Hi list
Thinking a bit about my Zynq board having now a working Linux supported AXI connection
with it's FPGA that can be used (thus far simple) examples with Vivado 1015.3, I would
like to plan using the right niche or corner or "edge" when making some examples that use
FPGA acceleration for music digital signal processing I have in mind. The current
connection that works properly and Open Source between the ARM cores and the decent FPGA
isn't very fast (little over 10 Mega Byte per second I suppose) but it's possible to run
quite some data amounts over the AXI bus, like the working (also FOS) HDMI interface
interrupt/DMA example (from the Parallella project/ADI), and possibly special DSP
functions that can make use of the (small) on chip memory bank.
It used to be so that a number of interesting DSP functions, just like other functions
including mathematics on supercomputers and computer graphics, are more memory bandwidth
limited than CPU/DSP power limited, which today is a bit of an obfuscated subject because
you have to take pipelined memory access into account as well as average and worst case
cache performance.
I was wondering what the devices currently in use in various DSP setups have as essential
memory bandwidth. I now my very fast PC has dynamic memory access bandwidth of about 80
Giga byte/sec (4 banks), and big graphics cards with DDR5 memory get what, say a few
hundred GB/s easily enough. Of course graphics cards aren't often used for audio DSP and
in the case of the PCs memory banks, we have to probably count considerable latency for
single, random-bank accesses.
Then there's the dedicated DSPs, I know about some fast multi-core TI DPSs and have used
some Analog Devices ones, which essentially use dynamic RAM chips for main memory of their
core(s), and there are dedicated ASIC/FPGA chip designs with internal memories up to a
part of or a few megabytes, and some had external fast static memories, but that appears
to be not fashionable.
What's the range of bandwidth people compute to use for serious dSP applications or is
there even an objective measure ? The big dream of computing for instance live FFTs for
studio quality audio is achieved for PCs depending on what you want but where is the
analysis ?
Theo V.
_______________________________________________
dupswapdrop: music-dsp mailing list
music-dsp@music.columbia.edu
https://lists.columbia.edu/mailman/listinfo/music-dsp
