FPGA history, current situation and future

Since its inception, FPGA has been entangled with the ASIC community. In the mid-1980s, Ross Freeman and his colleagues purchased the technology from Zilog and founded Xilinx for the ASIC simulation and education market. Zilog comes from ExxonMobil Petroleum Company, and its founding originated in the 1970s when people worried that oil would be depleted within 30 years-even though this statement is still popular today. At almost the same time, Altera with similar technologies as its core was established.

FPGA is a chip that supports circuit programming and realizes "simulation" of the circuit. For the implementation in ASIC, the operating performance of this simulation is slower than the actual circuit. It has a lower clock frequency and higher energy consumption, but it can be reprogrammed every few hundred milliseconds.

FPGA is used to make photomasks at ASIC manufacturer and submit to factory to simulate ASIC before manufacturing. Companies such as Intel and AMD use FPGA simulation chips before chip production.

Competition in the telecommunications field

FPGAs have been widely used in the telecommunications industry. As the continuous changes in telecommunications standards have increased the difficulty of manufacturing telecommunications equipment, companies that can take the lead in providing telecommunications solutions often occupy the largest market share. ASICs have a long manufacturing cycle, and FPGAs provide a shortcut. Telecom equipment began to adopt FPGA in the initial version, which caused the fluctuation of FPGA price. Although the ASIC emulation market is not affected by the price of FPGAs, the price of chips is very important to telecommunications companies. Many years ago, AT&T and Lucent manufactured their own FPGAs called ORCA (optimized reconfigurable cell arrays). But compared with Xilinx or Altera, they do not have a competitive advantage in silicon speed and specifications.

Today, Huawei has become the largest customer of FPGAs. The US-made FPGA may have been the fuse of the recent tension between China and the US. These chips give Huawei an advantage in the delivery of 5G telecommunications equipment, leading any other supplier in the world that is ready to compete for two years.

FPGA price war

FPGAs have long been used for SDR. SDR technology can support radios of multiple communication standards at the same time, similar to a telephone that can speak multiple languages. This time FPGA ran into trouble because SDR technology has embarked on two different adoption paths. On the one hand, commercial suppliers have developed many solutions based on cost-benefit considerations and have deployed SDR technology in all base stations on the planet. On the other hand, in the field of defense, large defense contractors build SDRs to protect profitable traditional product lines. This has caused the price of FPGA-based radio products to remain high, and parts of the US defense market have been resisting their application.

In the next step, FPGA tried to enter the DSP and embedded market development, and began to introduce some FPGAs using hard-core microprocessors. But the pressure to sell these new FPGAs is so great that if customers reject this new series of chips, they will be blacklisted by chip manufacturers, and sometimes even refused to provide services for several months. In view of the frequent failure of FPGA companies to conquer new markets, the growth pressure of the FPGA market is still huge. Because the FPGA chip area is huge and the intellectual property rights involved are numerous, it is difficult to reduce the price of FPGA products.

Hit a wall in HPC and data center

In the past few years, FPGAs have tried to develop in the HPC (high-performance computing) and data center markets. In 2017, Microsoft announced the use of Altera FPGAs in the data center, and Intel acquired Altera. In 2018, Xilinx announced its "Data Center First" strategy. When facing a large number of analysts, its CEO declared that Xilinx is no longer a pure FPGA company. This is quite dramatic, but it is a historical necessity.

The main obstacle to using FPGAs in HPC and data centers is place & route, which is the time it takes to run FPGA vendor-specific software to map circuits to FPGA components. For large FPGAs, using fast CPU servers can take up to three days to place and route. And in many cases, the software still cannot find the mapping after three days.

Hit a wall in the oil and gas field

Around 2007, applications in the oil and gas field formed a niche market. The time it takes to simulate the drilling of the earth to find oil on a traditional computer is longer than the actual construction and drilling time on site. Using FPGA accelerators greatly changed this time-consuming upside-down situation. The first FPGA used in the data center of an oil company to calculate seismic images was manufactured by Maxeler Technologies and delivered to Chevron.

The application of FPGAs in the oil and gas field has been expanded for several years, and it was not until the pressure from the ASIC industry that standard CPU technology returned. At the moment, prediction and simulation are still important in the oil and gas field. Seismic imaging is mostly done using CPU and GPU, but FPGAs still occupy a place. We know that "the current new things will become tomorrow's yellow flowers." Of course, artificial intelligence and the focus on data are current new things.

Nevertheless, FPGA is still a quick way to enter the market, a simple way to gain competitive advantage, and an essential technology in many key tasks. The price of each chip of FPGA is more expensive than that of ASIC, but for HPC and data center, compared with CPU and GPU, it requires fewer FPGA chips and lower cooling overhead. Therefore, the operating cost of FPGA is significantly lower than that of CPU. Or run software on the GPU. FPGAs make data centers smaller, which will make operators feel uneasy because they worry that their data centers may shrink.

ASIC vs. FPGA

Another use of FPGA is as a supplement to ASIC. The purpose of constructing ASIC is to realize fixed functions, and adding FPGA can provide some flexibility for the latest changes of products and adapting to different markets.

Modern FPGAs integrate more and more hard-core functions and become more and more like ASICs. And ASICs often add some FPGA structures to the design to facilitate debugging, testing, field repair, and increase the flexibility of adding small functions.

But the ASIC team has been fighting against the FPGA concept. The ASIC designer asks "What function does the user need?" and loses patience after getting the answer "I'm not sure".

The driverless car industry is such a new battlefield. Due to constant changes in algorithms and laws and regulations that may change when cars enter the market, driving techniques need to be constantly adjusted accordingly, which requires flexible and variable solutions. FPGA has a lower clock frequency, smaller heat sink, and physical size smaller than CPU and GPU. Lower power consumption and smaller size make FPGAs an obvious choice. Nevertheless, the GPU is easier to program and does not require three days of layout and routing.

Another crucial consideration is that for simulation and testing considerations, the same code needs to be run on the car and in the cloud. This requires that the FPGA must be available in the cloud before it can be used in the car. Due to the above problems, many developers prefer to choose GPU.

The evolution of FPGAs

FPGAs are under continuous development. Modern interfaces are making FPGAs easier to program, more modular, and easier to collaborate with other technologies. FPGA supports the AXI (Advanced Extensible Interface, Advanced Extensible Interface) bus, making it easier to program, but it will also introduce a lot of serious efficiency losses, which will reduce the performance of the FPGA and ultimately lead to a decline in its competitiveness. Some academic work has proposed research to solve wiring problems, such as Eric Chung's paper on FPGA dynamic networks, but these advanced concepts have not yet been accepted by the industry.

How are FPGAs connected? For HPC workloads with large data streams, PCI Express can be used and communication hiding technology can be deployed. But what about small-scale workloads like NFV (network function virtualization) that can serve a large number of users at the same time? VMware’s recent survey results point out that for NFV and virtual machine acceleration, FPGAs must be directly connected to CPU, and use cache coherency as a communication mechanism. Of course, a key feature is that the crash of the FPGA does not cause the CPU to crash, and vice versa. Large technology companies are re-examining the requirements of the IBM mainframe era, intending to use standardized platforms to cover increasing complexity.

There are also opportunities in the mass enterprise market. When providing an FPGA platform, even if companies do not have the budget for ASIC development and do not understand the latest silicon manufacturing challenges and solutions, they can develop circuits and establish a competitive advantage in their products. For example, the emerging edge computing of the Internet of Things (IoT) realizes calculations in the proximity of sensors, displays, and even when data flows through.

At the same time, FPGA companies are pushing the technology stack up to the CPU socket. Intel occupies a dominant position in this market, and it has mastered NFV special instructions and other technologies. The main obstacle to adding new CPUs and FPGAs in the data center is not only speed and cost, but also the availability of software and drivers for all possible I/O devices.

The key to implementing FPGAs in the data center is ease of use. For example, use automated tools to drive FPGA applications and avoid placement and routing problems. Microsoft is the first to use FPGAs in large data centers to accelerate Bing, NFV, and artificial intelligence algorithms. It also builds abstract, domain-specific languages and flexible hardware infrastructure. In business, the main problem with FPGAs is the market entry strategy.

It's too late to think about software after building a new chip. How to adapt the hardware to the software and benefit from the existing software? This also provides an opportunity to rethink the FPGA architecture. But there is a caveat: the silicon industry is a golden beast. Building ASIC is a poker game whose stakes have been rising for many years. This is a winner-takes-all game, and the threat of FPGAs is eliminated early in the game.

FPGAs are bringing additional bad risks to silicon projects.

Niche Market

As software designers often say, "What software can accomplish should be realized by software." ASIC designers will say, “What ASIC can do should be done by ASIC.” The most interesting statement is, “If it can be done in software, then you don’t have to deal with all FPGA-thinking people.” Compared with ASIC. The size of the team and the size of software developers worldwide, FPGA company is very small, the community is also very small, of which there are only some even weird programmers.

Intel is advancing the flexibility of FPGAs. Among companies that follow the principle of "build hardware to run existing software," Intel is the most successful company.

FPGA performance may be faster than CPU and GPU, but the tangible lessons from industry and investment circles are that for most of the time since the advent of computers, speed and real-time performance are not that important. Few people buy computers just for high performance. Although this happens from time to time, it is impossible to establish a business market based on such random events. In addition, FPGAs lack standards, no open source code, and no pleasant programming model. Therefore, there is no standard market that supports FPGA programs that can work on all FPGA chips or are easily cross-compiled. Maxeler Technologies has advanced solutions that provide such interfaces, but broad industry adoption requires trust. Trust can promote the development of technology from the plaything of early adopters to benefit everyone, but this requires the promotion and support of existing data center vendors.

In reality, the user of the application will say: "I don't care about the specific method, as long as I can accomplish what I want to do." In the application areas that have not been widely explored, which FPGAs can show their skills? For real-time computing, FPGAs Can be used in industry. For computer vision on drones, FPGAs have advantages in weight and power consumption. Hardware upgrades on satellites are expensive, and FPGAs provide vital long-term flexibility. What FPGA needs are products that share weal and woe. Such products must be easy to program. Hardware or software alone is not enough. An ecosystem and a complete solution are also needed.

Real-time compilation and automatic FPGA program generation are a good way to broaden the current market limitations. It's easy to say, but difficult to do, but with the breakthrough of artificial intelligence in the application space, more and more opportunities have emerged. Currently, everything can be done by artificial intelligence, and even traditional algorithms such as seismic imaging in the oil and gas field use artificial intelligence. Processing artificial intelligence modules requires scientific and engineering solutions. FPGA can provide a good starting point, starting from connecting the AI module and then integrating into the FPGA architecture. For example, Xilinx's next-generation chips integrate artificial intelligence architecture, CPU, 100G interface, and FPGA units into the same 7-nanometer chip.

From another perspective, as artificial intelligence chips generate and process large amounts of data, FPGAs are required to provide large amounts of input and quickly take output away. With the advent of new ASICs for artificial intelligence processing, FPGAs will show their talents in artificial intelligence chip companies.

Prediction

1. A successful CPU+FPGA server chip will appear, or an FPGA that directly accesses the CPU cache hierarchy. Some people agree, some people deny it.

2. SoC FPGA chips will continue to grow and develop, driving the medical, next-generation telecommunications and automotive industries.

3. Developers will use FPGAs to achieve magical things and promote the progress of the world, but the fact that FPGAs exist inside must be kept secret.

4. The name of FPGA will be retained, and chips called FPGAs will continue to appear, but their internals will be very different.

5. Once we give up (data flow) optimization in order to simplify FPGA programming, the performance of the FPGA will decrease and it will not be able to compete with the easy-to-program CPU.

6. FPGA will have dynamic routing, evolving interconnection, and flexible data movement at runtime.

7. Like the complete software stack on FPGA, the place and route software will be open source. Yosys and Lattice FPGA are already working on this.

8. All semiconductor architectures will be combined into a single chip that combines TPU, GPU, CPU, ASIC and FPGA. All technologies can be integrated in the chip, or some technologies can be integrated.

9. More chips will focus on specific application space, and only a few will achieve general purpose. In a sense, everything will become a SoC.