the best SRC 20 years ago when it was often cleaner to converter to analog and convert it back to digital in the required sample rate.

 

You shouldn't have spikes at the same level as the passband signal with zero-stuffing. The picture looks more like a true zero-order-hold (repeat each sample 2^N times) to me. Zero stuffing should give you about 12dB of attenuation on the first reflection of the passband. I'd post a q&d FFT but I'm not enabled for attachment upload.

The first reflection of that 10KHz signal is at 34.1KHz, so 4100Hz is a second-order intermodulation product of those two. The magnitude of that product will also include other aliasing that generates the same intermodulation product, including the reflection of the entire oversampled DAC output via the D/S modulator, and the spikes at the oversampled Nyquist from the true zero-order-hold of the DAC output. 2050Hz shows up as a higher order intermodulation product in the same way.

It's very unlikely that they're actually oversampling to a zillion megahertz. There's no public datasheet for this part but other related parts (e.g. ES9069) show a fairly standard 8x oversampling, meaning you're going to 352.8KHz from 44.1KHz and then the delta-sigma loop operates on that. Notably that a delta-sigma loop doesn't do anything to attenuate the aliasing of that oversampled rate, at least not in any design I'm familiar with. The only thing it's concerned with is keeping the quantization noise out of the passband, and it does that by pushing it way out.

So why does all this show up? In a word, non-linearity. Something in this DAC is behaving non-linearly with at these frequencies and it's creating a spray of crap. ESS might have tweaked their oversampling filter and D/S modulator over time to increase the attenuation of some of these products over time, but if their reference design is optimized for "good enough" at a minimum eBOM and with a minimal board layout then companies that make their living regurgitating reference designs will have these problems forever.

Here's my rant on this point: measurements will not actually serve the end user until both reviewers and manufacturers agree on a common set of parameters and either stick to them or document when they deviate from them. There actually is a perfectly serviceable definition for these measurements in the form of AES17-2020, and it doesn't seem that anybody actually cares to follow it. In particular, AES17 dictates 20KHz bandwidth for 44.1KHz and 48KHz input sample rate measurements, and if a different bandwidth is used the measurement needs to be qualified with that. Amir uses 22KHz, you're showing 90KHz, and both of these might be interesting numbers but they should be documented. For all I know some reviewers or vendors are using a non-standard notch filter on their measurements too, or measuring at 1KHz instead of the standard 997Hz.

AES17 also dictates how to present measurements, and surprisingly enough "DAC rankings by SINAD at 1KHz" isn't listed. In fact, THD+N at a single frequency and level isn't actually an AES17 measurement at all. I encourage anyone who cares about this and who isn't an AES member already to go sign up for the basic membership ($125/yr, cheaper for students) and at least read through AES17. There's an old copy floating around online but the 2020 version has a lot of sensible changes in it and is worth reading. Disagree with it if you want, but even in that case I'd rather see "test performed per AES17-2020 except using bandwidth of 90KHz and 0dBFS signal level" than just "THD+N" or "SINAD" with f**k-all in terms of detail about test procedure.

 

Rant part deux: Amir loves to wave away concerns about these kinds of distortion products as "not an audible concern" because they're down at -120dBFS next to a pure 0dBFS sine wave. The error in doing so is assuming that nonlinearity will behave, well, linearly. The only thing you can neatly extrapolate from seeing these intermodulation products on a simple measurement like this is that you have measurable nonlinearity that affects the audio band.

 

I just looked across a swath of ESS datasheets and the ones that discussed oversampling seemed to indicate an 8x oversampling rate. 8x oversampling might mean "up to 8x" as in "everything goes to 352.8 or 384KHz", FWIW. Now the result you showed for 44.1KHz is interesting because unless there's a measurement artifact it looks like a spike at 384KHz. I wonder if there is some kind of sample rate conversion going on here. The noise spectrum looks shockingly similar too and that ought to be be affected by the sample rate as well.

EDIT: ESS's "Patented Time Domain Jitter Eliminator" is... an ASRC. I bet it's converting everything to 384KHz.

I'm also not sure what you mean by "upsamples in the traditional multibit sense". Anything other than a NOS DAC oversamples for the same reason and more or less in the same way.

They might have just gone backwards in the reference design. Unless the product is using an I/V conversion stage that's built into the IC, the most likely source of these nonlinearities is external to the DAC.

To be clear my rant was more or less in agreement with what you're saying. If everyone is using the same term to refer to different measurements, the whole thing is utterly pointless. My ire, such as it is, is directed at the companies engaged in the meaningless SINAD race and the "objective" reviewers who enable them.

No disagreement here. My point is that standards are a common language. They might be meaningful or they might be meaningless, but they should neither be blindly followed or discarded without good reason. If they're not meaningful and a common language is needed, they should be improved or replaced. Chesterton's fence very much applies here.

The 997Hz thing is just a brown M&M of measurements to me. Either you're trying to follow a standard (plus or minus deviations) or you're documenting your disagreement and calling that out when you post measurements. You've documented your disagreement and that's fine.

FWIW, I've spent a whole bunch of time deep in cryptography, hardware security, etc. including in solutions for movie studios. I think you'll find that we're pretty much on the same page with respect to the role of standards.