When is an FPGA Worth it and When is it NOT – when developing an Industrial Embedded System

FPGA Strengths/Best Suited:

Much of what will make it worthwhile to utilize an FPGA comes down to the low-level functions being performed within the device. There are four processing/algorithm attributes defined below that FPGAs are generally well-suited for. While just one of these needs may drive you toward an FPGA, the more of these your application has, the more an FPGA-based solution will appeal.

1. Parallel processes – if you need to process several input channels of information (e.g. many simultaneous A/D channels) or control several channels at once (e.g. several PID loops).
2. High data-to-clock-rate-ratio – if you’ve got lots of calculations that need to be executed over and over and over again, essentially continuously. The advantage is that you’re not tying up a centralized processor. Each function can operate on its own.
3. Large quantities of deterministic I/O – the amount of determinism that you can achieve with an FPGA will usually far surpass that of a typical sequential processor. If there are too many operations within your required loop rate on a sequential processor, you may not even have enough time to close the loop to update all of the I/O within the allotted time.
4. Signal processing – includes algorithms such as digital filtering, demodulation, detection algorithms, frequency domain processing, image processing, or control algorithms.

Weaknesses/Non-Optimal:

With any significant benefit, there’s often times a corresponding cost. In the case of FPGAs, the following are generally the main disadvantages of FPGA-based solutions.

1. Complex calculations infrequently – If the majority of your algorithms only need to make a computation less than 1% of the time, you’ve generally still allocated those logic resources for a particular function (there are exceptions to this), so they’re still sitting there on your FPGA, not doing anything useful for a significant amount of time.
2. Sorting/searching – this really falls into the category of a sequential process. There are algorithms that attempt to reduce the number of computations involved, but in general, this is a sequential process that doesn’t easily lend itself to efficient use of parallel logical resources. Check out the sorting section here and check out this article here for some more info.
3. Floating point arithmetic – historically, the basic arithmetic elements within an FPGA have been fixed-point binary elements at their core. In some cases, floating point math can be achieved (see Xilinx FP Operator and Altera FP White Paper ), but it will chew up a lot of logical resources. Be mindful of single-precision vs double-precision, as well as deviations from standards. However, this FPGA weakness appears to be starting to fade, as hardened floating-point DSP blocks are starting to be embedded within some FPGAs (see Altera Arria 10 Hard Floating Point DSP Block).
4. Very low power – Some FPGAs have low power modes (hibernate and/or suspend) to help reduce current consumption, and some may require external mode control ICs to get the most out of this. Check out an example low power mode FPGA here. There are both static and dynamic aspects to power consumption. Check out these power estimation spreadsheets to start to get a sense of power utilization under various conditions. However, if low power is critical, you can generally do better power-wise with low-power architected microprocessors or microcontrollers.
5. Very low cost – while FPGA costs have come down drastically over the last decade or so, they are still generally more expensive than sequential processors.