How can a microcontroller be more efficient than an FPGA?

Question

If we don't count the cost of both (MCU, FPGA) are there any applications where a microcontroller can be more efficient than an FPGA? It is easier to program a microcontroller than an FPGA (embedded C vs VHDL, for example) but if someone spent time and implemented the MCU design inside an FPGA, then the FPGA could be used for the same application as the MCU but more efficient because we can implement more than a single MCU in an FPGA.

I was wondering if the only application that an MCU can be more efficient than an FPGA is when we connect an analog sensor to the MCU (for example an analog source that harvest entropy in case of producing random numbers etc)? Is this correct or I am totally wrong?

edit(as gbulmer suggested): My criteria for efficiency are speed of execution and throughput/watt

Answer 1

See When can FPGA's be used and Microcontrollers/DSPs not?

Microprocessors will nearly always win on throughput/watt. For total throughput, you can get quite good results with DSP chips, larger microcontrollers, or ARMs with onboard graphics cores. Although it may be hard to repurpose the graphics cores to other applications if the platform doesn't support CUDA or similar.

FPGAs win on latency, especially if you want to guarantee low latency response to lots of inputs. You can only have one highest priority interrupt on a microcontroller. FPGAs are also good if you want a fixed-function process that doesn't use a lot of RAM, such as a compression algorithm from some sort of signal capture.

Answer 2

The key difference between FPGA's and microprocessors/microcontrollers is that a µp is a specialized circuit, and an FPGA is a general circuit that can be 'run-time' configured. In a fair comparison (same die size and process parameters) a µp/µc will win for what it is designed for: running vastly different sequences of instructions, with little repetition and lots of decisions and data dependencies. In the extreme, think of running Linux or Windows, or a scaled down equivalent for a smaller µc.

An FPGA will win when it can use its configurability and parallellism: on simple and repeated transformations, that do not translate effectively to µc/µp instructions.

Take for instance IEEE floating point calculations. A µp/µc will often have dedicated hardware for that, which is more efficient than an FPGA configured to do the same. But change a few trivial details in the IEEE specification, and the µc/µp hardware is useless, but the FPGA can be configured accordingly, and it will be as efficient as it was with the original specs.

Note that a µc/µp can trade throughput for complexity: if you make the task more complex, it can still do it, but it will take more time. An FPGA will need more FPGA real estate if the job gets more complex. In other words: in an FPGA the functionality is spread over (requires more) the FPGA real estate, in a µc/µp it is spread over (requires more) time.

Summary: a µc/µp will be more efficient for what it is designed for (which is general serial computation), an FPGA will be more efficient for a 'small' parallelizable task.