nVidia CUDA GPU computing PC FIR

[mace1337]

Do you think this can be implemented in OpenCL? So that it can run on ATI and Nvidia cards? (they both support OpenCL now afaik)

[stegard]

Originally Posted by novec

FIR needs a certain number of samples to do its thing, so it will introduce some delay. Lowering the number of samples will lower the delay, but at the cost of a less accurate filtering.

I see, any idea on how to change the number of samples?

[novec]

Judging by the previous posts it seems doable, but I have no experience with it.

[critofur]

Hello,
Could anybody tell me if any of the small integraded nVidia "Ion" MicroITX solutions might be powerful enough for at least a stereo 2-way realtime FIR crossover?

Have Koon's developments been made compatible with Foobar 2000 or Jack?

(I found the nVidia 460 GTX for $159, I'm sure that is more than powerful enough, but I'd like a small all-in-one-box efficient micro-PC solution that would be palatable to speaker customers, and it seems that most currently available inexpensive DSP solutions are far from powerful enough to do things such as phase correction along with a realtime multi-way crossover)

[sdims]

Here the comparison of low-end CUDA cards, but in Russian:
??? NVIDIA / ?????????????? / ??????????? ?????????? GeForce8 ? ????? ????? / ????????? GeForce*8600*GT, GeForce*8600*GS, GeForce*8500*GT, GeForce*8400*GS*rev.2, GeForce*840

We have numbers of single-precision perfomance in 9th row from the top, begining with number of 113,3 GFlops for 8600GT.
As you can read at koonlab.com article, 8400GT is capable to

So even GeForce 8400 can reach 8-9 G TAPS calculation / sec.

and that for 4-way 44100 KHz we need

44100 sample/sec * 2 (L,R) * 4 (way) * 8192(Taps) = 2.9 G tap calculation / sec

So, perfomance of the ION2 must be enough for 44100 Khz 4-way crossover with 8192 TAPs.

But I don't think that ION2 is a good idea, for 2 reasons:

1. Small ION2 boards pricey against small LGA775 boards with Intel chipsets.
2. Digital filters in most of the modern monolithic delta-sigma DAC's is a little bit crappy in comparison with old standalone digital filters such as PMD-100/NCP5842/CXD1244 or in comparison with CPU resampling in SoX resampler of Foobar, so, the more amount of resampling job you do on CPU with SoX instead of built-in delta-sigma digital filters - the more precise sound you will have.

Just my opinion, that far better and precise results must be in next chain:
1. Resampling on CPU with SoX to 192/24
2. After that - CUDA FIR filtering to 3-4 ways

So, amount of processing data will increase strongly, I think that you need for this
- LGA775 Mini-ITX board with at least 2 PCI-e slots, one for soundcard, one for CUDA-card,
- LGA Core2Duo 45nm,
- and GTS450 of better card
- and of course - agreement with software author.

Try to look at Palit PCI-E NV GTS450 1024Mb 128b DDR5 783 / 3608 Low Profile. I don't know anything about how noisy this card, but it the only known to me low profile GTS 450 card with external blower.

For example, my Audigy2ZS sounds like **** without SoX, but with SoX&ASIO in Foobar I can listen to this sometimes. I know, that part of **** is that 2ZS only 48 kHz clock, but I also compare SoX resampling @192 kHz on Asus HDAV 1.3 Deluxe with BB DACs and X-Fi Elite Pro with CS4398 and rather good hardware upsamling in X-Fi chip. SoX was always more precise, clear and more dynamic, but it reveal any loosy stage in system - post DAC circuits, preamps, power amps, speakers. If any stage has compromises and flaws - with SoX sound will be "cold" and "hard", but in reality SoX sounds just like other system sounds, without any dullnes from built-in digital filters of delta-sigma DACs.
Again, this is just my opinion and taste, but it seems like: no gigaflops - no precision & dynamics.

[terranigma]

My soundcard accepts only 48khz multiplier sampling rates after clock oscillator modification. I'm resampling my tracks via SoX on the fly. Will this issue cause a problem for using GPU FIR?

[SoNic_real_one]

Originally Posted by sdims

Digital filters in most of the modern monolithic delta-sigma DAC's is a little bit crappy in comparison with old standalone digital filters such as PMD-100/NCP5842/CXD1244 or in comparison with CPU resampling in SoX resampler of Foobar, so, the more amount of resampling job you do on CPU with SoX instead of built-in delta-sigma digital filters - the more precise sound you will have.

+1 to that...