FPGA Central

Hello...

Thanks for reading my question.

I am trying to determine whether a commercial grade mini-ITX style
computer can be used for the DSP job that I describe below. The mini-ITX
has dual Intel cores (Core 2 Duo "Santa Rosa" T9500) using Intel chip sets
(965GM Northbridge and 1CH8M Southbridge) with 4 GB DDR-2 RAM.

We plan to use a Debian LINUX distribution (64 bit).

The DSP project entails handling 48 audio streams (DS0) from two T1
connections. Each of the 48 DS0s presents as 8 bit u-law compandored PCM at
64 kbps. The T1 interface card will have hardware to convert the
compandored PCM to linear 16 bit PCM so we will not need to do that in
DSP.

Twenty-four of the incoming streams will include two frequency bands that
must be separated. We plan to use IIR dual biquads to provide the necessary
filtering (estimated 4 biquads per stream). In addition, we need to
implement an FSK decoder for a 300 bps non-MSK format (200 hertz frequency
separation).

The other twenty-four incoming streams will have a similar dual-biquad
filter as above (estimate 4 biquads per stream). A 300 bps modem will then
be added to the stream.

I am trying to find some benchmark for MFLOPS per biquad. And some idea of
the MFLOPS available from a mini-ITX computer as I described above so that
I can decide whether this is practical or whether I need to plan on using a
DSP daughter card to handle the filtering, etc.

MY QUESTIONS:

1. How many MFLOPS does a dual-biquad require?
2. What is the approximate DSP performance (ie. MFLOPS) of a PC style
computer as I described above?
3. Should I consider an off-the-shelf DSP PCI board? Can you recommend
one? Or are there other options such as a USB connected external DSP
engine?

Many thanks for any insights you may be able to share.


Regards,
Larry

On Dec 10, 8:55*am, "lhicks" <lhi...@pantelintl.com> wrote:
> Hello...
>
> Thanks for reading my question.
>
> I am trying to determine whether a commercial grade mini-ITX style
> computer can be used for the DSP job that I describe below. The mini-ITX
> has dual Intel cores (Core 2 Duo "Santa Rosa" T9500) using Intel chip sets
> (965GM Northbridge and 1CH8M Southbridge) with 4 GB DDR-2 RAM.
>
> We plan to use a Debian LINUX distribution (64 bit).
>
> The DSP project entails handling 48 audio streams (DS0) from two T1
> connections. Each of the 48 DS0s presents as 8 bit u-law compandored PCM at
> 64 kbps. The T1 interface card will have hardware to convert the
> compandored PCM to linear 16 bit PCM so we will not need to do that in
> DSP.
>
> Twenty-four of the incoming streams will include two frequency bands that
> must be separated. We plan to use IIR dual biquads to provide the necessary
> filtering (estimated 4 biquads per stream). In addition, we need to
> implement an FSK decoder for a 300 bps non-MSK format (200 hertz frequency
> separation).
>
> The other twenty-four incoming streams will have a similar dual-biquad
> filter as above (estimate 4 biquads per stream). A 300 bps modem will then
> be added to the stream.
>
> I am trying to find some benchmark for MFLOPS per biquad. And some idea of
> the MFLOPS available from a mini-ITX computer as I described above so that
> I can decide whether this is practical or whether I need to plan on usinga
> DSP daughter card to handle the filtering, etc.
>
> MY QUESTIONS:
>
> 1. *How many MFLOPS does a dual-biquad require?
> 2. *What is the approximate DSP performance (ie. MFLOPS) of a PC style
> computer as I described above?
> 3. *Should I consider an off-the-shelf DSP PCI board? Can you recommend
> one? Or are there other options such as a USB connected external DSP
> engine?
>
> Many thanks for any insights you may be able to share.
>
> Regards,
> Larry

When doing DSP on a general-purpose computing platform running a non-
realtime OS like Linux, it's hard to specify how many "MFLOPS" you can
get. Your application is going to be sharing time with many other
system services, other processes on the system, etc. There are a lot
of variables in the system that affect your performance, like memory
bandwidth, I/O bus bandwidth, etc. Also, your performance can vary
widely based upon other factors that you might not consider, such as
how well your code interacts with the CPU cache. On modern x86
processors, you can see large performance differences just by changing
the order of some operations to make them more cache-friendly.

With that said, I've found that given some effort you can often wring
more performance out of these CPUs than you might initially think. Use
parallelism wherever possible (multiple threads to take advantage of
multiple cores, SIMD instructions to take advantage of execution unit
width). I would recommend doing some prototyping and benchmarking to
get an idea of what your platform might be able to support.

Jason

lhicks wrote:

> Hello...
>
> Thanks for reading my question.
>
> I am trying to determine whether a commercial grade mini-ITX style
> computer can be used for the DSP job that I describe below. The mini-ITX
> has dual Intel cores (Core 2 Duo "Santa Rosa" T9500) using Intel chip sets
> (965GM Northbridge and 1CH8M Southbridge) with 4 GB DDR-2 RAM.

I would strongly unrecommend using commertial PC for industrial tasks,
especially if you are planning mass deployment of your solution. This is
a dead end; it seems like attractive idea initially, however many
people regret about that later.

> We plan to use a Debian LINUX distribution (64 bit).
>
> The DSP project entails handling 48 audio streams (DS0) from two T1
> connections. Each of the 48 DS0s presents as 8 bit u-law compandored PCM at
> 64 kbps. The T1 interface card will have hardware to convert the
> compandored PCM to linear 16 bit PCM so we will not need to do that in
> DSP.

Converting uLaw into PCM is a trivial 8-bit table lookup. It takes nothing.


> Twenty-four of the incoming streams will include two frequency bands that
> must be separated. We plan to use IIR dual biquads to provide the necessary
> filtering (estimated 4 biquads per stream). In addition, we need to
> implement an FSK decoder for a 300 bps non-MSK format (200 hertz frequency
> separation).
>
> The other twenty-four incoming streams will have a similar dual-biquad
> filter as above (estimate 4 biquads per stream). A 300 bps modem will then
> be added to the stream.
>
> I am trying to find some benchmark for MFLOPS per biquad.

Why can't you simply try this for yourself?

> And some idea of
> the MFLOPS available from a mini-ITX computer as I described above so that
> I can decide whether this is practical or whether I need to plan on using a
> DSP daughter card to handle the filtering, etc.
>
> MY QUESTIONS:
>
> 1. How many MFLOPS does a dual-biquad require?
> 2. What is the approximate DSP performance (ie. MFLOPS) of a PC style
> computer as I described above?
> 3. Should I consider an off-the-shelf DSP PCI board? Can you recommend
> one? Or are there other options such as a USB connected external DSP
> engine?
>
> Many thanks for any insights you may be able to share.

A LOT depends on how your code and data fits in cache, and if/how do you
utilize advanced features such as SSE. For non-SSE code entirely in
cache you can expect something like 3-4 CPU clocks per filter tap.
Look for Intel Signal Processing library; it has some benchmarks as well.


Vladimir Vassilevsky
DSP and Mixed Signal Design Consultant
http://www.abvolt.com

Thanks for the prompt replies. I have done some additional research and
have decided to off-load the DSP calculation to one of our standard DS
products. Its a somewhat more expensive approach but I can be sure I won'
run out of DSP horsepower.

Much Appreciated,
Larry


>Hello...
>
>Thanks for reading my question.
>
>I am trying to determine whether a commercial grade mini-ITX style
>computer can be used for the DSP job that I describe below. The mini-ITX
>has dual Intel cores (Core 2 Duo "Santa Rosa" T9500) using Intel chi
sets
>(965GM Northbridge and 1CH8M Southbridge) with 4 GB DDR-2 RAM.
>
>We plan to use a Debian LINUX distribution (64 bit).
>
>The DSP project entails handling 48 audio streams (DS0) from two T1
>connections. Each of the 48 DS0s presents as 8 bit u-law compandored PC
at
>64 kbps. The T1 interface card will have hardware to convert the
>compandored PCM to linear 16 bit PCM so we will not need to do that in
>DSP.
>
>Twenty-four of the incoming streams will include two frequency band
that
>must be separated. We plan to use IIR dual biquads to provide th
necessary
>filtering (estimated 4 biquads per stream). In addition, we need to
>implement an FSK decoder for a 300 bps non-MSK format (200 hert
frequency
>separation).
>
>The other twenty-four incoming streams will have a similar dual-biquad
>filter as above (estimate 4 biquads per stream). A 300 bps modem wil
then
>be added to the stream.
>
>I am trying to find some benchmark for MFLOPS per biquad. And some ide
of
>the MFLOPS available from a mini-ITX computer as I described above s
that
>I can decide whether this is practical or whether I need to plan on usin
a
>DSP daughter card to handle the filtering, etc.
>
>MY QUESTIONS:
>
>1. How many MFLOPS does a dual-biquad require?
>2. What is the approximate DSP performance (ie. MFLOPS) of a PC style
>computer as I described above?
>3. Should I consider an off-the-shelf DSP PCI board? Can you recommend
>one? Or are there other options such as a USB connected external DSP
>engine?
>
>Many thanks for any insights you may be able to share.
>
>
>Regards,
>Larry
>
>
>