With the development of artificial intelligence, high-speed wireless communication, medical and life sciences technology, and the increasing demand for data flow optimization in complex chip architectures, FPGAs are ushering in new application scenarios.

The biggest feature of FPGAs (Field Programmable Gate Arrays) is their ability to reprogram and reconfigure digital logic after chip deployment. In the world of AI, the speed of algorithm updates far exceeds that of chip architecture updates, making this flexibility a key advantage.

Moreover, FPGAs have lower upfront costs and are often used for prototype validation of fixed-function chips like ASICs, or as a transitional solution before stabilizing a workload. However, due to manufacturing costs, they are more commonly found in specialized markets with low shipment volumes and high performance, such as fighter jets or laboratory high-precision measurement equipment.

However, the FPGA market is expanding, with emerging applications including:

• Cloud servers accelerating complex algorithms, particularly suitable for medical and research scenarios involving 3D mathematical computations;

• Low-cost virtual hardware validation, used for hybrid prototype systems;

• Alleviating memory and I/O bottlenecks in SoCs and NoCs;

• AI data pipeline management and preprocessing;

• 5G/6G base stations and core network devices;

• Horizontal product differentiation and custom hardware functions;

• Responding to changes in security protocols as a reprogrammable insurance mechanism.

The main high-volume users of high-end FPGAs include communications (wireless and wired), data centers, network devices, as well as military aerospace and government sectors.

Venkat Yadavalli, head of business management at Altera, stated: “The annual shipment volume of these applications ranges from thousands to hundreds of thousands. As FPGA integration features enhance, such as embedded Arm cores and AI Tensor Blocks, the applicability of FPGAs is continuously expanding into embedded and edge AI.”

01Robotics and Medical Imaging Become Rapidly Growing Fields

Robotics requires deterministic latency and real-time decision-making at the edge, integrating a large amount of heterogeneous data from cameras and sensors. Yadavalli said: “Voice, video, sensor fusion… these are all very suitable for FPGAs.”

The applications of medical imaging are even broader, ranging from ophthalmic retinal scans to MRIs. Rob Bauer, senior manager of adaptive and embedded product marketing at AMD, stated:

“You first collect a large amount of analog data, which must be filtered, processed, and undergo matrix mathematical calculations to reconstruct images. FPGAs excel at these tasks.”

The AI engines within FPGAs are also very suitable for volumetric imaging.

02FPGA vs ASIC: Not a Choice Between Two, but a Hybrid Strategy

In today’s world where high-speed interconnects and AI acceleration are becoming the norm, choosing between FPGAs, ASICs, or other ICs is not a simple binary choice.

Andy Nightingale, vice president of product management at Arteris, said:

“Designers ultimately adopt hybrid solutions. Interconnect architectures allow for the mixed use of FPGAs and SoCs, for example, for network security or alleviating I/O bottlenecks. Some modules will retain FPGAs even if SoCs require custom hardware, as protocols and standards will change.”

For instance, if network security standards change, FPGAs can be reprogrammed directly without the need to spend months redesigning ASICs. This is particularly important in rapidly evolving industries.

However, in terms of logic unit costs, FPGAs can never achieve the extreme optimization of ASICs. Yadavalli pointed out:

“ASICs or ASSPs are optimized for a single fixed function, while FPGAs need to maintain a general programmable structure, which inevitably sacrifices area and power consumption.”

Ultimately, it is a trade-off between market scale, algorithm stability, and ROI.

03Increased Application Complexity Makes FPGAs More Valuable

Bauer stated:

“In AI in the physical world, increased throughput of edge-side sensor data, lower latency requirements, and higher security demands are all typical strengths of FPGAs. It is a matter of continuous optimization.”

04AI Models are Constantly Evolving, and the Value of FPGAs Fluctuates Accordingly

As AI/ML models continuously change, the reprogrammability of FPGAs brings advantages, but also makes the question of “FPGA or ASIC” a dynamic one.

Mo Faisal, CEO of Movellus, explained:

“When workloads change frequently, you tend to favor general-purpose computing. When algorithms stabilize, you customize. This cycle will repeat continuously.”

For example, large scientific experiments (such as CERN or Fermilab) may deploy thousands of FPGAs, but when scaling to more systems, cost and efficiency may become insufficient, necessitating a reevaluation of the architecture.

According to Zhang Kexun, head of research at ChipAgents:

“In fields like life sciences, model structures have not yet converged, requiring hardware that is both fast enough and customizable. FPGAs will shine in these scenarios.”

In the consumer market, the focus is more on cost and efficiency. Nandan Nayampally, CCO of Baya Systems, pointed out:

“In scenarios like high-frequency trading where algorithms change rapidly, FPGAs are more suitable. However, if there isn’t enough volume to support ASICs, FPGAs will be maintained.”

The advantage of FPGAs lies in their customizable parallel processing structure, allowing for the construction of dedicated hardware logic based on algorithms, rather than using general cores like GPUs.

05Cloud-Based Virtual FPGAs: New Tools for Complex Algorithms and Validation

Cloud-based FPGAs (such as AWS EC2 F1/F2 instances) can be used to offload compute-intensive workloads from data centers.

Russell Klein from Siemens EDA introduced:

“AWS’s F2 instances include a PCIe card with eight Xilinx FPGAs, which customers can program through PCIe. We have already enabled high-level synthesis tools to directly generate acceleration modules that can be deployed on F1/F2 instances.”

The life sciences field has extensively used FPGAs to accelerate complex 3D mathematical operations such as DNA or chemical reaction analysis.

At the same time, cloud FPGAs have also become a low-cost hardware validation method.

For example, SiFive used to rent FPGAs on AWS for only $6 per hour to run the operating system of a multi-core RISC-V SoC on virtual FPGAs.

Klein stated:

“This is a low-budget ‘self-built simulator’ that can be quickly started and stopped, making it very suitable for complex system validation.”

AMD’s Bauer also pointed out that in an era of skyrocketing costs at the 2nm node, FPGAs allow software teams to conduct large-scale validation before chip tape-out, which is a key driver for shift-left practices.

06Reducing Data Bottlenecks in SoCs/Chiplets to Enhance AI System Efficiency

FPGAs can also be directly placed in data paths to reduce memory or I/O bottlenecks and improve data movement efficiency.

The latest collaboration between Arteris and Altera is aimed at this purpose.Nightingale stated:

“FPGAs reduce buffering and enhance throughput through stream data management, while also completing preprocessing before data reaches the CPU/GPU, thereby alleviating critical performance bottlenecks in AI systems.”