This is a follow up post in via cheater cheater:
"I just wanted to let you know that the Parallella kickstarter which I wrote to you about last month (http://m.matrixsynth.com/2012/09/parallella-supercomputer-for-everyone.html) is going strong. The project has broken 50% of the required goal; with 10 days to go it just might work out. There have been a lot of updates, and they showed some DSP demos, which directly apply to performance in music applications:
http://www.kickstarter.com/projects/adapteva/parallella-a-supercomputer-for-everyone/posts?page=2
Demo #1: Using a Scalable Parallel 2D FFT for Image Enhancement
http://www.adapteva.com/white-papers/using-a-scalable-parallel-2d-fft-for-image-enhancement/
- FFT, that's reverb, equalizers and cabinet simulators for you! This demo uses the much more computationally intensive two-dimensional FFT for images.
Demo #2: Scalable Parallel Multiplication of Big Matrices
http://www.adapteva.com/white-papers/scalable-parallel-multiplication-of-big-matrices/
- matrix multiplication is a staple in many advanced DSP techniques. Among others, analog component modeling can be implemented with it (think Virtual Analog synths the kinds of FXPansion's Synth Squad/ D-CAM, Arturia, or the u-he Zebra/ACE/Diva)
Demo #3: Approaching Peak Theoretical Performance with Standard C
http://www.adapteva.com/white-papers/approaching-peak-theoretical-performance-with-standard-c/
- what this says is that, MHz for MHz, nothing can be faster than the Parallella. It's fairly amazing given that, when comparing clock speeds, the thing is about 5x faster than desktop CPUs.
There are also new pledge rewards and pledge levels. A lot of comments have been posted and the discussion is very lively: http://www.kickstarter.com/projects/adapteva/parallella-a-supercomputer-for-everyone/comments
It looks like the project is on the verge of getting funded - but it'll need any bit of help - so if anyone reading this is interested in the project, consider backing it!"
Wednesday, October 17, 2012
2 comments:
To reduce spam, comments for posts older than one week are not displayed until approved, usually same day. Do not insult people. For items for sale, do not ask if it is still available. Check the auction link and search for the item. Auctions are from various sellers and expire over time. Posts remain for the pics and historical purposes. This site is meant to be a daily snapshot of some of what was out there in the world of synths.
PREVIOUS PAGE
NEXT PAGE
HOME
© Matrixsynth - All posts are presented here for informative, historical and educative purposes as applicable within fair use.
MATRIXSYNTH is supported by affiliate links that use cookies to track clickthroughs and sales. See the privacy policy for details.
MATRIXSYNTH - EVERYTHING SYNTH
© Matrixsynth - All posts are presented here for informative, historical and educative purposes as applicable within fair use.
MATRIXSYNTH is supported by affiliate links that use cookies to track clickthroughs and sales. See the privacy policy for details.
MATRIXSYNTH - EVERYTHING SYNTH
Subscribe to:
Post Comments (Atom)
I can see a few architectural niggles where DSP is concerned. For one thing, it only supports single-precision floating point operations, which will lead to error accumulations (in IIR filters, for example); on the integer side, it doesn't appear to support saturated arithmetic or, more problematically, a mixed-precision multiply (32x32->64 bits) or a fractional multiply. Working round those constraints isn't going to be impossible, but it will take more instructions - to the point where an operation that would take a single cycle on a DSP (or a chip with DSP-optimised instructions) could end up taking as many as a dozen cycles on one of these cores. I'm not saying it'll prevent some interesting DSP uses, but it is something to factor in when comparing this to a conventional DSP.
ReplyDelete(Having said that, the only reason I'm not backing it is that I'm going to have too much month left at the end of this money.)
Hey Gwen,
ReplyDeleteI believe they're working on bringing 64-bit FP to a future iteration. However, this is more geared for operations that use loads and loads of parallelizable data (e.g. volterra kernels) rather than stuff that a desktop cpu can do anyways.
Don't forget that most hardware VA synths run on 32-bit DSP chips anyways. Using 64-bit is not necessary if you're careful about your algorithm's convergence and stability. Anything from the Waldorf Q+ to the Virus C and Nord Lead 3 uses 32-bit arithmetic, iirc. Those synths don't exactly sound too bad :)
The good thing is that it's an open platform - and, better yet, you have a live wire to the developers - so suggestions like this will have an easier time coming through. I don't think that's ever been possible in the past or will ever be possible in the future.