The growing portable speaker segment is resulting in some solid Class D innovation from silicon vendors. A new part from Infineon is especially interesting if your focus is battery life. Class D offers great efficiency at high power, usually exceeding 90% at the max rated power. But as you drop to 1/100th or 1/1000th of rated power, typical class D amps can fall to 50% efficiency or worse. For professional sound reinforcement, this isn't much of an issue. But for battery powered speakers, efficiency at low levels becomes important if you want to provide the 10-20 dB of headroom expected in higher quality products. And make no mistake--this segment is huge with Forbes reporting Smart Speakers growing at nearly 50% YoY. In the US, this translates to about 60M Amazon Echo products per year (the rest being 30M non-Amazon products). The silicon vendors are likely pulling out all the stops to get the win in the next products from Amazon and Google because the rewards are huge.
Infineon's latest offering in this space is the MA12070. The MA12070 uses multi-level modulation to try and improve efficiency at lower power levels. In automatic mode, the transition between the modes of operation occurs seamlessly as input levels change--there's nothing special you need to do to benefit from the new architecture.
The multi-level works by adjusting the rail voltage that is feeding the output bridge. Normally, you provide a rail voltage to the bridge and the output MOSFETs switch between ground and that rail voltage. If you have a 50V rail voltage, then a full bridge can drive +50V across the speaker or -50V across the speaker. That would be the peak of your sine, and the RMS would be 50/sqrt(2) = 35.3V. That means from a single 50V supply you can optimistically push 311W into 4 ohms.
At high power levels, this works well and the efficiency is usually stellar: north of 90%. But at lower power levels, your switching losses tend to dominate. If you pushing 300W at the peak, and if your average is 100X below that, it means your average power is just 3 watts, with the transients in the music program hitting 300W.
Now, the transients don't occur often--most of the time you are averaging 3W. But the output rail is still slamming +/-50V even during the quiet passages. And those excursions are having to charge and discharge stray on-chip capacitances resulting in wasted power.
In the MA12070 during periods when less output power is needed, the MA12070 generates a new rail voltage that is half of the supplied voltage. It does this with a "flying cap." Switched capacitor (aka flying cap) DCDC circuits are very potent for doing integer voltage conversions (that is, multiplying or dividing by integers). Taking 5V and generating 2.5V, -5, or +10V via switched cap topology is pretty straightforward (but 3V, -6 or +11 would get messy). A second rail also gives the silicon maker the option to build a second bridge from smaller transistors (half the rated voltage and current) with far less stray capacitance.
If you cut the rail in half, then you cut your max power out by 1/4th. So, a 20V rail could generate 50W into 4 ohms, and a 10V rail could generate 12.5W. Switched cap voltage converters are generally good for a few hundred mA, and the voltage divider topology does give you 2X more current for free. Building a flying cap voltage divider with 400 or 500 mA out seems readily doable today as existing parts today are doing 200 mA. But to build a flying cap voltage divider with the current needed to hit 10W seems a bit of stretch. But maybe that's part of Infineon's secret sauce--it would take some deep digging into the part to understand exactly what they've done. For now, let's just assume they are smart people and have done their homework and made all the appropriate trade-offs between losses in the bridge and losses in the PVDD/2 rail generation.
TI takes another approach to all this with their "Class D Boosted" architecture. Instead of generating a rail at half the input voltage, TI uses an inductor + DCDC to double (or more) the supply voltage. So, Infineon uses a capacitor + DCDC to to generate a 0.5V rail and TI uses a inductor + DCDC to generate an arbitrary higher voltage rail.
How is the efficiency of the MA12070 and a state-of-the-art Class D part such as the TPA3255 that doesn't employ these fancy tricks? Below is a plot showing from the respective data sheets showing the efficiency curves of the MA12070 and a TPA3255. The MA12070 delivers 2x80W @ 10% distortion into 4 ohms peak output power, or 160W total. The TPA3255 delivers 2x315 @10% distortion in to 4 ohms. Thermally, the TPA3255 is designed to be punished and it can push the near the rated power levels for some time. The MA12070 treats the peak power as true peak power and expects to be far away from the peak most of the time. And this makes sense, they are different animals.
In the plot, the red dashed lines indicates a decade reduction in power from the previous. The MA12070 shows about 80W max output, and thus a decade below that would be 8W and we can see the efficiency is about 73%. On the TPA3255, a decade below max power the efficiency is around 80%.
At two decades below max, the MA12070 is about 60% and the TPA3255 is about 60%. And at 3 decades below max the MA12070 is about 17% and the TPA3255 is about 17%.
So, at first blush when compared to a very popular high-power part that uses a simple full-bridge that is always swinging rail to rail, the difference seems negligible. But improvements will likely come as Infineon learns to better harness the technology.
While the difference might not seem huge, remember that Infineon has achieved this at lower power levels--levels where the quiescent current usually dominate. In this regard, and considering the idle current of the MA12070 is under 20 mA @ 18V input while driving a -40 dBV signal into 4 ohms, what Infineon has achieved is impressive considering the output power that can be achieved.
Measuring the Noise
The MA12070 is filterless, save for some modest EMI control that might be added by the engineer. The cost savings and size savings from Infineon's design is substantial, and it also reduces the load sensitivity present on all class D amps that require LC output filters.
But it does pose a challenge when measuring into a resistive 4 or 8 ohm load. Absent some inductance to slow the switching slew rate across the load resistor, the amount of switching noise induced on the output is enormous and that makes measurement difficult. The solution, according to Infineon, is to use use 22 uH inductors in series with the output when making measurements. This will tame the fast edges, and make the signal more palatable to the signal analyzer. Of course, when driving into a real speaker, you'll get the inductance for free which helps to keep your EMI edges under control.
The plot below shows the inductors placed in series with the MA12070 evaluation board.
In addition to the series inductors, Infineon is using the class AES17 brick wall filter, as well as the AUX-0025 input filter. The AES17 filter needs to deliver 60 dB of attenuation at 24 kHz while providing no more than 0.1 dB of ripple at 20 kHz--a very tall order! The AUX-0025 filter is more focused on attenuating the switching action of the modulator, with 50 dB of attenuation at 250 kHz (see Designing Audio Power Amplifiers, page 756, for an excellent treatment of how difficult these filter and the measurements of Class D can be).
Absent these specialized (and expensive) filters, you need to be very aware of how your measurement equipment will react to the out of the band energy.
Below, let's take a look at how the switching products are attenuated as the signal (-60 dBV input to the MA12070) flows from the MA12070, through the series L, through the QA450's output LPF, and finally through the front-end of the QA401. The plot below shows the progression, with time scales changed to highlight portions of the waveform. Note that grounding was done through the standard ground clip to the front panel. A much shorter ground wire directly to the ground plane near the signal of interest would likely yield further improvement. The scope has 300 MHz of bandwidth, 2GS sample rate and no bandwidth limiting was used.
The first plot (#1) shows the MA12070 before the series L. Note that peaks are nominally a few volts, but can hit 10V. After the series L (#2), the peaks are hammered down substantially (almost 10X).
Plot #3 shows the output of the QA450. Recall the QA450 has a first order LPF with a corner around 33 kHz at the output. We can see some more substantial reduction (about 20X).
Finally, going into the ADC (measured by probing inside the QA401), the spikes are hammered down again by another factor of 20X.
The last stop for filtering is inside the ADC itself. The AK5397 offers about 100 dB of out-of-band noise rejection. The key, in order for it to work, is to ensure the switching spikes are not pencil-thin pulses of energy (narrow in time, high in amplitude), but instead are broader, more spread out pulses of energy--that means the slew rate has been limited, and that means the input CMRR of the QA401 can help and the AK5397 can help too.
So, what does it look like if we make a noise measurement with the QA450 and QA401 with and without the 22uH inductors?
In the plot below, we can see the yellow (left) trace and the red (right) trace. The left channel has a -60 dBV signal going into the MA12070 board, and with the default 20 dB gain of the amp, the output is -40 dBV. The right channel is muted.
We can see the N+D measurement for both channels. This measurement is all of the energy from 20 to 20 kHz with the signal subtracted. For the left channel, that is -83.7 dBV (A-weighted). This is 65.3 uVrms.
For the right channel, this is -70.9 dBV (AW) which is 285 uVrms. Clearly, the inductor is making a big difference. The noise levels are almost unchanged for both 4 and 8 ohms. But the 13 dB gap is substantial.
The noise spec on the MA12070 is show below:
With the inductors present, we measure 65.3 uVrms, which is very close to the typical spec of 60 uVrms. Given this, we can feel reasonably comfortable moving ahead without the high-order filters in place. But make no mistake--without the recommended 22uH series L, the results are pretty brutal.
When measuring frequency response, the series L pose a problem because at 20 kHz the inductors will measure nearly 3 ohms of impedance, and combined with the load resistors this will result in a substantial roll-off as you can see in the plot below. The red trace has no series L, while the blue trace has the 22 uH series L.
So, for confirming frequency response, if we use the series L we will need to understand the measurement will be knocked down at higher frequencies by a fixed amount. In an automated test environment, this is easy to account for: we'd measure a few golden units, and then compare to a mask.
The plot below shows the frequency response into 4 and 8 ohms without the series L inductors present. The gain shown is the nominal 20 dB of the MA12070 minus the 6 dB attenuation of the QA450. Note the QA450 has about 1.4 dB of rolloff at 20 kHz (due to its output LPF) and so the 1.6 dB rolloff shown at 20 kHz is mostly from the QA450. Again, this is easy to compensate for in a test environment.
At the low end, the 3 dB point is about 14 Hz. This is easily fixed if needed with larger input caps.
And the filterless design of the MA12070 shows no load-dependent frequency sensitivity. In Class D devices with LC output filters, there is an issue at higher frequencies.
The THD+N of the MA12070 with the recommended inductors is shown below:
Figure 9 of the reference board user's manual shows the following (not the X axis is a typo--they obviously mean Power in watts):
The Infineon measurement is stereo for 4 ohms, taken at 18V. The QA401 plot above is left channel only for 4 and 8 ohm loads taken at 18V. You can see at 1W the Infineon measurement is showing about 0.07% THD+N (-63 dB) which is about a 3 dB discrepancy with the QA401 measurement.
The MA12070 is a solid part, with an innovative architecture. The wide efficiencies we've seen before, but the quiescent current consumption and current consumption at low power make this part very interesting for segments where battery life is important. And the flatness into both 4 or 8 ohms is also nice. It'd be nice to see how this part might scale upwards to hit the 200 or 300W at 1% THD+N for a few mS while still doing 20 or 30W average without the need for an external heatsink and external LC. The spec indicates 30W continuous without a heatsink at 22V input. With a bit more headroom, this part could form the guts of a 5.1 or 7.1 home theater amp that is incredibly small.
Next we'll focus on building an automated test for this part.
If you liked the post you just read, please consider signing up for our mailing list at the bottom of the page.