[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Discuss-gnuradio] Full MA paper available
From: |
Vincenzo Pellegrini |
Subject: |
[Discuss-gnuradio] Full MA paper available |
Date: |
Thu, 15 Jul 2010 18:04:14 +0200 |
As one must keep his own promises (and I promised indeed a few months ago),here is the link to the full MA paper
available for anyone who might be interested in trying to apply these ideas to SDR, especially within the typical GNURadio, GPP-based processing paradigm.
Just two very small "release notes"
1. MA Design is not yet a "machine task", i.e. no MA-compiler can be written so far. Still we believe this will be possible.
The MA design section called RTAR is indeed already algorithmic, while what we call AS, remains for the moment a
human design task. In our opinion, by fixing algorithmic rules for AS, something like an MA-compiler could be rather
reasonably written.
2. Still, our current priority is not to make MA "automatic", as this will surely be done in the future, in case MA proves to be an
effective technique for SDR implementation.
Our aim is instead to prove such effectiveness by showing how big achievable MA acceleration factors may be.
I.e. how significantly an SDR might benefit from memory-implementation of some of its critical parts, as long as this
can strongly increase performance (energy-efficiency) without causing "losses in system reconfigurability and generality",
thus preserving the most valuable peculiarity of an SDR.
So far, we've boosted our implementations through MA by something slightly bigger than a factor 10. We obtained this with
some months of work, and by applying the concept of "memory space specialization" as described within the attached
paper, on an ultra cheap (150 € Intel Q9400) standard CPU.
We strongly believe that:
A. Obtained acceleration factors can still grow by further applying techniques described within the paper on ordinary
commercial GPPs
B. Much larger acceleration factors are available if we use computational back-ends being designed ad-hoc for
memory based techniques like MA. (i.e. prioritizing memory access wrt pure computation)
Our research effort is currently concentrated on proving such two concepts to be right.
Upon success, we believe interesting possibilities could open up for SDR as a radio-implementation technology, eventually
leading to substantial broadening of its application spectrum.
thanks for interest to those interested ;-) ,
apologies to those that are not.
Best Regards to Everybody
vince
PS.
small repost of MA-related video demos / conference presentations, just in case:
[Prev in Thread] |
Current Thread |
[Next in Thread] |
- [Discuss-gnuradio] Full MA paper available,
Vincenzo Pellegrini <=