Scribd
Upload
27 views

Documentation FPGA ClassD

This document discusses digital Class-D amplifier design using sigma-delta conversion. It explains that a multibit sigma-delta converter can shift quantization noise out of the audio baseband, allowing high resolution audio with a lower digital clock frequency than would otherwise be required. The TI TAS5701 is presented as an example that likely uses a mixed-signal design and multibit sigma-delta modulation to achieve high quality audio output from a relatively low-frequency digital section. Implementation details are discussed such as upsampling, noise shaping order and parameter selection, and fixed-point arithmetic considerations for an FPGA implementation.

Uploaded by

Lê Xuân Tình
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
27 views

Documentation FPGA ClassD

This document discusses digital Class-D amplifier design using sigma-delta conversion. It explains that a multibit sigma-delta converter can shift quantization noise out of the audio baseband, allowing high resolution audio with a lower digital clock frequency than would otherwise be required. The TI TAS5701 is presented as an example that likely uses a mixed-signal design and multibit sigma-delta modulation to achieve high quality audio output from a relatively low-frequency digital section. Implementation details are discussed such as upsampling, noise shaping order and parameter selection, and fixed-point arithmetic considerations for an FPGA implementation.

Uploaded by

Lê Xuân Tình
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 19

Fully digital Class-D amplifiers

by using a multibit sigma-delta converter (noise-shaper)

https://github.com/YetAnotherElectronicsChannel
https://www.youtube.com/c/YetAnotherElectronicsChannel
Audio PWM modulation theory

Picture from:
https://www.fairaudio.de/lexikon/puls-weiten-modulation/
Audio PWM modulation theory - digital

- Overall concept stays the same


- Drawback: Triangle waveform is implemented as running counter
no perfect triangle waveform
Audio PWM modulation theory - digital

- PWM frequency needed for Class-D power-stage ( ~ 200-500 kHz)

- For good audio quality : PWM-resolution needed >= 16 bits

- Triangle-generator (counter) must run with every PWM period one time up and one
time down (2 * 2^16 steps = 131072 clock-cycles needed for one period)

- For 200 kHz switching frequency, the digital clock-frequency needed is 200000 *
131072 = 26.2 GHz

- 26.2 GHz is impossible to realize on a low-cost FPGA (50 - 100 MHz)


Audio PWM modulation theory - digital

Assuming a 100 MHz clock on FPGA modulating a 200 kHz PWM

-> 100Mhz / 200 kHz = 500 clock cycles available for each PWM period

-> Counter must run one time up and down with 500 clock-cycles -> resolution is ~ 250
steps (approximately 8 Bit)

-> 8 Bit is a very bad audio quality


Example digital Class-D (TI TAS5701)

- Full digital Class-D amplifier with I²S input


- TI most probably uses a mixed power / digital technology to combine
power mosfets and digital logic on the same silicon die
- Drawback: Those technologies can run digital logic comparable to a
low-cost FPGA (100 to max. 150 MHz) -> how is it possible to achieve a
good audio quality?
Example digital Class-D (TI TAS5701)

- Audio chain contains a “SRC” -> Sample-Rate Converter


- Audio chain contains a “4th order Noise Shaper and PWM”
- How is it working in detail? -> Let’s implement it on a FPGA!
Introduction - Sigma Delta Conversion
- Sigma Delta converts an analog signal into a 1-bit bitstream (same signal represented in 1’s and 0’s)
- Output bitstream is based on a clock
- The output signal is ideally the same as the input signal
- Practically the signal quality depends on the “oversampling rate” (baseband-frequency of the signal
vs clocking speed of the modulator) -> the more oversampling, the better the output signal regarding
signal-noise ratio
Introduction - Sigma Delta Conversion
- Reducing the resolution of a signal (bit-width) is introducing “quantization noise” (also in the
baseband of the signal)
- A sigma-delta converter is shifting or “shaping” the quantization-noise in the baseband signal into
higher-frequency ranges if the signal is oversampled -> therefore called “noise-shaper”
- As the noise is shifted out of the baseband-signal, it’s not hearable anymore for humans
Multibit Sigma-Delta / Noiseshaper
- The output-comparator (only 1 bit resolution) of the sigma-delta converter is replaced
now by a x-Bit “quantizer”
- The output data-width is reduced from 24-bit to e.g. 4 bits
- Signal quality on the output ideally the same as on the input (depends on
oversampling-ratio)
- Multiple integrators can be connected in series, reducing the amount of oversampling for
the same signal quality (TI TAS5701 example = 4th order = 4 integrators)
High-res audio PWM modulation

- The 96 kHz audio signal is converted to fs=384 kHz with a SRC


- The noise-shaper reduces the bit-width of the audio-signal from 24 bit to 5
bit (with almost the same audio quality as on the input)
- Oversampling ratio is = 384 kHz / (2* f_baseband ~40 kHz) = 9,6 ~ 10
- PWM modulator needs (2 * 2^5 * 384k clock cycles = 24,576 MHz)
- 24,6 MHz is realizable on a cheap FPGA or mixed-signal silicon technology
:-)
High-res audio PWM modulation

Remember: This is only an example how I implemented it

- Improving audio quality is possible by e.g. running the clock with 4* 24.6
MHz and using a 7 Bit PWM modulator instead
- But the iCE 40 FPGA on my TinyFPGA-BX board is not really capable of
running higher then approx. 40 MHz for math-operations with > 24 bit
values (improvements with more expensive FPGAs possible)
- My humble information I have about the TI device is, that it even works
internally with an 8-Bit PWM (not confirmed 100%)
Image from:
https://forums.cockos.com/showthread.php?t=215201

Upsampling (SRC)

- Between every incoming sample, 3 zero-samples are introduced and finally


low-pass filtered by a FIR filter with cut-frequency at ~ 20 kHz
- This process reduces the signal-energy by the upsampling-factor (4)
- Incoming samples must be multiplied by 4 before to keep the same
audio-level at the output
Upsampling (SRC)
- In my implementation, the FIR filter has 124 coefficients, but as 75% of all
samples are 0, the zero-samples must not be processed actually
- => effectively 124/4 = 31 multiplication- and summing operations per
incoming sample needed (with a 24,6 MHz clock, there are 64 clock-cycles
available for each of the incoming 384 kHz samples)
4th order Noise Shaper (CIFB structure)

Parameter c is always 1
4th order Noise Shaper

- Calculation of all parameters ( b1-b5, a1-a4, g1,g2) is very critical to obtain optimal
signal/noise performance and being stable
- Tools / Documentation available for determining the parameter-set
- Matlab Delta-Sigma Toolbox (Link)
- Delta-Sigma Toolbox by University of Ulm (Link)
- Application Notes on the Web (Link)
- In my FPGA example, I even hand-adjusted the parameters of the g1 and g2 resonators for
improving audio-quality
- I wasn’t lucky with the Matlab-tool / tool from University of Ulm
- Parameter-set of University of Ulm was always unstable (distorted noise on the output)
- Parameter-set of Matlab-tool was stable, but signal/noise performance was terrible
- I used the parameter-set of the app-note above for b1-b5 and a1-a4 ... g1+g2 I was adjusting “by hand” via a SPI interface
in the FPGA for manipulating the parameters in realtime (not present anymore in my published GitHub project now)
Actual FPGA implementation

MCLK = 25 MHz / 2 = 12.5 MHz


Real clocking & BCLK = 25 MHz / 4 = 6,25 MHz
frequencies: WS = 25 MHz / 256 = 97.6 kHz
PWM = 25 MHz / 64 = 390 kHz
Implementation hints
- Your system MUST run fully synchronous
- E.g. the 64 counter steps of the PWM modulator must be in sync with the 384 kHz
audio-stream from the noise-shaper
- Therefore not implementable on a µC (e.g. STM32)
- Upsampling by a factor power-of-2 is easy to implement because of
multiplication of power-of-2 by just bit-shifting to the left
- PWM frequency is therefore also a power-of-2 multiple of the input sample
rate (check out switching frequency of the TI TAS5701 device)
Implementation hints

- I used fixed-point arithmetic for all math-operations on the FPGA


- Check out my video about IIR filters on FPGA where I explained how
fixed-point arithmetic on a FPGA is working
https://www.youtube.com/watch?v=eE6Qwv997cs

https://github.com/YetAnotherElectronicsChannel/FPGA-Audio-IIR

You might also like