A Case Study: Using FPGA for Audio DSP
When we started working on the concept of a pedal effect based on a Shepard glissando, one of the first things we needed to decide was how many filters would be necessary in its parallel filter bank. This vision, fully realized in Spira, led us to design for up to a maximum of 24 bandpass filters and a maximum slope of 18 dB per octave.
That’s a lot of processing, and we knew we needed a platform to deliver that kind of power with minimal latency. So we are excited to be debuting our FPGA signal processing platform, which makes Spira possible.
When we started developing our FPGA platform early in 2016, the advantages were hypothetical and, for us at least, not yet realized in concrete form. Now that we’re on the other side of it, I want to share with you what we’ve learned and what FPGA technology can do for both professional and consumer audio applications.
Xilinx has published data showing FPGA performance metrics over the past several decades.
Granted, this chart misses recent years. But if your assessment of FPGA viability in a production setting is based on what you saw ten or more years ago, it’s probably time for a second look.
FPGA technology has matured to the point where pricing per equivalent amount of work is competitive with current generation DSP chips, and volume pricing is generally available. Much of this has been driven by non-audio industries such as networking, communications, and automotive. FPGA benefits from being more general purpose, despite its reputation for being esoteric. This leaves the door open for niche uses outside the big industries driving this growth.
Most people considering an FPGA solution for audio will currently be working with traditional DSP. The two kinds of platform have a lot in common. Both have convenient SoC features, and for applications about the size and cost of a guitar pedal, have about the same footprint, power consumption and unit cost tradeoffs. Both offer parallelization via embedded multipliers, but the FPGA has vastly greater opportunity for parallelization due to its architecture and its paradigm of implementation.
Spira’s audio pipeline is five audio clock cycles (at 96 kHz) or a total latency of about 50 microseconds. The total measured platform latency from analog in to analog out is 250 us, with the delta-sigma ADC pipeline accounting for the majority of the latency. For the given workload, this is considerably better than even the best traditional DSP products we have evaluated.
Successful development for an FPGA target does require skills, tools, and awareness that are different from DSP environments. However, the size and complexity of the code itself is similar. More sophisticated techniques, specifically better parallelization/pipeline density, will translate to even better performance. But just getting the basics right is already enough to realize an advantage.
FPGA also offers a much more flexible board topology, since the majority of pins are general purpose and assignable. This may be a convenience, or a way to reduce board size, or may help ensure optimal placement of high speed traces, such as for DDR RAM.
The ability to put a soft-core processor on an FPGA is often emphasized, and for some things it’s a great convenience to have access to a traditional microcontroller on the same chip. But we think it would be a mistake to handle any audio there, and Spira uses no such soft-core processor. Spira’s code is written wholly in System Verilog. This leads to greater dev efficiency, better resource utilization, and ensures that all the work performed on the chip can maximize the inherent benefits of the FPGA architecture.
There are some interesting announcements recently for products using FPGA in synthesizers, studio gear, and other larger and more expensive products. That seems like a natural place for the wave to start, in products with higher COGS and overall greater processing demands. But we’ve seen firsthand that the numbers now make sense for smaller products at a lower price point. We are excited about the future and the potential for smaller, portable, and price competitive products like Spira.