Parallella Community

I was thinking about using my Parallella system for audio processing when it arrives,
and thought about using, N cores of the 16 cores for processing audio sample streams.

For N>1, that leads to the possible solution of routing audio on the Epiphany chip,
between cores, for serial or parallel processing.

My question is: what is the achievable bandwidth and latency in practice,
for such a basic scheme of stream processing? I read there's a form of
"broadcast in parallel" as a hidden option, which is great, but is it possible
to do a sort of "systolic" transfer, in parallel ? I mean can the processors
be programmed to have fixed communication patterns In and Out, going on
continuously, and simultaneously ?

Thanks in advance

T. Verelst

Theo, the Epiphany chip is very well suited to the kind of systolic processing you are thinking of. The basic technique is for a core to write intermediate results into the internal RAM of the next core in the pipeline. This can be done in increments of as little as 1 byte (though 4 or 8 bytes makes the most efficient use of mesh network bandwidth), and with a latency of 1.5 cycles if the communication is nearest-neighbour.

In the real world things aren't quite that simple because some sort of coordination between cores is needed to prevent buffer overruns. You'd want to use something like the eport primitive invented by forum member notzed and described here: http://a-hackers-craic.blogspot.com.au/ ... again.html

I've been playing with different primitives for computational mesh applications, so while I don't know what the latency of an eport write might be, it's possible to hazard a guess that it could be around 50ns (30 cycles).

While peak core-to-core bandwidth is in excess of 1.4GB/s, the coordination overhead dominates the bandwidth of small transfers. Using small fixed-size blocks, sized and aligned in multiples of 8 bytes, should result in the optimal balance of bandwidth and latency because the overhead code remains simple and all transfers can be in the "natural" units of the e-mesh.

The external RAM is something of a bottle-neck - about 300MB/s I believe - would that be adequate for the number of audio streams you are contemplating?