[sdiy] DSP on GPU?

[karl dalen]

Interesting, thank's for the link.

KD

>ASSI <Stromeko at nexgo.de> 

> It is still non-trivial to use a GPU for audio.  One
> problem is to find 
> enough parallelism while keeping the memory access pattern
> in line with the 
> GPU architecture (get this part wrong and the penalty is a
> three orders of 
> magnitude or worse slowdown) and another one is getting the
> data back out 
> the GPU (the fastest way out actually is through the HDMI,
> but that comes 
> with its own problems).  Latency is a problem,
> especially if some data goes 
> the "wrong" direction (i.e. back to the CPU).  Last
> but not least, none of 
> the articles talks about efficiency - the GPU that have
> general enough 
> programming facilities are very power hungry and the
> low-power GPU that 
> exist today are not very programmable for things other than
> graphics.
> 
> Just in, some experiments on the most general GPU
> architecture avilable 
> today (Nvidia Fermi, on a GTX480):
> 
> http://www.aes.org/journal/online/JAES_V59/1_2/
> 
> "As a result we conclude that GPUs are well suited for
> audio applications that need heavy computation, can be
> parallelized, and can tolerate some latency. It is
> essential
> to get enough concurrent threads running so that the GPU
> is fully utilized. In practice this often means increased
> buffer sizes and thus increased latencies. However, in
> many practical cases this cost is acceptable, since the
> latency is only a few milliseconds. Now that it is
> possible
> to use GPU hardware to do audio signal processing, more
> real-time performance is available than ever before."