Your browser is out-of-date!

Update your browser to view this website correctly. Update my browser now


FPGA: the future is fixed, not floating

Calrec Audio's Patrick Warrington talks to Kevin Hilton about fixed point FPGA circuits and why floating point DSP might not be the all-conquering technology it's held up to be.

There’s nothing like a good eye-catching title, so anyone with an interest in audio technology would have been drawn to a seminar called “Why FPGAs and Fixed Point Architecture are Kicking Floating Point DSP to the Kerb”, writes Kevin Hilton. That was certainly the idea behind Patrick Warrington’s presentation in The Audio Room at last week’s Broadcast Video Expo (BVE), although Calrec’s technical director acknowledges the billing would not be the first choice of any engineer. “It wasn’t my title,” he says, “but it was intended to provoke a bit of controversy.” There is always going to be something contentious about suggesting a technology that has served the audio industry well, and was instrumental in its shift to digital, might not be up to the job any more – and could even be replaced. But Warrington approaches the potentially thorny issue with an engineer’s objectivity, rather than a marketing person’s hyperbole. “I would never suggest FPGA and fixed point would replace floating point DSP wholesale,” he says. “There are applications where using FPGA couldn’t be justified economically but it comes into its own when there are hundreds of channels of audio to be processed.” FPGA (field-programmable gate array) is an integrated circuit (Apollo Bluefin2 board pictured) that can be set up by the product designer or manufacturer to match the exact requirements of a user. In broadcast audio this means dealing with multiple channels – not just surround sound but incoming feeds, alternative languages and communications. Warrington says that when he and the design team at Calrec set about designing the Bluefin high density signal processing system in 2004 they realised floating point DSP was not going to be able to cope. Bluefin was designed initially for the Alpha console, with the aim of allowing it to handle large amounts of surround inputs by doubling the desk’s ability to process information. “It became apparent fairly quickly that what we wanted to do was going to use fixed point rather than floating point,” he says. “We looked at the mathematics and that led us to a bit of a revelation – it was possible to improve on the performance of the benchmark floating point, off-the-shelf DSP cards.” Probably the most popular of these is the SHARC (Super Harvard Architecture Single-Chip Computer), which has been produced by Analog Devices since its development in the early to mid 1990s. It is noted for its good performance and low cost, both of which have made it a firm part of modern computer-based systems. Audio equipment manufacturers have relied on SHARCs for many years but Warrington feels the chip, while not bad technology, has been used as a selling point beyond its capabilities.” “Sometimes product literature holds up the SHARC as a guarantee of good quality audio,” he says. “It isn’t.” Understanding why, he explains, comes down to mathematics. This, for some of us, might be where the whole subject becomes hard to grasp. Essentially the problem lies in the nature of floating point technology; real numbers – or mantissas – can support a range of values by being scaled up using an exponent or multiplier. In digital audio resolution and dynamic range are the key factors. Achieving the best performance for both of these depends on selecting the appropriate number of bits, or word length, to do the job. Floating point, Warrington says, theoretically gives this flexibility; in practice digital audio processing introduces a number of errors, including audible noise, unless there is extra resolution. The key functions that can be affected are gain, mixing and, most important of all, equalisation. “The weakness of floating point is that although it is good for expanding a system it does not increase the resolution,” says Warrington. “There is nothing wrong with floating point but it doesn’t benefit you when you need additional resolution.” Calrec, Fairlight and Digico have all produced circuits for their consoles based on FPGA technology. Warrington says fixed point delivers the necessary capacity and capability to produce the levels of resolution demanded by broadcasters from today’s digital mixing desks. He concedes, however, that this comes at a financial cost to the customer: “FPGA chips are more expensive than SHARCs, so unless someone is dealing with multiple channels it is hard economically to justify using them.” As with all things, the demands of the market and the job will determine what is used – engineers merely provide the options. “People won’t abandon floating point DSP,” concludes Patrick Warrington, “but if someone is processing hundreds of channels then the benefits of FPGAs will force them down that route.”