FFT in TI's DSP

Hello,

Does anyone implement 256pt fixed point FFT on TI's DSP chip? If so,
would you please let me know the model, response time and etc?

Thanks in advance,

Mike

Mike wrote:
> Hello,
> 
> Does anyone implement 256pt fixed point FFT on TI's DSP chip? If so,
> would you please let me know the model, response time and etc?

Mike,

"TI's DSP chip" is about as specific as "GM's motor vehicle". Can you 
narrow it down?

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

Hi Jerry,

How about Spartan? Actually I am seeking the chip as cheap as possible.

Thanks,

Mike

A Spartan as a Xilinx DSP--much different than a TI DSP.  I suggest you do a bit
of background research and then re-phrase your questions.  Also, try to include
as much detail as possible--the better your question, the better our answer can
be.

-- 
Jon Harris
SPAM blocked e-mail address in use.  Replace the ANIMAL with 7 to reply.

"Mike" <mail2mz@gmail.com> wrote in message
news:1116947808.644643.52820@z14g2000cwz.googlegroups.com...
> Hi Jerry,
>
> How about Spartan? Actually I am seeking the chip as cheap as possible.
>
> Thanks,
>
> Mike
>

Sorry to confuse everyone. Actually I am trying to compare the Pros and
Cons for the implemetation of FFT using DSP or FPGA.  For DSP,  I would
like to us TI's chip; For FPGA I would like to use Xilinx's chip. I
think both of them are able to complete a 128 pt or 256 pt fix point
FFT. What I don't know are the price, speed and any other concern.

Thanks,

Mike

"Mike" <mail2mz@gmail.com> wrote in message
news:1116948847.214232.269800@g49g2000cwa.googlegroups.com...
> Sorry to confuse everyone. Actually I am trying to compare the Pros and
> Cons for the implemetation of FFT using DSP or FPGA.  For DSP,  I would
> like to us TI's chip; For FPGA I would like to use Xilinx's chip. I
> think both of them are able to complete a 128 pt or 256 pt fix point
> FFT. What I don't know are the price, speed and any other concern.

Still not good enough in the information dept to be able to provide
meaningful answers.
What is the input sample rate of the data, what is the max clock speed you
are going to be able to run your FPGA and your DSP?
It feels like you have a lot more learning to do before our answers can
really be of any use to you (even after you provide all the information). If
it feels like we are belitttling you, our apologies.

Cheers
Bhaskar

> Thanks,
>
> Mike
>

Mike wrote:
> Sorry to confuse everyone. Actually I am trying to compare the Pros and
> Cons for the implemetation of FFT using DSP or FPGA.  For DSP,  I would
> like to us TI's chip; For FPGA I would like to use Xilinx's chip. I
> think both of them are able to complete a 128 pt or 256 pt fix point
> FFT. What I don't know are the price, speed and any other concern.
> 
> Thanks,
> 
> Mike
> 
An n-point FFT requires something on the order of n*log(n) multiplies 
and adds.  Figure that the TI chip is going to require something on the 
order of 5-20 clock ticks * n * log_2(n) to get it done (someone will 
want to refine that figure, which is good because I haven't actually 
coded an FFT on a DSP).  An FPGA will do whatever you design it for, 
with a _BIG_ tradeoff between speed, logic utilization, and engineering 
time -- e.g. with a big enough FPGA you could do an n-point FFT in 
log_2(n) steps, but it'd have to have 2n multipliers to do it.

Check the archives of places like Electronic Design News, Electronic 
Design, possibly look for whitepapers on the TI and Xilinx websites -- 
there's a lot more information that you need than we can answer here in 
reasonable length posts.

-- 
-------------------------------------------
Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

Bhaskar Thiagarajan wrote:

> "Mike" <mail2mz@gmail.com> wrote in message
> news:1116948847.214232.269800@g49g2000cwa.googlegroups.com...
> 
>>Sorry to confuse everyone. Actually I am trying to compare the Pros and
>>Cons for the implemetation of FFT using DSP or FPGA.  For DSP,  I would
>>like to us TI's chip; For FPGA I would like to use Xilinx's chip. I
>>think both of them are able to complete a 128 pt or 256 pt fix point
>>FFT. What I don't know are the price, speed and any other concern.
> 
> 
> Still not good enough in the information dept to be able to provide
> meaningful answers.
> What is the input sample rate of the data, what is the max clock speed you
> are going to be able to run your FPGA and your DSP?
> It feels like you have a lot more learning to do before our answers can
> really be of any use to you (even after you provide all the information). If
> it feels like we are belitttling you, our apologies.
> 
> Cheers
> Bhaskar
> 
> 
To rephrase your question:

"I have some cargo to take across the lake.  Do I need a rowboat, a 
powerboat, or a freighter?".

To rephrase Baskar's reply:

Is your cargo a cup of sugar or 100 shipping containers thereof?  Is the 
lake is 100 yards across or is it Lake Superior?  How soon does your 
cargo need to get across the lake?

Do you see our problem here?

-- 
-------------------------------------------
Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

Mike wrote:

> Sorry to confuse everyone. Actually I am trying to compare the Pros and
> Cons for the implemetation of FFT using DSP or FPGA.  For DSP,  I would
> like to us TI's chip; For FPGA I would like to use Xilinx's chip. I
> think both of them are able to complete a 128 pt or 256 pt fix point
> FFT. What I don't know are the price, speed and any other concern.

It is exactly the price and speed you need to know.

If the DSP is fast enough, that will most likely be the best solution.
If not, then you either need multiple DSP or FPGA.   For the FPGA case,
either a fairly large one or multiple chips.  The FPGA solutions will
tend to scale to larger arrays fairly easily, for example a systolic
array solution should be possible where a longer array will do longer
transforms at a give clock rate.   If you don't need one result per
clock cycle, you might be able to do some sharing of hardware between
different parts of the algorithm.

-- glen

 > For FPGA I would like to use Xilinx's chip.

hmm ... I hope you are at least talking about Spartan2 (or above - 
Spartan 2E for example)

 > What I don't know are the price, speed and any other concern.

the chip is cheap (XC2S150E is 22$ at digikey@90pc) but do you have the 
developement tools and the knowledge?
can you live without a multiplication?
can you build the interface to you data-source? (unfortunately this is 
usually more complicated)

Price:
I think you can get Spartan 2E chips from XC2S50E (few logic) up to 
XC2S600 (lots of logic elements) ... the price/availability has more 
fluctuation between vendors than in the family ...
(last year we had a factor 2 in price and 18 weeks of waiting difference 
for a bigger Xilinx FPGA part)



Speed:
for the FPGA you can have everything ...
- you could implement a general purpose CPU and calculate the fft on 
that (will be really slow - but there are free CPUs and you will soon 
have results ;-)
- you could take the xilinx Core generator and use a ready made fft 
module (license $$)
- there is a fft core on opencores.org ... "a 1024 point 12 bit FFT runs 
at about 97MHz in a Spartan2e100 -6 device and occupies about 52% of the 
device." ... but they don't say how long it will take for one block
(and I haven't seen a 128 point fft)
- you can write your own, fully pipelined parallel fft - you can almost 
any throughput - the limit is memory bandwidth (and the fact, that the 
latency is not as easy to be reduced)


other concerns:
- If you are new to hardware, you need 1/2 year for learning the FPGA 
developement ... thats more expensive than the chip



the DSP is easier for getting it going ... if you need more speed then 
you need the fpga (and a lot more time)


are you actually trying to build something? just a prototype? a real 
product with > 100k? or is it just a comparison for a paper ..?


bye,
Michael