I've been a silent observer for some time now, but since GPU processing is 'close to my heart', I thought I'd jump in... So there goes my first post in the pd-list...

In general, GPUs are really beneficial for parallelisable algorithms involving heavy-computations, such as FFTs, fast convolution, BLAS with huge matrices, finite difference modelling etc... To maximise performance, the GPU kernel needs to unanimously operate over a large enough data set which needs to be copied into the device's memory, as GPUs generally can't access the host memory. This would mean large buffers --> increased latency. So doing 'real-time' DSP on a GPU would probably make more sense for stuff like physical modelling, complex additive synthesis etc..., rather than to 'generally reduce the system latency'.

*However*, if the SoC platform physically shares memory between the GPU and the CPU, then this could, in theory, help reduce the inherent latency (as no memory transfers would be required), but without having detailed documentation, this would be difficult to assess. 

Cheers,

Jon.