they make this seem like a huge deal, anyone know why? is this going to like put HD content on an iPhone? I don't expect that the signal processing hardware is the bottleneck here.
> I don't expect that the signal processing hardware is the bottleneck here.
Computational resources is always the bottleneck in bioinformatics, quantum chemistry, or any sort of high data volume analysis or simulation, and the FFT is a fundamental and commonly used transform in all fields.
At least for people who still use computers to, well, compute things, a faster FFT is a huge deal.
How many domains still use FFT specifically though as opposed to some other transform? Most of the signal processing papers of the last decade that I've read or read about have used either wavelets, FFT on a small window such that this isn't really applicable, or some arbitrary non-orthogonal basis.
When I was doing EEG signal analysis in grad school, I used the FFT all the time. While there's some cool stuff involving wavelets and using interesting basis sets (matching pursuit looks cool), if you're primarily looking at power and frequency over time, the FFT is sufficient, and usually faster than the other algorithms. (And if you're looking at power/phase, the common Morlet wavelet choice is mathematically equivalent to an FT with a Gaussian taper.)
I'm not sure what you mean by "on a small window such that this isn't really applicable"; can you give an example? As long as you accept the inherent time-frequency resolution trade-offs, there's no obstacle to using FFT on a small window. It's called the short-time Fourier transform (STFT), and it's used everywhere; it's probably used more than analyzing an entire signal, since we frequently want to know how power and phase change over time in a signal, and a full-signal FT can't tell you that.
>I'm not sure what you mean by "on a small window such that this isn't really applicable"
FFT is very useful on a small window, but algorithms that improve the asymptotic efficiency are unlikely to be useful at that scale. With n=100 the asymptote doesn't matter. We'll still be using the FFT forever, I'm just skeptical that the frontiers of technology will be advanced by a faster FFT, since it seems like the coolest stuff is happening elsewhere. It's increasingly becoming the quick-and-dirty counterpart to the sophisticated-but-slow methods.
If the hardware isn't the bottleneck, you can simply make your compression/encoding algorithms more complex, because 10% saved traffic is probably much more valuable than 10% saved CPU cycles (remember on phones traffic means a lot of energy drainage for sending/receiving). Besides that, though, FFT is old, so whatever codec you use to view HD videos on the iPhone, it probably already uses FFT or some other comparable thing like wavelets or so.
Computational resources is always the bottleneck in bioinformatics, quantum chemistry, or any sort of high data volume analysis or simulation, and the FFT is a fundamental and commonly used transform in all fields.
At least for people who still use computers to, well, compute things, a faster FFT is a huge deal.