DAC chips with lower noise

For many years I was content to parallel TDA1387s to get lower noise. Its datasheet SNR - the maximum signal divided by the noise at digital zero - is 98dB typical, A-weighted. Since I'm not a fan of using weighting to goose the numbers, I tend to just subtract 3dB from A-weighted numbers to arrive at an approximate (unweighted) figure of 95dB. When comparing chips' specs, its important to filter out marketing fluff like the use of A-weighting.

Since doubling the number of chips doubles the current but the noise only rises by 3dB (because its uncorrelated between chips), doubling the chips improves the SNR by 3dB. 256 chips is 2^8 so should in theory get us an improvement of 24dB, or to 119dB SNR. Fine in theory, but the assumption that noise is uncorrelated won't hold if the source of that noise is the power rail since that's going to affect all chips equally. So attention needs to be paid to the cleanliness of supplies, or alternatively ensure that each DAC chip's supply is separate and hence not correlated with the others.

There are a few practical issues with 256 chips, not least of which is soldering that many chips down to boards. The current demand from the PSU isn't trivial either, around 1.4A. If the chips are packed too close, heating becomes an issue too since the array of chips is dissipating about 7W. Then there's the question of how to deal with a current output of +/-128mA. All these difficulties gradually pushed me in the direction of seeking out other DAC chips. I started at the cheapest multibits (based on Taobao pricing which often means recycled chips) and have begun to work my way upwards....

PCM56 has a DS typical output noise of 6uV. That's specified in micro-volts because it has an internal opamp to give it a voltage output of 2VRMS. Dividing the max output by the noise number takes us to 110dB. Already its equivalent in SNR to an array of 32 TDA1387s but the noise from the current output (which is what I'll use) is going to be lower because the opamp has to add some noise. How much noise is hard to estimate but we can put some bounds on it. Seeing as the opamp sees the 1200R Zout of the DAC to GND and its feedback R is about 3k, its noise gain is (1+3/1.2) = 3.5. So for output noise of 6uV, the noise RTI (referred to input) must be 6/3.5 = 1.7uV. If the opamp were equivalent to an NE5534, that has noise RTI of 0.6uV. (I'm excluding resistor noise here). Such a noise contribution is negligible compared to the DAC's own noise but if the opamp were significantly noisier than 5534 then its contribution would start to matter.

The next chip of interest on Taobao is PCM58. Its datasheet noise is unbelievably low, specified as -126dB, A-weighted. This figure merits some deconstruction as its far above all other R2R DACs I've looked at. Reading the small print its measured with an OP27 (pretty low noise) and a 10k feedback resistor. I wonder why they chose 10k as its not at all typical in an audio application. Perhaps its their way of gooseing the figures? Since the feedback R is 10k, the maximum output level must be 10V peak, or 7VRMS. Taking 123dB as the unweighted number, the output noise must be 5uV. The OP27 noise gain is (1+10k/1k2) = 9.3 so RTI noise must be 0.5uV, pretty impressive. OP27 has 0.4uV RTI noise. I don't think this claim holds up under examination since we've also not included resistor noise. Or perhaps I'm missing out something vital in my analysis. Anyway, even if the real figure is 6dB worse than claimed its still a good deal lower noise than PCM56.

There's something else very interesting about PCM58 which isn't a feature of many other R2R DACs, that's the presence of the 'BPO' pin. PCM63 also has this but its not shown in any other DS for BB DACs to my knowledge, not even PCM63's successors PCM1702 and PCM1704. What 'BPO' stands for is bipolar offset. It turns out the internal DAC architecture isn't bipolar, its single-ended. Meaning it only sources (or sinks) current, it doesn't do both. To make PCM58 output a bipolar current, it needs an additional current source combining with the DAC proper output, this current source is connected to pin BPO. Even AD1862's internal architecture drawing shows a resistor feeding the output from the Vref, making the output bipolar. ( I'm guessing here no DAC is truly bipolar, they're all unipolar and need an additional current source tweaking the output). On AD1862 there's no BPO pin, you cannot disconnect the offsetting resistor. On PCM58 (and PCM63) you can. I think the ability to disconnect BPO is a valuable tweak in the quest for ever lower noise.

 

[abraxalito]

PCM1702/4 are the subject of today's short post. Both headline numbers for SNR are 120dB, A-weighted typical. So 117dB to me.

For PCM1702 :
The measurement is made with 5k Rfb and an OPA627 which is a very low noise JFET. But JFET opamps don't get as low noise as bipolars. The OPA627 DS helpfully gives an integrated noise in a defined bandwidth plot, no other opamp DS has this. From this graph we can expect OPA627 to have 0.85uV input noise in the band 0.1Hz to 20kHz. Going down to 0.1Hz seems a tad pessimistic for audio so let's try another approach - the input noise density at 1kHz is 5.2nV/rtHz. It goes down at 10kHz but integrating the 10k number will give a too optimistic total noise given the LF rise. So the 1k number seems a happy medium, and this gives a figure of 0.74uV input noise, slightly better than the figure including contributions down to 0.1Hz.

The DAC Zout is 1k and so the noise gain is 6 giving an output noise of 4.4uV as against a max output level of 1.2mA*5k = 6V peak or 4.2VRMS. This arrives at 119.6dB SNR. So its quite believable the DAC achieves 117dB, this would have to include the Johnson (current) noise of its 1k output impedance.

For PCM1704 :
The DS says dynamic performance data (which includes SNR) are tested with a 5534 and 7.5kohm feedback resistor. Integrating the 5534's input noise (4nV/rtHz) gives us 0.57uV. Noise gain is 8.5 and max OP level is 9V peak or 6.4VRMS. So 6.4/4.8uV = 122.4dB. Using the 5534 and 7.5k is giving the PCM1704 a nearly 3dB advantage over PCM1702 in the I/V but still the claimed SNR is the same. Hmmm.

 

[abraxalito]

More information about PCM58 has come to light. This paper from IEEE Bipolar Circuits, 1988 describes in detail the design features of PCM58 and ends with a mention of the SNR - '122dB, A-weighted'. The 122dB figure measured with an OPA27. Another reference of note is the PCM58's Japanese datasheet, which gives a -120dB figure here. So I think we can safely ignore the -126dB claim in the English version of the DS.

Thanks go to Nazar Shtybl for hosting these invaluable resources.

 

[abraxalito]

PCM58 update - yesterday after running for about a day without any issues, my balanced PCM58 DAC started generating clicks at random times. After a few more minutes, it started distorting. After power cycling these issues disappeared but came back again after a short while. I first suspected that maybe I'd overlooked some set-up or hold time in the design of the digital logic to interface them with I2S. That checked out OK after examination. The problem was only showing up on one of the four chips so I probed around the various pins on that and happened to notice the voltage level on pin15 wasn't stable, it was decreasing slowly over time. Checking against the other working chip on the same board, its voltage also wasn't stable.

What I eventually took note of was the pin15 voltage was positive wrt GND and the cap connected to it (a 4.7uF tantalum) had its negative terminal to pin15. This is how the DS shows the pin15 decoupling cap, but based on the voltmeter reading, the DS is in error. Now few people I suspect will encounter this error because the value shown on the DS for the cap is 100nF which means designers are unlikely to select a polarized cap. But then again - so are all the other caps doing decoupling duty on the DS and its clear from reading the text that 100nF isn't sufficiently large for LF noise reduction. So if anyone else is inclined to try a larger, polarized value for the pin15 decoupler, connect it the opposite way to the way the DS shows. With all my pin15 caps duly reversed, no more clicks or distortion and the pin15 voltages are stable over time.