In 2015, Intel acquired Altera for $16.7 billion. And just this year AMD officially acquired Xilinx, which is an important milestone for the FPGA field, because Xilinx and Altera are the main suppliers in FPGA. After the two major acquisitions, the industry began to have huge concerns about the future of FPGAs.
The main goal of AMD's acquisition of Xilinx is to build an industry-leading high-performance computing company. According to AMD, "in the various growth markets where Xilinx has established a leading position, it will significantly expand the breadth of AMD's product portfolio and customer set." "But what is the main market Xilinx currently leads?
Xilinx released a powerful FPGA platform called Alveo three years ago. Alveo is the first FPGA accelerator card originally developed by Xilinx to compete with GPU. Prior to this, Xilinx was mainly developing FPGA chips and relied on other vendors to provide FPGA cards. With the emergence of Alveo accelerator cards, Xilinx hopes to provide a powerful platform under its brand name as an accelerator card. Xilinx has developed an impressive ecosystem around the Alveo platform. They use the performance of Alveo cards to accelerate some applications in the fields of machine learning, deep neural networks, databases, natural language processing, genomics, and quantitative finance. The main goal of the Alveo card is to provide a powerful alternative to GPU for deep learning.
In order to use FPGA accelerator cards quickly and efficiently, Xilinx company proposed SDAccel development environment, the purpose is to make CPU/FPGA programming as convenient as CPU/GPU programming. The advancement of FPGA accelerator cards has always been limited by the limitations of hardware programming languages. The SDAccel development environment is to facilitate the use of FPGAs. This platform allows the use of high-level programming languages such as OpenCL and C/c++ to write FPGAs without the need to use VHDL or Verilog. Although the use of OpenCL to develop efficient hardware accelerators requires a deep understanding of FPGA technology, these tools allow software developers without FPGA expertise to develop their own accelerators.
About a month ago, Nvidia announced the acquisition of ARM for $40 billion, vowing to "build the world's top computing company in the era of artificial intelligence." The purpose of the acquisition is to achieve Nvidia's goal to develop a data processing unit (DPU), which includes:
Programmable multi-core CPU (ARM)
High-performance network interface (SmartNIC) (Mellanox)
Rich flexible and programmable acceleration engine (Nvidia GPU)
So after acquiring Xilinx from AMD, what will the future of FPGA look like?
Intel has a rich portfolio of hardware accelerators, including powerful Xeon processors, GPUs, FPGAs and ASICs for deep learning. So what are the main markets for FPGAs in Intel? In a speech at Intel, their FPGA targets were in 3 specific markets:
Natural language processing (BERT)
Fraud detection (LSTM)
Smart city (reasoning).
It seems that Intel is mainly focused on low-latency applications, and FPGAs can provide lower latency than other platforms. Previously, Intel acquired Habana Lab for US$2 billion, and their high-performance deep learning inference and training technology is excellent. However, the difference between these FPGAs for deep learning applications and Habana-based ASIC deep learning platforms remains to be seen.