Are IIR filters are worth it, given the fact that FIR filters can use SIMD very efficiently with good implementation? Latency isn't that much of a problem. 64 tap filter (2x) is like 1.5ms roundtrip and is pretty much negligible, unless there are lots of oversampled plugins in a chain. Also, going past 128 taps for oversampling is way beyond diminishing returns, so might aswell not bother. High order IIRs are prone to numerical issues and are way harder for CPUs, because they include long dependency chains in it's computations. FIRs don't have that problem at all IF implemented correctly. Ah, and yes, min-phase FIRs are possible as well, and their implementation is pretty much identical other than the fact that the kernel must be min-phased. But latency benefits are not all that great. Also, multistage oversampling is kinda not worth it, single stage is simpler to design and performance is pretty much the same if not better.
I love your content and energy. Can't wait to see more!
Is this presentation available for download?
GRRRRRRRREAT Stuff, Sam. Thanks!!
Got questions? Leave them here and I’ll do my best to answer!
Are IIR filters are worth it, given the fact that FIR filters can use SIMD very efficiently with good implementation?
Latency isn't that much of a problem. 64 tap filter (2x) is like 1.5ms roundtrip and is pretty much negligible, unless there are lots of oversampled plugins in a chain. Also, going past 128 taps for oversampling is way beyond diminishing returns, so might aswell not bother.
High order IIRs are prone to numerical issues and are way harder for CPUs, because they include long dependency chains in it's computations. FIRs don't have that problem at all IF implemented correctly.
Ah, and yes, min-phase FIRs are possible as well, and their implementation is pretty much identical other than the fact that the kernel must be min-phased. But latency benefits are not all that great.
Also, multistage oversampling is kinda not worth it, single stage is simpler to design and performance is pretty much the same if not better.
I’d like to remind the author that oversampling and upsampling are different concepts