[sdiy] Hardware convolution box?

[Neil Johnson]

If you can work with the slightly odd floating point format in single
precision GLES shader language then the shader engines in the Raspberry Pi
(1, 2 & 3) will give you about 24 GFLOPS.

Neil


On 13 Feb 2017 20:38, "cheater00 cheater00" <cheater00 at gmail.com> wrote:

On Mon, Feb 13, 2017 at 8:55 PM,  <paula at synth.net> wrote:
> I looked at starting the Chameleon II project with Terry Shultz, but sadly
> this was around the time freescale started pulling out of the DSP world.
>
> Your choices for "DSP" now are to look at something like the Analog
Devices
> Blackfin/Sharc family, look at FPGA (who wants to roll their own DSP) or a
> fast ARM Processor (like the Axoloti).
>
> I guess the problem is that ultimately it comes down to the market
appetite
> for such products. The Chameleon was great, but never really got the
market
> it needed to survive. the Axoloti looks fab, but I doubt it's being made
in
> serious volume (>1000 per year).
>
> What is nice is to see a resurgence of love for DSP and DSP programming :)
>
> Paula

I think just the first three uses are going to be good enough to find
enough market share between them to warrant a board run. Many people
already have lots of stomp boxes and they will pay $2500 to get an
Axe-FX and use it only for cab simulation... and then they only get a
few msec worth of impulse response time budget... that's fairly weak,
and it's not difficult to beat.

But either way if we can agree upon some sort of platform we can at
least go and buy similar dev boards and start plugging away at it. I
think the TI TMS320C is the best platform because the upper end is
very high and you can start with very cheap chips; the most powerful
board from them is $600, and it's an amazing powerhouse.

If it turns out we have a nice platform going that we could share with
the world we can think about spinning up some boards.

As Veronica points out the most difficult part is to set up a company
and stay afloat. So don't. Develop for dev boards, we can afford them,
and it's for us, so we don't care about mass market appeal. Rather
than set up a company that'll stock this stuff, go the Arduino route,
and create a generic platform anyone can "make" with. A board like
this could have appeal for computer vision, robots, quad copters, sdr,
etc. This should keep enough people interested who would pay for e.g.
group buys.

The #1 objective should be that anyone who hasn't done any embedded
programming before should be able to get into this within days,
especially without hardware. It would be a great idea to have
something like a qemu based emulator, even if it doesn't work at full
speed. An even better idea is something that simply runs in the
browser. This way anyone who wants to keep the platform alive, can.

It would be a good idea to keep this future minded as 5 years from now
nVidia will have entered the dsp business full-scale and their chips
will kill anything that's out there. I spoke to John Carmack and he
said they are certainly going to do this. Tim Sweeney confirmed that
the GPUs as they are now are 10x more powerful per dollar as the best
of the best DSPs in my survey earlier on in this thread. TI and AD
have no contender in the GPU based dsp space so they will lose out in
the end; nVidia will dominate the market for automotive applications.
They've already started building self-driving-car embedded modules
with their GPUs on them, and eventually they'll do their own silicon.
So it would possibly be a good idea to keep the algorithms modular, so
that the concrete implementation of anything that needs to be in
assembler can be replaced in the future. This would also prevent
"platform death".
_______________________________________________
Synth-diy mailing list
Synth-diy at synth-diy.org
http://synth-diy.org/mailman/listinfo/synth-diy
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://synth-diy.org/pipermail/synth-diy/attachments/20170213/2efd3087/attachment.htm>