Low Latency Audio Transport and Processing with standard PC Hardware and Custom FPGA Hardware

Type: Master thesis
Student: Fabian Schenkel
Advisors: Attila Karamustafaoglu, Mårten Sterngren, Clemens Lombriser

In the last decade, analogue audio mixing consoles have gradually been
replaced by their digital equivalents. Digital audio processing leads
to a higher signal-to-noise ratio, allows more complex processing
algorithms, and can handle larger scaled systems. In most professional
digital mixing consoles, Digital Signal Processors (DSP) are used to
perform the audio processing tasks. DSPs feature high-performance
processing and induce low audio latency, which is a key factor
especially in live applications.

While digital audio offers increased possibilities in audio
processing, the customer also asks for additional features and more
processing power for new digital mixing systems. To satisfy this
growing demand, one possibility is to implement the increased
requirements by using higher performance DSPs. However, it seems that
this approach is not feasible for the near future. While still valid
for general purpose CPUs, the famous Moore's Law has
lost it's validity for DSPs, therefore higher performance based on
DSPs could only be achieved by increasing the number of DSPs used in
the system. Since it is clear that this can not be done to an infinite
extent, research for other devices than DSPs to perform the audio
processing in a mixing system is required.

The goal of this thesis was to evaluate the use of standard computers
featuring general purpose CPUs as a DSP replacement for audio
processing. The performance of modern CPUs is still constantly growing
according to Moore's Law. Every 18 months, CPUs featuring twice the
performance compared to the old generation are available on the market
and this trend is expected to continue at least for the next
decade. Because of this, CPUs seem to be a highly suitable replacement
for DSPs. CPU based processing also has further advantages, for
example the effort of moving to a new generation processing system is
much lower. Instead of designing new hardware and firmware as required
when changing to a newer DSP family, a new CPU-based system could be
developed by just using a new computer off-the-shelf and modifying the
processing software accordingly. Maintaining the processing software
is also easier since it is written in a high level language
(e.g. C++), which furthermore improves algorithm portability.

However, using a CPU-based processing system exhibits some
disadvantages as well. The audio data to be processed needs to be
transferred into the computer and back with a high throughput
requirement, especially when a large scale system is demanded, but
audio mixing applications allow only very low latencies. Furthermore,
the increasing need for processing power could require systems where
one single computer is not sufficient to do all the processing. The
audio interchange between several computers needed for such a system
would additionally affect the minimal possible latency. Moreover, many
audio mixing applications require a fail-safe system. Full
availability of the system can only be guaranteed by introducing
redundancy into the system, which would require special mechanisms
handling a system failure in a CPU-based system.

As an evaluation, the audio processing mechanism of a state-of-the-art
digital audio mixing system from Studer based on DSPs has been replaced
by a standard computer in this thesis. To achieve this goal, the
corresponding I/O system has been modified to allow audio data
exchange with the computer via Ethernet Layer 2. The modified I/O
system has been connected to a computer with a standard network
interface card, where the audio data has been received, processed, and
returned. The processing software has been designed to be controllable
from the mixing desk by emulating a DSP processing core. Measurements
have been made to evaluate the audio processing capabilities of a
modern multi-core processor with special attention to the achieved
overall audio latency. As a result, the limitations and opportunities
provided by a computer based audio processing solution have been
shown.