How to know if i need fixed-point of floating-point?

[Antonio_Magma]

My project involves implementing an OFDM transciever using a DSP.

I have currently narrowed down my list to 2 DSPs:
TMS320C6416 DSK (Fixed-point) Cost: USD500
TMS320C6713 DSK (Floating-point) Cost: USD400

Question:
How do i determine whether my application requires a fixed-point or floating-point processor?

 

[luancao]

It depends on the algrithm of your OFDM transciever.

Is there any floating pointing calulation in it? May be the answer is always yes. However, If the there is not many floating pointing caculation, you could still use the fixed point DSP. Since fixed point algrithm is easy to handle, most designer like to use fixed point DSP in any case that the complexity is still within his/her ability to handle.

To make things simple, if you could port the algrithm easily into a fixed point algrithm use the fixed point DSP.

 

[sph3r3]

Whether you can use fixed- or floating- point depends on your application. You have to consider the dynamic range (i.e. the difference between the largest and smallest numbers that can occur in your computation) and noise floor that your application requires. Both of these will probably be limited by your ADC/DAC stages. That will then determine whether you need floating-point arithmetic (high dynamic range, but typically slower than integer arithmetic for the same $$$) or fixed-point, and how many bits of precision you need.

You could simulate your application in e.g. Matlab to find out how fixed/floating point limits the dynamic range & rounding noise in your application. Another alternative is to try to do it theoretically using something like interval arithmetic (finds the upper & lower error bounds due to rounding/truncation). This will probably be quite difficult to do for a complicated system like an OFDM transceiver.

 

[Antonio_Magma]

I plan to only implement a 64 or maybe 128 FFT only, without any complex coding. I have done some simulations in Matlab and they are some floating point values, since it involves complex numbers.

What i plan to do is to construct a Simulink model of my system and then generate C code from the Embedded Target for TI6000.

However, alot of blocks for DSPs and Embedded Target for TI6000 are fixed-point in Simulink.

Does it mean in this case it will be safer to get a fixed-point processor?