Probability successful in FPGA if done in modelsim!!

almost 100% success if you know the proper design methodology.

Differences occur for various reasons, e.g.:

Improper simulation coverage or stimuli

Your simulation must be automated to remove human error

overflow/underflow of buffers

timing violations

wrong startup logic and stuck state machines or stuck handshake signals

wrong operating temperature

undriven inputs during development.

and the list goes on...

occasionally device or ips or your compiler are to blame

What kaz said. Plus heres my list of ways to make it work as well as possible from the start:

1. Make an algorithm in Matlab

2. Architect the algorithm so you can match the algorithm architecture in simulink to the architecture on the FPGA. (eg fixed point, data flow etc)

3. Run test vectors.

4. Convert algorthm to HDL. Make sure you use good design practice (all synchronous - any clock domain crossings done with FIFOs where possible, or at least double register)

5. Create a self checking test bench that uses the same input as the simulink model, or even better, use the simulink model as the testbench (you can use modelsim co-simulation with simulink). Do black box testing to ensure HDL results match the simulink model. Also during testing, make sure you hammer the design with all edge and corner cases and a random scattering in between, plus stress the control inputs with random control (assuming you have thinks like memory bandwidth and contention to take into account).

Only when you've done this, can you try and compile the design. (but of course, you were constantly checking your modules through the synthesisor to make sure they met the FMax you need)

6. Compile the design with timing constraints. Iterate until it fits.

(7. If you have the resources, use the simulink model as the testbench again with the real hardware running for black box testing)

8. Use a real input source, and check the design.

Using these steps should help minimise any design problems. Then you can look at actual hardware faults at the end.

--- Quote Start ---

What kaz said. Plus heres my list of ways to make it work as well as possible from the start:

1. Make an algorithm in Matlab

2. Architect the algorithm so you can match the algorithm architecture in simulink to the architecture on the FPGA. (eg fixed point, data flow etc)

3. Run test vectors.

4. Convert algorthm to HDL. Make sure you use good design practice (all synchronous - any clock domain crossings done with FIFOs where possible, or at least double register)

5. Create a self checking test bench that uses the same input as the simulink model, or even better, use the simulink model as the testbench (you can use modelsim co-simulation with simulink). Do black box testing to ensure HDL results match the simulink model. Also during testing, make sure you hammer the design with all edge and corner cases and a random scattering in between, plus stress the control inputs with random control (assuming you have thinks like memory bandwidth and contention to take into account).

Only when you've done this, can you try and compile the design. (but of course, you were constantly checking your modules through the synthesisor to make sure they met the FMax you need)

6. Compile the design with timing constraints. Iterate until it fits.

(7. If you have the resources, use the simulink model as the testbench again with the real hardware running for black box testing)

8. Use a real input source, and check the design.

Using these steps should help minimise any design problems. Then you can look at actual hardware faults at the end.

--- Quote End ---

@ Tricky, Thanks for your valuable information,

I'm wondering about these steps, are they specified to some sort of algorithms such as DSP, or image processing. Or it's more general method do an algorithm?