Las FPGAs, del prototipado rápido al corazón de los centros de datos

Field Programmable Gate Arrays (FPGAs) are reconfigurable devices made up of thousands of logic elements, which functionality can be modified as many times as needed. Due to their hardware nature, FPGAs are able to exploit the parallelism available in multiple algorithms, as is the case of the Application Specific Integrated Circuits (ASICs). However, while ASICs cannot be modified once manufactured, reconfigurability provides FPGAs with a flexibility similar to the software technology, while their development costs are also well below the figures corresponding to ASICs. At the beginning, FPGAs were mainly used for the fast prototyping of integrated circuits. However, their capacity and efficiency have been growing exponentially since Xilinx launched the first FPGA to the market in 1984. As a consequence of this evolution, state of the art devices are usually used within military and aerospace systems, as well as in any area of the professional and even of the consumer electronics.

In the last months, efficiency and capacity improvements have been accelerated with the emergence of the last generations of devices manufactured in 28 and 14 nm technologies. In addition, companies have bet heavily on high level synthesis languages, which bring the design of FPGA-based systems closer to the software. As a consequence, different reports appearing in the last weeks show that the journey for the FPGAs is not over yet. First, Microsoft announced that they will install next year FPGAs in all the datacenters giving support to the Bing search engine. FPGAs were tested experimentally in one of their centers, showing a throughput improvement of 95%, while increasing the power consumption by a 10% and a 30% the cost. Moreover, the improvements achieved in the search algorithm can be translated straight to the server, similarly to what happens with software. In the same direction, Baidu has also confirmed that they are using FPGAs to boost up the execution of deep neural networks to improve data searching, but also other tasks such as speech and image recognition, within their servers. FPGAs are also gaining a strong position in the financial market. Recently, investment funds and banks such as Deutsche Bank or JPMorgan have described how they have integrated FPGAs in their datacenters to accelerate the risk assessment, and this way being able to reduce the latency when sending stock purchase or sales orders. This business strategy is called high frequency trading.