Let's Build a Phono Stage

Discussion in 'DIY' started by peef, Aug 14, 2016.

  1. peef

    peef Almost "Made"

    Joined:
    Oct 5, 2015
    Likes Received:
    108
    Dislikes Received:
    0
    Trophy Points:
    33
    Earlier last month, a friend decided she was getting a turntable and asked me for advice. Looking at phono stages, I drew a blank. So, we’re going to make one.


    Some general requirements

    1. There should be no hum, hiss, or buzz.

    Kind of ruins the experience, no? This means the circuit should have excellent PSRR, or an excellent power supply, or ideally both. This also means no ground loops and no EMI.

    2. It should also sound like a single ended phono stage.

    That means more second order harmonic distortion than third, more third than fourth, and so on. There should be no oscillation either, and we’re going to do our best to minimize sources of distortion that can impart a “solid state” sound.

    3. RIAA equalization should not rely on feedback.

    Feedback can be a very lovely thing, but not so much in RIAA equalization. Notably, feedback RIAA works by varying the feedback ratio according to frequency, such that you have a higher ratio at high frequencies and a lower one at low frequencies. An unfortunate consequence is that your input and output impedances are also frequency dependent. So is your distortion. This seems like a bad tradeoff.

    4. The user should have full control over cart loading, and be able to plug into any reasonable load.

    Alternatively, this means that input impedance should be very high, and output impedance should be low. We can then tailor the input impedance to our needs with loading resistors and capacitors. But, we should have the freedom to load the cart with no more capacitance than what the cables provide.

    5. The whole thing shouldn’t draw more than a few watts.

    For goodness’ sake, it’s a phono stage. Radiated noise is proportional to current, so it’s really in our best interest to keep AC currents down anyway.

    We’re also going to be building to a price point, and strictly for MM. With compelling options in the $200-500 range, it does not make sense to blow that much on fancy Lundahls or Duelunds. Even though Lundahls are lovely.

    Meet the cascode
    For high input impedance and gain, you’d be hard pressed to do better than a cascode. A cascode is a common emitter/source/cathode stage working into a common base/gate/grid stage. They might look like this:

    [​IMG]

    The lower device is a transconductance amplifier. The current through the cascode is modulated by applying a voltage signal between the +In and Ref (In) nodes. The current is then converted back to a voltage using a resistor, and we can sample the output voltage across it on the -Out and Ref (Out) nodes.

    The top device is a transimpedance amplifier. Its role is to provide a low impedance load for the lower device such that no voltage signal appears on its collector/drain/plate. It typically takes the bulk of the dissipation, and it’s this device that’s going to swing the signal voltage; so, figures of merit would be a low input impedance, a high output impedance (i.e. its current is not affected by the signal), and good linearity vs current.


    Here's an example of an all-triode cascode. The Muscovite is one of the few phono stages that does this.

    [​IMG]

    Let’s say the triodes are ECC88s, which have a mu of 33, a plate resistance (rp) of 2.7K, and a transconductance (gm) of 12.5 mA/V. In this circuit, the cascode gain is 8 kΩ * 12.5 mA/V, or about 100. Not bad, considering mu is a third of that.

    Tube folks like to solve their circuits with curves. We can do that too, but probably not with the curve that you have in mind; we use the Ia/Vg curve instead of the familiar Ia/Vp curve. Linked.

    Manufacturers of fine semiconductors will also include a similar graph in their datasheets. Below is the Vgs/Id curve for the 2SK170. For both devices, linearity goes up with current, but the tube really is the more linear of the two devices. Mind you, this isn’t really a problem for a phono stage; the signal magnitude is so small that we shouldn’t need to swing more than a few hundred microamps.

    [​IMG]

    If you mix device polarities, you can throw together a folded cascode. In its simplest form, it might look something like this.

    [​IMG]

    The principle is the same: the upper device (PNP) presents the lower device (NPN) with a low impedance such that the lower device develops no voltage gain. Instead, it conveys signal current into an I/V resistor, which, in contrast to the regular cascode, is referenced to ground.

    An interesting benefit is that the output can be at the same DC voltage as the input, which makes for easy DC coupling. If I were Nelson Pass, I might connect the -Out back to the +In through a feedback resistor and call this SuSy. Or if I were Rod Coleman, I might replace the bottom BJT with a triode and call it a “shunt cascode,” then use that to drive a 300B.

    There are caveats, though. There are now two paths to ground, so you need twice as much current. Biasing the whole thing isn’t trivial, as it’s a balancing act between gain, idle current, and voltage drop across the devices and gain resistor. I’ve had troubles with oscillation that I couldn’t track down. Unlike the regular cascode, it can be a real cow to linearize, too.

    Our first stage: a “current shunting” cascode
    With all that in mind, we can throw together a first stage. We’ll be stealing a trick from Gary Pimm, and load our cascode with a current source. Typically, this isn’t possible; the cascode is a current source, so we would be putting two current sources in series. They would fight to define bias, until the lower impedance of the two is pinched off. The trick is to throw in a shunt resistor, which both defines the DC bias condition and limits the gain to something reasonable, and reasonably independant of frequency.

    Here's what it looks like, simplified.

    [​IMG]

    At the heart is a parallel JFET – BJT cascode. The rest is just there to make sure it turns on.

    The cascode sinks a current set by the JFETs and their source resistors. In order to define the plate voltage, we run a small current through the 20k load resistor.

    What’s the benefit over our normal cascode? There are two big ones. If we were to run a normal cascode at 40mA, with a 20k load, the resistor would need to drop 800V. There is nothing wrong with 800V supplies, but if we are even considering one, we should be doing something more interesting with it than heating our listening rooms. The other advantage is that the output signal and input signal are both referenced to ground; the power supply is completely removed from the current loop, and the power supply rejection goes from essentially zero— probably the biggest strike against the cascode— to something usable.

    Other than the noise benefit, parallel JFETs let us run the cascode at higher current, because the current that each JFET can sink is limited by Idss. Why is this beneficial? The impedance that the JFET drain sees is dominated by the BJT’s emitter resistance, Re, which is a function of current and temperature. For the lowest impedance, we want the current to be high, and the temperature to be cool.

    So how much gain can we squeeze out of it? With 68 ohm source resistors, we’re looking at a gain in the range of about 50dB. Just enough to use a single stage.

    A quick word on bias. A good bias supply should have a low impedance and noise across the frequency range. You might think the best bet is to use one of TI’s fancy shunt references, but it turns out they are unilaterally terrible.

    [​IMG]

    Instead, we’ll use a Zener. On a bad day, a 6.8V Zener will put out 20µV of broadband noise; this is actually less than an LM4040, which the datasheet specifies at 80µV. Depending on current running through the diode, its impedance is anywhere from 2 ohms to 700 ohms, but, unlike the LM4040, it is constant with frequency because it does not rely on feedback to keep its impedance down. Zener diodes do present a small inductance, so we’ll bypass it with a small capacitor for good measure.


    Next time: the RIAA network and second stage.
     
    Priidik, purr1n, bazelio and 11 others like this.
  2. uncola

    uncola Friend

    Friend
    Joined:
    Sep 27, 2015
    Likes Received:
    458
    Dislikes Received:
    1
    Trophy Points:
    63
    Location:
    Oahu, Hawaii
    preorder link missing
     
  3. peef

    peef Almost "Made"

    Joined:
    Oct 5, 2015
    Likes Received:
    108
    Dislikes Received:
    0
    Trophy Points:
    33
    uncola, I'll always have an early bird special for you. <3


    The RIAA network
    We need a filter, and we have sufficient gain to go passive. The only thing more dreadful than reading about RIAA filter calculations is writing about it, so we’ll discuss application instead.

    Our cascode’s voltage gain (Av) is the product of the cascode transconductance (gm) and load impedance (Zload or RI/V).

    [​IMG]

    In the last post, Zload was a constant resistive impedance. The trick here will be to replace it with a frequency-dependent load that will eat more gain at high frequency than at low frequency. We can do this with capacitors, which sound terrible, or with inductors, which are hard to get, cost more, and could make the amp oscillate ( (and oscillation will definitely sound terrible). I’ll take the caps.

    [​IMG]

    Note that the RIAA impedance appears in the gain and output impedance equations. The output impedance decreases with frequency. For a 20 kΩ gain resistor, we’re looking at 110 nF and 36 nF, and 2.87 kΩ and 87 Ω.


    Power supply rejection
    It’s worth noting that PSRR will now also be frequency dependent. How does that work? Well, think of the circuit as a voltage divider.

    [​IMG]

    We can approximate the constant current source as a very, very big resistor in parallel with a very, very small cap. Gary Pimm measured a few and 100 MΩ // 10 pF is fairly routine. My own measurements show a higher capacitance, but suffice to say that 10pF is already so small that differences may very well be due to my wiring.

    The equation for our voltage divider is:

    Vout = ∆V+ × ZRIAA / (ZRIAA + ZCCS)

    At low frequency, R dominates:

    Vout = ∆V+ × 20 kΩ / (20 kΩ + 100 MΩ) = 0.00020 = -74dB-ish

    For AC, it’s easiest to sim the result due to all the poles and zeroes.

    [​IMG]

    We can reasonably count on some PSRR up until 200 kHz. Not bad. That's more bandwidth than your typical regulator offers, and we can expect to kill the rest with a passive solution.


    The second stage
    For the filter to work correctly, we want to terminate it with a high impedance. A follower.

    [​IMG]

    No, no, not that kind of follower. That’s just silly. This kind of follower.

    [​IMG]

    By using a P-channel device, the power supply is completely isolated from the circuit. Even loaded with a CCS, the follower’s PSRR is only so-so. A BJT, a MOSFET, or a JFET would all work, but the BJT probably works a little bit better because of its lower input capacitance.

    Naturally, this will require a DC blocking cap on the output. But with at least two equalization caps in the signal path, the circuit has squarely crossed over into give-up-on-life territory; so, no harm in adidng a third, right?

    With that, we have a complete amplifier circuit.


    Next time: powering up.
     
    Last edited: Aug 27, 2016
  4. Hrodulf

    Hrodulf Prohibited from acting as an MOT until year 2050

    Friend
    Joined:
    Sep 28, 2015
    Likes Received:
    2,698
    Dislikes Received:
    4
    Trophy Points:
    93
    Well, this is meant for listening how a needle scrapes a spinning piece of plastic. So we're way past that!
     
    sphinxvc and peef like this.
  5. peef

    peef Almost "Made"

    Joined:
    Oct 5, 2015
    Likes Received:
    108
    Dislikes Received:
    0
    Trophy Points:
    33
    Powering up
    Selecting a power transformer is annoying.

    For the lowest radiated field, we want a toroid. But, toroids buzz, and do a phenomenal job of coupling junk from the line, that can be difficult to filter if you’re designing around 120 Hz.

    What we really want is a cheap split-bobbin transformer on a crappy laminated steel core. These have much lower bandwidth, but also a much larger stray field. Not great, with signal levels as low as these. A gapped C-core power transformer like Lundahl's strike a good between both issues, but, well, Lundahls.

    A third option is a switching supply. I know, gross. But maybe not. Diodes in “linear” supplies switch, too, and generate all sorts of noise in the MHz range. Full wave bridge rectifiers are the worst in this regard. We can deal with the junk, but it’s decidedly trickier when you’ve also got 120 Hz noise.

    There’s also nothing preventing us from regulating a switching supply. It’s less efficient, sure, but this is audio we’re talking about. Who cares if things get toasty?

    We'll be using the Mean Well RS-15-48. It's a 48V 15W switcher rated at 200 mV pk-pk of ripple at a switching frequency of 132 kHz.

    Quantifying the unknown
    Take a look at the datasheet, and you’ll quickly find it entirely useless for our uses. How much noise is there? Where is that noise?

    Here’s the output of a 12V Recom switcher into a 47 µF aluminum polymer electrolytic capacitor.

    [​IMG]

    Awful. Adding insult to injury, the period between the switching peaks corresponds to 20 kHz, not the 130 kHz typical switching frequency that the datasheet quotes. What gives?

    Well, maybe we’re not looking at the right place, and found noise that’s not from the SMPS oscillator. And perhaps the capacitor isn’t helping.

    Switching regulators rely on feedback to set their output voltage. Depending on the design, it may place the load in the feedback loop. So, place a cap directly across the output, and you’re effectively slowing down the loop. This is similar to loading an op amp with a capacitor. Removing the capacitor will raise the switching frequency, and doesn’t seem to affect the magnitude of switching peaks.

    So how does the Mean Well look? Here are some shots off the scope. It’s loaded with a 20 mA current source, and no capacitance. Note that this scope works by sampling the signal, based on the time scale; so, it will only show us the frequencies that we are looking for.

    [​IMG]

    The fundamental works out to 60 Hz, 30 mV pp. This isn’t our switching noise, but you can bet it’s audible.

    [​IMG]

    Now we move up to 8kHz or so. This signal was “riding” the 60 Hz wave, but we couldn’t see it in the previous shot. The dips likely extend well below the 3.5mV that the scope measures, because they’re at a much higher frequency.

    [​IMG]

    And here’s what’s going on around 100 kHz. I couldn’t get my scope to trigger, but there are at least two signals of interest here: peaks around 130 kHz, and a much smaller signal at about 500 kHz.

    So, two takeaways here. First, we have to deal with noise over at least five decades, which will require more than a simple regulator or bulk capacitance. And second, we have less than 50 mV of noise to eliminate on the raw DC supply.

    No single regulator or filter can reliably span five decades. It makes sense to split the burden between three stages:

    1. A regulator that spans the audio band;
    2. A passive filter for high frequency noise; and,
    3. A lossy filter for VHF noise.

    The regulator
    A short, but related digression.I often make cocktails for friends, but, being the experimental type, had my fair share of mishaps. A particular aviation of questionable proportions prompted a friend to advance what we call the Pepsi Criterion:

    “A cocktail is only worth making a second time if it’s at least as tasty as Pepsi.”

    It makes sense. Pepsi is one of the most engineered sensory experiences on the planet. It’s also ubiquitous, costs less than a dollar and comes in a can. If we can’t do better than that, why bother? The list of drinks that did not pass is a bit embarrassing.

    It makes sense to have a Pepsi regulator. We’ll use a Zener shunt regulator.

    What? Not an LM317? Well, no. For starters, it can hardly handle the ~35-40V supply we need—its regulation drops from 70 dB to 40 dB at the upper end of their range, possibly because they don’t benefit from the high feedback ratio of lower output voltages. Second, their dynamic characteristics are handily beaten by a Zener shunt regulator.

    Shunt regulators also ensure that the SMPS and its feedback loop are fully isolated from the load. If you believe in current loops, that means none of the audio signal is flowing through the SMPS. It also provides a constant current draw, allowing the SMPS to perform its best.

    [​IMG]

    Yep, another voltage divider. Regulation improves as the series device’s impedance increases and as the shunt’s impedance decreases. We can increase the series impedance considerably by replacing the resistor with a current source.

    A quick sim shows that for 50 mA, 39 V output, a BJT CCS works best because of its higher transconductance and lower capacitance. It provides constant impedance across the audio band, and about 20 dB greater regulation. More parts, but eh. It’s just sand.

    [​IMG]

    [​IMG]

    To push down the Zener impedance, we can replace the single 39 V diode with a string of lower voltage diodes.

    The obvious next step is to wrap the Zener string in a buffer. A BJT follower divides the Zener impedance by hfe. But this doesn’t work so well in practice. The dynamic impedance of a Zener depends strongly on the current going through it; if we add a buffer, we necessarily rob the diodes of some of that current. The net result is that, buffer or no buffer, regulation and output impedance are about the same. So, only Zeners today.

    The power supply filter
    The raw supply filter is going to require a bit more finesse, and few filters are quite as classy as a Bessel. Why a Bessel? These filters are the least prone to ringing, because they offer constant group delay at the cost of a bit less attenuation. I’ve had good luck with a simple third order filter, and got about 40dB attenuation at 100 kHz.

    [​IMG]

    We could replace the inductor with a common mode choke, for some common mode filtering as well. But it's not quite as robust-- it relies on the inductors being well coupled, after all-- and we can will get great performance without it. I don’t trust my measurements to be quantitative past that frequency, but the filter doesn’t show any sign of letting up well into the lower MHz. Simulated response below.

    [​IMG]

    There’s one last trick we can pull. Ferrite chips.

    A lossy filter
    Ferrites are frequency-dependent resistors (FDR). Like caps and inductors, their impedance is not constant with frequency. Unlike these devices, their impedance is mostly resistive. Here’s a plot of a ferrite’s Z vs f.

    [​IMG]

    What’s the difference between an FDR and an impedance? Capacitors and inductors store energy; resistors and ferrites dissipate it. Because a ferrite has a small reactive component, care must be taken to choose one where the reactive portion won’t cause resonance.

    [​IMG]

    The impedance plot shows a fairly high impedance over a very narrow range, or a high-Q ferrite. So, if our lossy filter is not correctly damped, switching noise will be amplified. This plot from AD’s application note shows a 20dB resonance. If there's one thing this forum's taught me, it's that ringing is a very bad thing indeed.

    [​IMG]


    The solution is twofold:

    1. Use a low-Q ferrite with a low crossover frequency;
    2. Decouple with a low-Q capacitor; or, better yet, synthesize one with a high-Q capacitor in series with a small resistor.

    The complete power supply filter is below.

    [​IMG]

    And, with the regulator, here is our simulated PSRR, hopefully accurate into the lower MHz. What's certain is that there will be no ringing in the passband.

    [​IMG]

    And output impedance. 1 Ω might seem high if you're a numbers person, but the amp presents a reasonably linear 10 MΩ+ impedance to the power supply, and that's being conservative.

    [​IMG]


    Next time: Boards boards boards.
     
  6. peef

    peef Almost "Made"

    Joined:
    Oct 5, 2015
    Likes Received:
    108
    Dislikes Received:
    0
    Trophy Points:
    33
    No big post this week, but I prototyped the power supply without ferrites, which aren't in yet.

    [​IMG]

    It is very quiet.
     
  7. Cspirou

    Cspirou They call me Sparky

    Friend
    Joined:
    Sep 27, 2015
    Likes Received:
    4,950
    Dislikes Received:
    5
    Trophy Points:
    113
    Location:
    Northwest France
    You plan on doing point to point wiring or designing a PCB?
     
  8. peef

    peef Almost "Made"

    Joined:
    Oct 5, 2015
    Likes Received:
    108
    Dislikes Received:
    0
    Trophy Points:
    33
    Definitely a PCB-- point to point would be a disaster. Just threw this together as a quick test before finalizing boards. Don't want to add to the pile of things that didn't work outside of Spice. The phono will require fewer clip leads, too. :)
     
    HitmanFluffy likes this.
  9. peef

    peef Almost "Made"

    Joined:
    Oct 5, 2015
    Likes Received:
    108
    Dislikes Received:
    0
    Trophy Points:
    33
    Boards
    Finally, boards. Three of them. And a quick word or two on part selection. The “division” is a bit different from what’s typical. Instead of a power supply board and an amp board, I opted to split the whole thing between three boards. The DC filter is on a standalone board; the shunt regulator’s current source gets its own board; and the amp and shunt leg of the regulator go together on the same board.

    There are a few benefits to doing it this way. Shunt regulators are prone to self-immolation if the connection between the load and the supply is severed. Here, if the connection is severed— say you’re using a crappy clip lead—, the current source will pinch off, and nothing can really go wrong at all. We no longer need to worry about the impedance of the wire between the supply and the load or anything it might pick up, because the low impedance connections are all on the amp board.

    The DC filter could go on the CCS board, but it’s not much more expensive to put it on its own board. And besides, standalone current sources are worth keeping around.

    Perhaps most importantly, boards <10 cm x 10 cm can be had for about $2 a pop. Things get pricy if you go bigger.

    The DC Filter
    So, the schematic in the last post won’t really work. The problem is that, at high frequencies, capacitors become inductors, and inductors become capacitors.

    With inductors, things are easy. The figure of merit is the self-resonant frequency (SRF), or the frequency at which the inductance and parasitic capacitance form a resonator. Most datasheets will more or less tell you “past x Hz, you’re on your own!” As you might expect, a lower value L will have a higher SRF. Typically, shielded parts have a lower SRF because of the shield-to-inductor capacitance.

    Caps are trickier. Most manufacturers don’t differentiate between ESR and ESL. Typically, the two are lumped together and called “dissipation factor” (DF, or tanδ). Q describes the same phenomenon, but with a different number. We don’t really care about ESR—it just limits the maximum attenuation we can get from the filter. We do care about ESL, though, because that’s what determines the cap’s SRF.

    With inductors, our only option really is to pick a smaller one. While ceramic caps sound like garbage, they’re more or less purpose-made for bypassing power supplies. NP0/C0G are really the best for this, but don’t come in the sizes and voltages we want. X7R do; however, their capacitance will vary with voltage and applied signal, so it’s important to derate them.

    Film would also work, but probably isn’t better. The ESL of a film cap is typically around 10 nH, because they are physically large. Electrolytics actually aren’t a bad choice. They have a high storage density and can be smaller. Newer Al-poly electrolytics manage as little as 500 pH (!), but also don’t come in the voltages and sizes we need.

    [​IMG]

    [​IMG]

    What’s with R7? An inductor’s inductance will increase just before reaching the SRF, and that could mess up the filter response. With R7 in place, the series impedance is limited to R7 // L1.

    The current source
    Not much to add here. There is no ground plane, because the current source is a floating two-node device; there should be no connection to ground. Adding a ground plane would simply add capacitance and worsen performance.

    [​IMG]

    [​IMG]

    It’s the same size as the DC filter board, so they could be stacked. The TO220 part works with most BJTs or MOSFETs.

    The important part
    The amp board is connected differentially from input to output. What that means is that, in lieu of a ground plane or a star ground, connections are done in such a way that each stage will amplify the difference between two nodes; each device has an input and a reference. There is a single connection to the ground plane, allowing it to act as a shield but not a signal carrier.

    [​IMG]

    [​IMG]

    There are a few options for the input device.

    BF862: High gm, high Vds part. There have been reports on diyAudio about parts from certain fabs having worse 1/f noise performance than others, but it is very quiet

    2SK3557: High gm, but lower Vds. This part is designed to be cascoded. It has better curves, runs hotter, and draws less gate current.

    2SK932: Highest gm of the lot, and also low Vds. Otherwise very similar to the 2SK3557.

    BF861: The BF862's less popular, lower transconductance, higher noise sibling. It seems to be the more linear of the two.

    They should all sound good, but different.

    The RIAA caps are also surface mount. It uses the very lovely Panasonic ECHU PPS caps.

    They are supposedly better than PP, and, unlike most fancy caps, come in 2% tolerance for accurate channel-to-channel RIAA compensation.

    Next time: Stuff gets built.
     
    Last edited: Sep 24, 2016
    uncola, Hrodulf, Jeb and 3 others like this.
  10. uncola

    uncola Friend

    Friend
    Joined:
    Sep 27, 2015
    Likes Received:
    458
    Dislikes Received:
    1
    Trophy Points:
    63
    Location:
    Oahu, Hawaii
    [​IMG]
    Pretentious Designs Phonostage logo prototype ;)
     
    yotacowboy, HitmanFluffy and Smitty like this.
  11. Cspirou

    Cspirou They call me Sparky

    Friend
    Joined:
    Sep 27, 2015
    Likes Received:
    4,950
    Dislikes Received:
    5
    Trophy Points:
    113
    Location:
    Northwest France
    @peef - How did this phono workout for you? Are PCBs available?
     
    Last edited: Dec 12, 2016
  12. peef

    peef Almost "Made"

    Joined:
    Oct 5, 2015
    Likes Received:
    108
    Dislikes Received:
    0
    Trophy Points:
    33
    Oof, it's a work in progress still. Something I didn't consider is that the bias is extremely sensitive due to the current source working into a fairly high impedance node, so it will stick to the rail if you look at it funny.

    I prototyped one with a gyrator load instead of a CCS; no trimming, just not quite as elegant in my view I'll probably update the board over the holidays. The other option is to split the gain between two stages, but what's the sport in that?

    The power supply on the other hand is awesome.
     
    sphinxvc and Cspirou like this.

Share This Page