Compute accelerators: GPUs have a serious alternative

FPGA is much older than GPU, this segment was formed in the late eighties of the XX century and was practically monopolized by four manufacturers. Among them are two recognized leaders - Altera, founded in 1983 and bought by Intel in 2015, and Xilinx, founded two years later. As of 2017, they own 31% and 36% of the FPGA market, respectively. There are two other major vendors - Microsemi and Actel, bought by it in 2010.

Ross Freeman, inventor of FPGA (right in the picture) - author of ideas for programming finished products and a semiconductor company that has no own production (fabless)

Historically, FPGA was preceded by Programmable Logic Arrays (PLA) and Complex Programmable Logic Devices (CPLD). The difference between FPGA and PLA and CPLD is both quantitative and qualitative - they did not have sufficient flexibility.

Traditional FPGA arrays consist of many interconnected, fairly simple blocks capable of performing certain logical operations (say, AND and XOR), and routing channels that communicate between blocks. Channel routing and the blocks themselves are programmable, programming in this context is reduced to the creation of LUTs for blocks and routes along which blocks communicate .

In the canonical version, FPGAs consist of millions of identical blocks, the state of which is reflected in the LUT (Lookup Table) tables. The LUT is located in a small programmable memory fragment where the logical function of the block is written. The supervisor connects the blocks and routing to a single system.

Fig. 1 FPGA and LUT units with three inputs

As the FPGA evolves, the block architecture becomes more complex, increasing the number of inputs to 8 or more, allowing more complex logic to be programmed. In the newest FPGAs, blocks are even more complicated, they implement not simple logic, but some specialized functions. Such complex blocks are called slice, that is, a slice or slice. Blocks of this type work faster than assembling their simplest sides. Examples of specialized units include multipliers or DSP signal processors. If multiplication of 32-bit numbers, implemented on simple blocks, requires about 2,000 operations, then a specialized block will require only one operation. The nomenclature of specialized slices is constantly expanding.

Fig. 2 Dedicated FPGAs

FPGAs, along with GPUs, can be used to speed up image processing, cloud computing, broadband communications, big data, robots and others. There are a lot of materials on the network that oppose GPU and FPGA. The results of numerous comparative tests have been published, but it is difficult to be sure that the results obtained are absolutely objective.

As accelerators, both GPU and FPGA have their own advantages and disadvantages that have historical and architectural roots. GPUs come from gaming computers, they have been widely produced for a long time and in large print runs, which provides a relatively low price. Programmable arrays were used most often in military applications, where price is not the main criterion. But FPGAs use transistor mass more efficiently. It follows that the advantage of GPUs in higher performance in terms of price, but FPGAs have significantly better energy efficiency.

An objective comparison of GPU and FPGA in terms of performance is still difficult because they differ in nature and different tests are used for them, and the results are expressed in the case of GPU in well-known flops, while for FPGA in less known maxa (MAKS, Multiply-Accumulate Operations per Second). A broader, comprehensive assessment of nine parameters (Figure 4) gives a score of 5:4 in favor of the GPU, but this score can be interpreted in different ways.

Fig. 3 Comparison of GPU and FPGA efficiency by unit performance of computation per watt and computation per euro

Fig. 4 Comparison of GPU and FPGA efficiency by 9 parameters

Until there is sufficient experience, estimates have to be limited to speculative comparisons. Among the advantages of FPGA is the hardware implementation of algorithms, they are faster and have a shorter delay time, they are measured in nanoseconds, and not microseconds, as in the case of GPUs. The GPU uses a traditional algorithm execution cycle - fetching a command and data from memory, queuing, executing and returning data to memory. On the FPGA side, there is significantly less physical size, power consumption and a greater variety of interfaces.

The advantage of the GPU is greater flexibility, better adaptability to floating point operations, support for older software versions.

However, despite the lack of strict argument, Microsoft and Intel prefer the FPGA. It is possible that the advantage on the FPGA bowl is caused by the significant reduction in the cost of these devices observed recently. FPGAs become available for installation in serial products. Previously, due to the high cost, FPGAs were used only in the development of new systems, at the layout stage, after which the algorithms implemented in them were transferred to ASIC (Application Specific Integrated Circuit) specialized integrated circuits. In serial products, ASIC was already installed, but at the same time they lost the ability to reprogram "in the field." This was the case before, but as of 2017 there is no reason not to pack serial devices with FPGA modules. It is easy to imagine what will follow from the possibility of saving reprogramming, for example, in computers focused on machine learning and other similar applications.

Realizing the importance of FPGA, Intel bought Altera in 2015, after which its vision of processors began to look similar to that shown in Figure 5. Although in 2017 there was a new Xeon Scalable processor family and Purley platform, and the new version of UPI replaced the QPI bus. In addition, a new memory has appeared, but the essence has remained unchanged - there is a CPU and there are FPGAs serving as accelerators.

Fig. 5. Intel processors with FPGAs

Microsoft demonstrates a special attachment to the FPGA, where a project with the ambitious name Catapult has been developed since 2010. The purpose of this projectile is to create a configurable cloud (Configurable Cloud), and the means to achieve this goal is a hyperscale acceleration fabric. Fabric, in this case, translates precisely as a structure, not a factory.

FPGAs in Configurable Cloud are used in two qualities. The first is as channel accelerators, what is called bump-in-the-wire, located between NIC network interface cards by top-of-rack switch (ToR) switches. In the second capacity, FPGAs serve as accelerators in servers designed for a particular class of tasks, such as bioinformatics, search ranking, deep learning, or heavy computing. They use the same servers, but with various programs (firmware) in FPGA.

Fig. 6 Configurable Cloud от Microsoft

Unlike Altera, which merged with Intel, Xilinx takes its own position, it supplies large hyperscalers (Huawei, Baidu) with a product stack (FPGA-powered Xilinx Reconfigurable Acceleration Stack) that supports services called FPGA-as-a-Service.

Fig. 7 Place of Xilinx Reconfigurable Acceleration Stack in the quadrant with axes "Application breadh" and "Accelerator utilization"

Services include libraries, framework integrations, developer boards, and support. OpenStack

Market estimates

2024: Global FPGA Chip Market Size Reaches $12.72 Billion

In 2024, the global user-programmable gate array (FPGA) market reached $12.72 billion. Almost half of the global sales of such products in monetary terms fell on the Asia-Pacific region. The corresponding data are provided in the Fortune Business Insights study, the results of which were published on November 27, 2025.

FPGAs are integrated circuits that can be reprogrammed after production to perform various functions. Such products consist of a large array of logic elements and configurable connections between them. FPGAs are programmed using hardware description languages, such as VHDL or Verilog, which define the logic and structure of the hardware circuit. The configuration can be changed repeatedly, which allows you to configure the device for new tasks.

One of the key drivers of the analytics market is the rapid introduction of artificial intelligence, which is accompanied by an increase in the load on data centers. To meet the needs of computing resources, hyperscalers and cloud providers are forced to actively expand their infrastructure by purchasing accelerators of various types, including those based on FPGA. Such solutions allow parallel processing of data, which allows them to be used for complex calculations, including machine learning. Due to their reconfigurability, FPGAs are a flexible alternative to specialized integrated circuits (ASICs), which are manufactured for one specific task and cannot be modified after production. In addition, FPGAs are capable of processing data at the hardware level, reducing latency compared to conventional processors.

Another stimulating factor is the continued development of 5G networks and preparations for future 6G services. The reprogramming capability allows FPGA to be quickly adapted to new telecommunications standards and protocols, including Open RAN architectures.

The direction of embedded FPGAs, or eFPGAs, is also developing. Such modules are integrated directly into "systems on a chip" (SoC) to increase flexibility. eFPGA is used in systems that require high performance and energy efficiency: these can be Internet of Things (IoT) devices, industrial automation, automotive electronics, etc.

By FPGA application, the market is segmented into telecommunications and network products, HPC data centers and platforms, consumer products and IoT, automotive, industry, aerospace, defense, healthcare, etc. In 2024, the largest share of revenue was provided by the first of the listed areas - $4.04 billion, or 31.9%. From a geographical point of view, the Asia-Pacific region dominates with $6.21 billion, which corresponds to 48.8% of global spending. Globally, the list of major players includes:

• AMD;
• Nvidia;
• Achronix Semiconductor;
• Intel;
• Lattice Semiconductor;
• QuickLogic;
• Gowin Semiconductor;
• Broadcom;
• Synopsys;
• Xilinx;
• Microchip Technology;
• Altera.

In 2025, the FPGA market size is expected to reach $13.92 billion. At the same time, North America will have $3.37 billion, Europe - $2.3 billion. Fortune Business Insights analysts believe that in the future, the CAGR will be 10.2%. Thus, by 2032, costs may increase to $27.51 billion.^[1]

Notes