This year, amd successfully completed the huge acquisition of Xilinx, a large FPGA manufacturer, with a transaction value of more than 30billion US dollars, which can be called the “century acquisition” of the chip industry.
So what does Xilinx do and why does amd buy it?
Xilinx is the inventor of FPGA. FPGA can be developed and shaped rapidly. Compared with other standard chips, FPGA does not need a development cycle of several years.
FPGA (field programmable gate array) is a semiconductor device that can be rapidly and dynamically reconfigured. They have some advantages over other types of devices, such as CPUs and GPUs, in a variety of workloads.
It has been 36 years since Xilinx invented FPGA. During this period, numerous large companies want to enter this field. Every ten years, a number of companies withdraw or fall (are acquired) until the FPGA dual giant pattern is finally formed.
As can be seen from the table below, FPGA companies have been disappearing in the past 30 years.
In the current FPGA market, Xilinx and Altera occupy the vast majority of the market, and the rest is distributed to several small companies, the third of which is lattice, which failed to acquire Ziguang that year.
In 2015, Intel purchased Altera, Xilinx’s main competitor FPGA manufacturer, for us $16.7 billion and integrated the company into its programmable Solutions Group (PSG). Since then, Xilinx has become the only child of FPGA company.
Among high-end chips, the top pearl undoubtedly belongs to CPU, GPU and FPGA. Since Intel acquired Altera, NVIDIA acquired mellanox and arm, and AMD acquired Xilinx, the high-end chip field will rapidly form a pattern of three pillars.
Intel’s CPU is very strong, but it can’t compete with AMD with TSMC’s 7Nm blessing within two years. Its GPU will also be released next year. Intel has Altera’s FPGA technology.
At present, AMD’s CPU and GPU are good, and the CPU performance can crush Intel for at least two years. GPU also has a place, the advantage is that it is cheap and sufficient. If Xilinx is successfully acquired, it will enter the high-end FPGA field.
NVIDIA’s GPU is very strong. If arm is successfully acquired, it will get the core technology of cpu/mcu in the field of Internet of things. Mellanox, which he has acquired, will have great advantages in the field of server and storage connection. NVIDIA currently does not have an FPGA, so digesting the current acquisition is the focus of the stage. The latest news is that NVIDIA has failed to acquire arm and has basically withdrawn from the competition in the core CPU market.
So where is the technical motivation for AMD to acquire Xilinx?
The two companies have a history of working closely together on deep learning projects, such as Xilinx deep learning solution on AMD epyc processor. In the data center server field, closely integratedCpu+fpga solutionIt is no longer a new thing.
Let me introduce the history of the fusion of these two things.
CPU is good at control, while FPGA is good at calculation. Therefore, it has long been thought of introducing CPU into FPGA development for logic control to make up for the defects of FPGA.
Altera first proposed SOPC (system on a programmable chip) technology, which uses FPGA logic and resources to build a soft core CPU system, and Xilinx followed up later. Altera’s main soft core is Nios II, while Xilinx’s main soft core is MicroBlaze.
Because they are a set of closed development environment developed by the manufacturer, they have not formed industry standards and are not very convenient to use, so they have not become popular.
With the development of big data, Internet of things and AI, the demand for computing power increases. Xilinx realizes that it is better to let FPGA do peripherals for CPU and directly integrate arm’s cortex-a series processors in FPGA. FPGA is specially used for computing. Everything else is done by powerful cortex-a series processors.
At the embedded system conference held in Silicon Valley in April2010, Xilinx released zynq 7000 series FPGA, which is composed of dual core a9+xilinx 7 Series FPGA. Please note that this is no longer a pure FPGA chip, but an embedded FPGA development board.
Zynq series has achieved great success, making it the mainstream solution for computing power acceleration. After the launch of zynq, many university research groups have used it to complete the deployment of convolutional neural network on FPGA platform, realizing the hardware acceleration of convolutional neural network. The research on FPGA accelerator has become a hot direction in the field of artificial intelligence.
This cpu+fpga mode has also been quickly followed up by Altera. They have also launched a dual core a9+altera FPGA development platform, such as DE1 SOC. Strictly speaking, the cortex-a9 processor is not a single chip microcomputer (i.e. microcontroller, MCU), but a microprocessor (MPU), but both of them are CPUs.
Next, arm became the key player. In 2015, arm finally decided to open source its own Cortex-M0 core. Subsequently, arm opened the Cortex-M3 core on june20,2017 and the cortex-a5 core on october22,2018. With the continuous promotion of designstart plan, arm has gradually established its advantages in the field of Internet of things.
The open source of cortex-m0/m4 core has enabled arm to take over the baton of SOPC from FPGA developers, and the traditional era of FPGA development has come to an end. FPGA development in the new era can be divided into two types: one is that there is a hard core CPU on the development board, which provides a cpu+fpga development environment; The other is that there is no hard core on the development board, but we can build a soft core CPU with FPGA logic and resources by using the Verilog file from arm open source, and we can also build a cpu+fpga development environment.
We can see that the current FPGA has been completely marginalized. In the sales of Xilinx products, the sales of pure FPGA chips are gradually shrinking, while the cpu+fpga is rising.
Considering that in the communication market, such as Huawei base stations, there is still a great demand for pure high-end FPGAs, Xilinx will not give up this market. However, with Huawei’s supply being restricted, Xilinx has excluded Huawei’s orders from the KPI and put its strategic focus on the new direction.
In fact, it is not just CPU. Under the current heterogeneous trend, MCU (MPU), DSP and FPGA have been converged.
A simplified DSP core is added to the M4 series of arm. The Da Vinci series of Ti is an arm+dsp structure, and the FPGA newly launched by Altera and Xilinx contains the core of arm. So the relationship between the three is more and more like the three circles of the three primary colors.
In recent years, AMD’s data center processor business has also been growing, and it is in increasingly fierce competition with Intel, which has long been a dominant player in this field. In the field of server computing power acceleration, both giants have adopted the core architecture of cpu+fpga. The addition of Xilinx will make AMD in a more favorable position in the competition with Intel and occupy a larger share in the rapidly growing Telecom and defense markets.
NVIDIA, another giant, can only strengthen its cloud GPU advantage to compete with its competitors’ cpu+fpga architecture, while Google, another big player in the data center, is quietly cultivating its TPU processor (its parallel processing architecture “pulse array” is similar to FPGA), which should not be underestimated.
In general, it is very clear that pure FPGA chips are moving towards the edge, and FPGA technology itself will not die out. It will be integrated and embedded into the new heterogeneous environment, cooperate with CPU and DSP, and become a necessary weapon in the mainstream market.
Since Altera and Xilinx were acquired, there are no independent FPGA companies in the world. Lattice and other small companies only account for a few market shares, which is not worth mentioning.
So, what is the enlightenment for Chinese local IC enterprises?
The view of AGM microelectronics is that in the future, it is likely that independent FPGA companies will no longer exist. FPGA technology will be deeply integrated with cpu/mcu to become a new mode of heterogeneous computing.
In the past two years, with the support of domestic replacement, the shipments of several major civil FPGA companies have far exceeded that of last year, and the proportion of pure FPGA chips is decreasing, mainly due to the high-energy SOC integrated with arm and FPGA.
AGM is an old fpga/cpld chip enterprise in China. This year, AGM not only launched ag32 series general-purpose MCU compatible with STM32, but also launched MCU 32 series integrated with FPGA, taking the lead in deeply integrating FPGA into cpu/mcu products in China.
If the customer uses both MCU and FPGA chip on the PCB, using this customized chip of AGM can help the customer reduce the purchase cost, save the space and layers of the PCB, reduce power consumption and reduce the manufacturing cost.