Why accelerated computing

The main reason for accelerated computing is speed. Either for a real-time response (when data must be processed as it is generated, for example in video processing for autonomous vehicle guidance or radar signal processing) or an increased speed to improve the overall performance. There are also situations of competitiveness, like in Financial Computing, Blockchain and military applications where 'the winner takes it all' and speed is the key.

While there are a number of ways to cater for a computing-intensive problem, FPGA offer a small footprint and low power consumption, making them ideal embedded and battery powered solutions. Being reconfigurable, they allow for upgrades as well as deep changes.

FPGA development has traditionally being regarded as 'difficult'. That is probably the picture of the wordy VHDL and Verilog times, kind of the assembly code ages in software. Today's FPGA development is mostly done with graphical interfaces and High Level Synthesis (HLS) tools that reduce the design and test cycle and improve the quality of the result.

How the existent solutions rate

CPUs offer the lowest cost with a limited performance. Software is readily available and deploying a new computing solution is easy.

Next step in performance is to add GPU cards to a computer. There is a notable boost in performance with a modestly higher cost. There is also an extra development effort for the GPU programming.

In an increasing number of cases, like Computer Vision and Machine Learning, a GPU solution is either too bulky or uses too much power. FPGA can offer better footprint and energy usage figures with extra performance because of customization to the specific problem. Learning curve and development is usually higher but the existence of prepackaged IPs and High Level Synthesis (HLS) tools is making FPGA more accessible.

Last is the ASIC (Application Specific Integrated Circuit) solution. While the increase in performance respect FPGAs is moderate, the necessary investment and risks are much higher. It only makes sense for medium-high production.