Due to setbacks in data centers, FPGA has been very quiet in recent years. But now, FPGAs have found new applications in fields such as artificial intelligence, high-speed wireless communication, medical and life science technology, and complex chip architectures, which can improve data flow.  
Field programmable gate arrays (FPGAs) enable designers to reprogram or reconfigure digital logic after chip deployment, which is crucial in the field of artificial intelligence because the update speed of algorithms far exceeds that of chip architectures. The initial cost of FPGA is relatively low, so it can be used as a prototype design tool for fixed function chips such as application specific integrated circuits (ASICs), or as a temporary solution before workload determination. However, due to its high manufacturing cost, FPGAs are typically used for small batch, high-performance special applications, such as fighter jets or laboratory measurement equipment.  
However, the FPGA market is still expanding. The new application areas include:
1. Cloud based servers can offload complex algorithms from data centers to medical and scientific applications involving 3D mathematics;  
2. Low cost virtual hardware verification through a hybrid prototype system; Relieve memory and I/O bottlenecks in SoC and NoC;  
3. AI infrastructure, data plane management, and preparation of data before it reaches the CPU or GPU;  
4. 5G/6G chips developed for base stations and core network equipment;  
5. Provide customized hardware functions through horizontal product differentiation;  
6. Prevent changes in security regulations and threats.  
The main large-scale users of high-end FPGAs include telecommunications companies (wireless and wired), data centers, networks, as well as military, aerospace, and government departments. The annual usage of these application areas ranges from thousands to hundreds of thousands, "said Venkat Yadavalli, head of Altera Business Management Group. In terms of application areas, FPGA integrates more functions, such as embedded Arm cores and AI tensor modules, which improves the utilization of FPGA in a wider range of embedded and edge AI applications. " 
Robotics technology is a thriving field that requires deterministic latency and utilizes heterogeneous data from cameras and sensors to make real-time decisions at the edge. It can be voice, video, or any data that requires sensor fusion, "Yadavalli said.  
Another important application scenario is medical imaging technology, which covers everything from optometrists scanning the retina to magnetic resonance imaging systems. You need to process raw analog data, filter and process it, and use it to reconstruct images, as well as perform a lot of matrix operations, "said Rob Bauer, Senior Manager of Adaptive and Embedded Product Marketing at AMD." FPGAs perform well in this regard. Artificial intelligence engines are highly skilled in matrix operations, making them an ideal choice for such volume imaging applications. " 
In today's rapidly changing, highly interconnected, and AI empowered environment, there is no unified answer as to whether designers should use FPGA, ASIC, or other types of ICs.  
Andy Nightingale, Vice President of Product Management and Marketing at Arteris, said, "You won't switch between two solutions at will. What you have is a combination or fusion of multiple solutions. Interconnection technology facilitates and supports hybrid applications of FPGA, such as for network security or mitigating I/O bottlenecks. FPGAs can be used for prototype design, but customers may retain certain FPGAs in the system. They may say, 'We want to use custom hardware in SoC, but we will retain some FPGA technology because the benefits of doing so outweigh the time required for custom hardware, even if specifications or standards are eventually found to have changed.'. 'If network security standards change, they must be re implemented and processed, and FPGAs may contain content that can be reprogrammed. Even though there may be slightly higher overhead in terms of power consumption, it may not be on the critical path, so this issue can be bypassed. " 
It is helpful to consider the cost of each logic unit in FPGA when weighing various options. Compared to ASIC, you need some architectural design to achieve cost-effectiveness, "Yadavalli said. Compared to ASIC or ASSP, you will always be at a slight disadvantage because their functionality is fixed, carefully designed and optimized to only achieve the functionality you need, rather than letting you create all the programmability yourself. In this case, there will be some additional overhead in implementing Ethernet or PCIe. You can switch all of these functions. You need all these circuits, and this will incur costs. Ultimately, it depends on the return on investment and whether the market has stabilized to the point where you need ASIC. " 

This is about making trade-offs in design to solve specific problems. If you observe FPGA users, you will find that they are trying to complete increasingly complex tasks, "Bauer said. In the field of physical artificial intelligence, people are trying to integrate artificial intelligence into edge systems, obtaining data from sensors and processing it immediately with extremely low latency. Data rates are constantly increasing, and security requirements are also constantly improving. This is an optimization problem

1.Continuously evolving algorithms

AI/ML models are a constantly changing goal, and although FPGAs can be reprogrammed to match them, selecting the best chip is a cyclical process as the situation changes.  
If the workload changes frequently, then I need general-purpose computing, "said Mo Faisal, CEO of Movellus. I can do more optimization and maximize its performance. But ultimately, you will realize that 'this highly flexible general-purpose computing is no longer valuable. I need to customize it. We are in a cycle where my problem statement has changed so much that I may be able to gain more benefits by adopting reprogrammable computing.' Eventually, we will exhaust the potential in this area and go back to the development history of ASIC and FPGA in the past. Of course, you can develop specific algorithms specifically for FPGA. For example, if you want to talk about the scale of particle accelerators, they use 10000 FPGAs to process data output at a specific rate. But if you want to achieve this on a large scale, currently only CERN and Fermilab are available. What if you want 10000 FPGAs? You need to reconsider FPGA. " 
Others also believe that this requires a specific analysis of the situation. Whether to use programmable FPGA depends on the type of workload you are facing, and the same goes for software development, "said Kexun Zhang, research director at ChipAgents. Researchers in the field of life sciences are designing different model architectures. I have seen that the exploration of model architectures in some fields and directions is far from consistent, which is why people still need compilers and hardware to run their customized model architectures at a reasonable speed. FPGA is shining in these fields. " 
For the consumer market and other mass markets, the cost and efficiency structure is not reasonable. In the data center, prototyping or targeting very specific AI changes may benefit from it, "said Nandan Nayampally, Chief Business Officer of Baya Systems. For example, if you engage in high-frequency trading and the algorithm changes regularly, then optimization is necessary. If I don't have enough business volume or cost to support ASIC, then I may insist on using FPGA. Many new AI architectures use programmable components and programmable engines, integrated on the same chip or multiple chips, which can improve efficiency while maintaining flexibility. " 
On FPGA, as many hard wired parallel computing units as possible can be built according to the size of the FPGA, as long as resources permit. Compared to general-purpose GPU cores, FPGAs can be customized according to your application, "Nightingale said," and you can also customize parallelism

2.Virtual FPGA machine for complex algorithms and validation

Cloud based FPGAs can also be used to offload computationally intensive workloads from data centers, for example, by using Amazon Web Services EC2 virtual servers. You can access AWS and configure a virtual machine, "said Russell Klein, Siemens EDA Project Director. This is a physical machine equipped with a so-called F2 instance. It has a built-in PCIe card with eight Xilinx FPGAs, all of which can be programmed with the main processor through the PCIe bus. I have been working with them to program these F1 instances using our advanced synthesis tools and telling them: 'Here is a function that allows us to push it onto the FPGA architecture through the PCIe bus.'. 'We have just implemented this feature, where I can obtain a function, compile it using our advanced synthesis tool, and then interact it with the software running on the processor using Xilinx tools. " 
In addition, F2 instances are also used to offload very complex algorithms. They mainly see this application in the field of life sciences - personnel engaged in DNA analysis or chemical reaction analysis need to handle very complex 3D mathematical operations. They are responsible for programming FPGAs and interacting them with processors running on the host, "Klein said. They can connect all these devices through PCIe connections, thereby accelerating these very complex algorithms. We will see the widespread adoption of this application in the field of artificial intelligence, as it is faster and consumes less energy. This is clearly the next step in building faster and more efficient inference and training environments. This ability is already in place, and we need the entire industry to start utilizing it. " 
Another application scenario for F2 instances is low-cost hardware verification. When SiFive was first established, they needed to boot the operating system on RISC-V design in order to manufacture chips, "Klein said. They want to use this as a validation step. They didn't purchase an FPGA prototype system, but rented FPGA boards on AWS for $6 per hour. They can integrate a CPU instance onto a chip and then program it onto multiple FPGAs and boards to run multi CPU, multi-core designs. They used these data center CPUs to build their own low-cost simulators. AWS is very interested in this and has developed related tools that can utilize the FPGA architecture modeling part of the system for logic simulation. They can combine all these functions to run faster. This is like a hybrid FPGA prototype system. " 
The entire card is connected to a processor on the server. They will insert multiple such cards on a server, each card in a different PCIe slot, so that multiple programs can be connected to multiple cards, "Klein explained. You can start and shut it down very quickly. These are FPGA cards for data centers that can accelerate the development of hardware validation, bioscience, and artificial intelligence. This is a feasible technology to address our current computing challenges. " 
At the same time, FPGA is helping to achieve left shift design. If you look at the explosive growth of new chip development driven by artificial intelligence and the cost of advanced process nodes such as 2nm, you will find that ensuring design correctness is more important than ever, "said Bauer from AMD. The software engineering team can use these simulation prototype platforms to develop corresponding software before the final ASIC chip is taped out

3.Reduce bottlenecks and prepare data

FPGAs can also help reduce memory and I/O bottlenecks in SoCs and chips by optimizing data transmission, thereby saving power consumption and improving performance in data centers.  
Altera has recently chosen Artemis to assist in resolving this issue. FPGAs can be directly placed in the data path to manage data streams, minimizing buffering and optimizing throughput, "Nightingale said. By managing data online, FPGAs can preprocess incoming data, reducing the workload of CPUs, GPUs, or other processing units, thereby alleviating the most significant bottleneck in AI system performance. Therefore, the occurrence of memory and I/O bottlenecks has also been reduced. " 
This is similar to online processing, which involves embedding FPGA directly into the data stream. Due to the large capacity of these devices nowadays, you can perform more processing on the same batch of data while it flows through, "Nightingale said. We see FPGA technology running through the entire data transmission process, closely integrated with data processing. " 
In data centers, FPGAs are used as smart network cards with local memory. You will encounter a very large mesh topology structure that must be able to quickly move data from one point to another, "Bauer said. The reconfigurable nature of FPGA, combined with its ultra high speed connectivity, memory capacity, and low latency, makes it highly suitable for parallel use with AI computing units (whether GPUs or ASICs). In this way, customers can accurately define the location of data movement. We will place the large capacity memory directly next to the computing unit. In AI applications, the proximity between memory and computing units is crucial. This is the high-end application. " 
On the other hand, very small FPGAs are also used on GPU clusters or server motherboards for board level orchestration, management, and power timing control. We call it a server I/O type use case, "Bauer said. The FPGA on the server controls each board level to ensure that all components are powered correctly. In addition, FPGA is responsible for transferring data to memory and communicating between different computing units. " 
Another new role of FPGA is AI infrastructure, which requires processing infrastructure data that enters GPU or CPU. After the data enters the system, the management of these data packets has high programmability, "Yadavalli said. You need a smarter way to manage network interface cards, smart network card functions, or some kind of data plane management, all of which need to be done before data enters the CPU or GPU. The role of FPGA is to receive this data and preprocess it in a way that can be used by all large GPUs and CPUs in the backend. " 
Data preprocessing and cleaning is another emerging field. The effectiveness of artificial intelligence applications depends on the quality of data. You may have the most advanced models or logical layer models (LLM), but if the data itself is noisy, whether it's business data or any commercial data, there will be a situation of 'garbage in, garbage out', "Yadavalli said. Data preprocessing refers to automatically converting all input information into a standard format. You can input PPT, voice commands, text, and so on. On the other hand, the logical layer model can only handle certain specific raw data that requires special processing to provide optimal information. FPGAs are perfectly suited for this job, as they can perform spatial processing on the diverse input data

4.Wireless communication infrastructure

Communication protocols such as 5G, 6G, Open RAN, and baseband applications are important markets for FPGA. Altera and BigCat have recently collaborated to expand FPGA based wireless access network technology. With the evolution of wireless standards, FPGA is typically used for deployment cycles of the first four to five years, "said Yadavalli from Artemis. Due to the incomplete standardization, major equipment suppliers such as Erickson, Nokia, Samsung, and others around the world are unable to develop ASICs in a timely manner. You need to pre order the entire specification two years in advance before you can start developing ASIC. Once the specifications are finalized, you need these chips to be launched in a timely manner in order to transition the network to the next generation of technology. " 
Artificial intelligence algorithms are also rapidly developing in the field of networks, which brings opportunities for FPGAs. In the field of 6G, AI processing technologies related to networks are widely discussed, but there is currently no standardized solution, and people are also full of doubts about the deployment time of 6G, "Yadavalli said. Many people may fall behind, lead, or go astray in certain areas. They need to have flexibility in AI computing. In addition, there are computers of various scales. Some computers can truly push computing to the baseband, where large cloud platforms have already been built. But there are also many technologies deployed on the radio side, as well as baseband equipment located below large antennas. All of these require FPGA as an auxiliary component. " 
In the 5G field, the usage of AMD's adaptive SoC in beamforming applications has significantly increased. This is an important component of standard promotion, "Bauer said.

5.Horizontal product differentiation

Vertically integrated companies can integrate specific accelerators or functionalities into application specific integrated circuits (ASICs), but most of the business in this industry is horizontally integrated. A company that produces various embedded processors is unlikely to develop a fixed accelerator function, as this would narrow its overall market opportunity.  
Siemens Klein stated that these companies can integrate embedded FPGA architectures into their devices or enable their SoCs to connect with external FPGAs. We have seen many manufacturers starting to do this. This can provide customized hardware functionality for their system. Although it is larger in size and consumes more power than ASIC level implementations, it is faster and more energy-efficient than software implementations. " 
For example, if a company wants the battery of a wearable device to last for a week without charging, they must consider transferring some functions from the processor. If they want to do horizontal integration and don't develop SoCs themselves, then they need programmable logic to transfer some functions, "Klein explained. Whether this programmable logic is embedded in the SoC they purchase or used as a standalone component, FPGA can help them achieve higher performance and lower energy consumption. " 
Embedded FPGA can also protect certain intellectual property rights through sparsity and obfuscation. Perhaps a software developer or software program has their own secret algorithm, "said Andy Jaros, Vice President of Intellectual Property Sales at QuickLogic." They want to have a hardware accelerator that matches it. They develop the algorithm and place it in eFPGA on ASIC without sharing it with anyone

6.security threat

Although FPGAs can be reprogrammed to cope with constantly changing regulations and growing threats (such as the ghost of quantum hackers), people are still concerned about emerging security threats.  
We are seeing the increasing use of FPGAs, aimed at ensuring the authenticity of the code running on these systems and preventing tampering, "Bauer said. In addition, for online encryption of data, some customers have their own proprietary technologies, and programmable logic is very suitable for these technologies. Other customers are satisfied with our integrated standard methods. We have a hard encryption module for implementing end-to-end online encryption
Other application scenarios include wireless or wired connections and firewalls. People are using FPGA for artificial intelligence based packet detection to identify threats, "Bauer said. You need a high-speed connection, and then you also need to integrate the model directly into the network architecture
However, there are also some drawbacks. Independent and embedded FPGAs provide great flexibility, but if the confidentiality and integrity management of bitstream configuration data is not handled properly, especially in shared environments such as data centers, this advantage may pose security risks, "said Scott Best, Senior Technical Director of Silicon IP at Rambus. Similar to the risks faced by data center ASICs and SoCs running semi independent virtual machines for different users, data center FPGAs also require strong configuration and access control to prevent unauthorized reprogramming or side channel attacks when users share programmable architectures. " 
Many FPGAs now have built-in encryption systems. The protection measures for bit streams representing FPGA programming have been improved. But they are still the most popular targets for attackers, who attempt to claim that they have invaded the system, modified a bit of the program, transmitted it to the FPGA, and made it run, "said Dana Neustadter, Senior Director of Product Management at Synopsys. He pointed out that attacking FPGA is similar to techniques such as secure boot. In the field of processors, we use secure boot to ensure that the correct program is running on the processor. In the FPGA field, the same task is completed by an encryption and authentication engine, which verifies the programming of the FPGA

7.Conclusion

FPGAs have low initial costs, so they will still serve as prototype design tools, but in today's high-tech field, they are also taking on more and more new roles. In order to keep up with the development of artificial intelligence and resist hacker attacks, designers may consider the best solution to be to mix fixed function devices and programmable devices