FPGAs matter today because they sit between processors and custom chips: they can be rewired after manufacture to build a hardware circuit for one job, so they deliver much of an ASIC's speed and efficiency without its cost, risk and long lead time. That makes them a strong fit where algorithms and standards keep changing, where latency must be low and predictable, and where volumes are too small to justify a custom chip. Their weaknesses are higher unit cost and power than an ASIC, and harder programming than a CPU or GPU.
Field-Programmable Gate Arrays (FPGAs) are becoming increasingly important in the new technological world. They offer a number of advantages over traditional Application-Specific Integrated Circuits (ASICs) and General-Purpose Processors (GPPs), making them a popular choice for a wide range of applications.
Key takeaways
- An FPGA is a grid of configurable logic, memory blocks, DSP blocks and I/O whose connections are set by a configuration file (bitstream) loaded after manufacture.
- Flexibility: the same chip can be reprogrammed in the field when a model, protocol or standard changes.
- Performance: it wins through custom parallel pipelines, not clock speed; FPGA designs typically run at a few hundred MHz.
- Efficiency: for a fixed, well-matched task it can use less energy per operation than a CPU, but usually more than an ASIC built for the same task.
- In data centres FPGAs are used as accelerators for networking, storage, security, video and ML inference.
What makes an FPGA different
A CPU runs one instruction stream (or a few) on fixed hardware. A GPU runs thousands of identical threads on fixed hardware. An FPGA has no fixed datapath at all: the designer describes a circuit in Verilog, VHDL or a high-level synthesis language, and the tools map it onto lookup tables (LUTs), flip-flops, block RAM and DSP slices. The result is hardware shaped exactly to the problem, with data flowing from stage to stage every clock cycle.
| Property | CPU / GPP | GPU | FPGA | ASIC |
|---|---|---|---|---|
| Hardware after manufacture | Fixed | Fixed | Reconfigurable | Fixed |
| Typical clock | 3 to 5 GHz | 1 to 2.5 GHz | 100 to 600 MHz | Design dependent |
| Latency | Variable (OS, caches) | High for small batches | Low and deterministic | Lowest |
| Energy per operation for a fixed task | Highest | Medium | Low to medium | Lowest |
| Upfront cost and time | None (software) | None (software) | Low to medium, weeks to months | Very high, many months to years |
| Ease of programming | Easiest | Moderate | Hardest | Not reprogrammable |
The clock figures above are typical ranges, given for orientation only.
Flexibility
FPGAs are known for their flexibility, as they can be programmed to perform specific functions after they have been manufactured. This makes them ideal for applications where the requirements are constantly changing, such as artificial intelligence and machine learning, where algorithms are constantly being updated. It also suits telecom and networking, where a new standard release or protocol can be supported by a bitstream update rather than a new board. Many devices also support partial reconfiguration, where one region of the chip is reloaded while the rest keeps running.
Performance
Another advantage of FPGAs is their high performance. They can process data much faster than traditional processors in some applications, making them ideal for applications that require real-time processing, such as video and image processing, high-speed networking and scientific simulations.
The speed comes from parallelism and pipelining. A worked example, with illustrative numbers: a 64-tap FIR filter needs 64 multiply-accumulates (MACs) per output sample. A CPU core doing, say, 16 MACs per cycle with SIMD at 3 GHz needs 4 cycles per sample, so about 750 million samples per second at best, and less once memory and other work intrude. An FPGA design that unrolls all 64 taps across 64 DSP slices produces one output every clock; at 400 MHz that is 400 million samples per second from one copy, and the design can be replicated several times on a larger device, with latency fixed at a few tens of clock cycles. The key point is not peak speed but sustained, predictable throughput with no cache misses or operating system jitter.
Power efficiency
In addition, FPGAs consume less power than traditional processors for specific applications, making them a more energy-efficient option in data centres. This is also particularly important in applications where power consumption is a concern, such as mobile devices, drones and wearable technology. Small, low-power FPGAs are used in these products for sensor hubs, display and camera bridging and always-on functions. The qualifier matters: an FPGA is more efficient than a CPU only when the task maps well onto custom hardware, and an ASIC built for the same task will normally use less power again, because programmable routing costs area and energy.
FPGAs in data centres
FPGAs are often used in data centres to accelerate various types of workloads and applications. Public examples include Microsoft's Project Catapult, which put FPGAs in Bing and Azure servers, and Amazon EC2 F1 instances, which rent FPGA accelerators by the hour. Typical uses are:
Customisable acceleration
FPGAs can be programmed to perform custom functions that are specific to a particular application, allowing for specialised acceleration that is not possible with fixed-function accelerators. See FPGA based customised acceleration.
Machine learning inference
FPGAs can be programmed to perform inference on machine learning models, with significant performance gains over CPU-based inference for suitable models, especially at low batch sizes and with reduced-precision arithmetic such as INT8. See FPGA based machine learning inference.
Cryptographic acceleration
FPGAs can be used to offload cryptographic processing from CPUs, freeing CPU resources and improving performance and security. See FPGA based cryptographic acceleration.
Database acceleration
FPGAs can be used to offload database processing tasks, such as query processing (filtering and aggregation close to storage), indexing and data compression, resulting in improved database performance.
Network acceleration
FPGAs can be used to perform custom functions such as packet classification, packet filtering and packet processing, improving network performance and security. FPGA-based SmartNICs and low-latency trading systems are well-known examples.
Video processing acceleration
FPGAs can be used to perform real-time video processing tasks, such as video encoding and decoding, transcoding and video analytics, improving the performance of video applications. See FPGA based video processing acceleration.
When an FPGA is not the right choice
- Very high volume and stable function: an ASIC wins on unit cost and power once the volume pays back its development cost.
- Large-batch training of big neural networks: GPUs and dedicated AI accelerators have far more memory bandwidth and a mature software stack.
- Branch-heavy, irregular software: a CPU is simpler and usually faster.
- Teams without hardware skills: high-level synthesis helps, but timing closure and debugging still need hardware engineers.
FAQ
Why are FPGAs slower in clock speed but faster in some tasks?
Because they do many operations in parallel every cycle in a pipeline built for the task, while a CPU does a few operations per cycle on general hardware.
Are FPGAs used for prototyping ASICs?
Yes. FPGA prototyping lets software teams run on a near-final hardware model months before silicon, and large multi-FPGA systems are built for this purpose.
Do FPGAs replace GPUs for AI?
Not in general. They compete well for low-latency inference and custom data types, while GPUs dominate training.
For a broader list of uses, see Top 10 applications of FPGAs, and for the cost trade-off see ASIC vs structured ASIC vs FPGA.
Your comments will be moderated before it appears here.