FPGA

FPGA. A field-programmable gate array is an integrated circuit designed to be configured by a customer or a designer after manufacturing. This configurability is achieved using hardware description languages to specify the desired digital logic, which is then synthesized and loaded onto the device. FPGAs are used across a wide range of industries, from accelerating algorithms in data centers to implementing control systems in industrial automation.

Overview

The fundamental concept is a semiconductor device containing an array of programmable logic blocks and a hierarchy of reconfigurable interconnects. Unlike fixed-function microprocessors or ASICs, the hardware functionality can be altered post-fabrication. This makes them invaluable for prototyping new designs, implementing evolving standards like 5G, and performing real-time signal processing in systems such as software-defined radio. Major companies driving the market include Xilinx, now part of AMD, and Intel following its acquisition of Altera.

Architecture

The core architecture typically consists of three primary elements. Configurable logic blocks, sometimes called slices or adaptive logic modules, contain look-up tables and flip-flops to implement combinatorial and sequential logic. A programmable routing matrix, comprising interconnect wires and programmable switches, creates pathways between these blocks. Additionally, specialized hardened blocks are embedded for performance, such as block RAM for memory, DSP slices for mathematical operations, and high-speed SerDes transceivers for protocols like PCI Express and Ethernet.

Programming and Design Flow

Designers use hardware description languages like VHDL and Verilog to describe the desired circuit behavior. This code is processed by synthesis tools from vendors such as Xilinx Vivado or Intel Quartus Prime, which generate a configuration bitstream. The flow often involves functional simulation with tools from Cadence Design Systems or Synopsys, followed by placement and routing and timing analysis. The final bitstream is loaded into the device via a JTAG interface or from an external flash memory, configuring the SRAM cells that control the logic and interconnects.

Applications

Their reconfigurable nature enables diverse applications across many fields. In telecommunications, they are used for network processing and baseband processing in equipment from Nokia and Ericsson. The aerospace and defense sector employs them for radar systems, electronic warfare, and flight controllers in projects like the Joint Strike Fighter. Within data centers, companies like Amazon Web Services use them for hardware acceleration of workloads including machine learning and genomics. They are also foundational in industrial control systems, medical imaging devices, and automotive systems for ADAS.

Comparison with Other Technologies

Compared to ASICs, they offer faster time-to-market and lower non-recurring engineering costs but at the expense of higher per-unit cost and greater power consumption. Microcontrollers and CPUs are software-programmable and excel at general-purpose tasks, but they lack the parallel processing capability and deterministic timing. GPUs from Nvidia are highly parallel but are optimized for floating-point operations on large data sets, whereas they provide fine-grained control over custom digital hardware. CPLDs offer similar programmability but are significantly less complex and dense.

History and Development

The invention is credited to Ross Freeman and co-founder Bernard Vonderschmitt, who founded Xilinx in 1984 and introduced the XC2064 in 1985. Early devices were used primarily for glue logic integration and prototyping for ASICs. The 1990s saw rapid architectural advances, including the introduction of SRAM-based configuration and larger densities, fueling adoption in telecommunications during the boom. The 2000s brought the integration of hard processor cores, such as the PowerPC in Xilinx Virtex-II Pro, leading to modern SoPC devices. Recent trends focus on heterogeneous integration for artificial intelligence, with companies like Achronix and Flex Logix pushing performance boundaries.