What are the reasons why China's AI chip industry is difficult to change the status quo of dependent survival?
A few days ago, Xilinx, the world's largest programmable chip (FPGA) manufacturer, announced the acquisition of Shenjian Technology, a star startup company in the field of AI chips in China. The news immediately caused a strong response in the industry. Although the two sides believe that this is a win-win outcome, we still see the dependent survival of China's current AI chip industry. What is the reason?
As we all know, the chip defines the basic computing architecture of the industrial chain and ecosystem. Just as the CPU is the core of the IT industry, the chip is also the core of the artificial intelligence industry. As of now, mainstream AI chips recognized in the industry include GPU, FPGA and ASIC in addition to CPU. Those in the industry who are familiar with the chip industry see this and know that the so-called final infrastructure (or genre) of AI chips is nothing more than this. When the above-mentioned infrastructure is based, the pattern has been determined.
It goes without saying that Intel has an absolute lead, and the possibility of basically breaking through on this architecture is minimal.
As for GPU, over 70% of the global GPU industry's market share is currently occupied by Nvidia. The GPU market that can be used for general-purpose computing in the field of artificial intelligence is basically monopolized by NVIDIA. It is reported that there are currently more than 3,000 AI startups in the world, most of which have adopted the hardware platform provided by NVIDIA.
Looking at FPGA again, although its market prospects are attractive, the threshold is unmatched in the chip industry. More than 60 companies around the world have invested billions of dollars to try to reach the top of FPGA highland. Among them, there are many industry giants such as Intel, IBM, Texas Instruments, Motorola, Philips, Toshiba, and Samsung, but only those who succeeded in reaching the top Four companies located in the Silicon Valley of the United States: Xilinx, Altera, LatTIce, and Microsemi. Among them, Xilinx and Altera occupy nearly With a market share of 90% and more than 6000 patents, the technical barriers constituted by so many technical patents are of course unattainable. And Xilinx has always maintained the dominance of the global FPGA.
It is precisely because the chip infrastructure pattern has been determined that the so-called domestic AI chip companies (including start-ups) are actually doing secondary development or optimization based on the above-mentioned basic architecture.
Take Shenjian Technology, which was acquired by Xilinx this time, as an example. Since its establishment in 2016, Shenjian Technology has been developing machine learning solutions based on Xilinx's technology platform. The two companies cooperate closely. Two DPU products of Aristotelian architecture and Cartesian architecture, which are the two underlying architectures of deep learning processors launched by Shenjian Technology, are based on the Xilinx FPGA platform.
In addition, since Xilinx was previously one of the investors of Shenjian Technology, we believe that Shenjian Technology is more like a manufacturer or partner that optimizes Xilinx FPGAs. The reason is simple. Once it leaves the Xilinx FPGA platform, Shenjian Technology will be a tree without roots and a source of water.
Of course, in addition to Shenjian Technology, it is said that the so-called AI chip BPU of another well-known Chinese AI chip start-up company Horizon is also based on secondary development on FPGA. Since it is based on FPGA, the core of the underlying architecture is inseparable from the reference and support of our aforementioned Xilinx, Altera, Lattice and Microsemi FPGA platforms. Even if it is really a disruptive innovation with the core architecture, since FPGA has been divided up by these four companies, it is difficult to have a foothold where it can survive.
I will look at ASIC again. Under the situation that large foreign manufacturers have almost monopolized the CPU, GPU, and FPGA markets, coupled with high technical barriers, Chinese AI chip manufacturers have always lacked key core independent technologies in the chip field. There have been breakthroughs in CPU, GPU and FPGA, and we can only find another way. From the current point of view, Chinese AI chip manufacturers are mostly small and medium-sized companies, combining with actual application needs, focusing on the development of AI ASICs on the device side, optimizing a certain vertical field, and winning with low power consumption and low cost. For example, Cambrian, a well-known AI chip start-up in China, is of this type.
Here we are not saying that ASIC has no prospects in the field of AI chips, on the contrary, Google TPU, which was well-known in the industry before, is based on ASIC. However, it should be noted that the reason why Google developed TPU is based on the application scale of its own data center, and scale is the key to determining the benefits of ASIC.
Although its own large-scale TPU has aroused praise in the industry, Google’s chief scientist Greg Corrado still put forward a different point of view at the Google AI technology sharing meeting held previously. He said, “At least so far, I have not seen the complete Different from the successful cases of traditional computing chips. On the contrary, we believe that existing chips should be specifically optimized in terms of AI, so that the current chips can complete AI tasks faster, with lower power consumption, and higher overall benefits." This is also the reason why Google has a TPU, but still uses CPU and GPU in its data center. The implication is that TPU is only a supplement and optimization for certain applications in the data center relative to the CPU and GPU, and cannot become the mainstream.
Specifically in China, in order to avoid the risks of long ASIC development cycle and large investment, the so-called AI chips based on ASIC development basically adopt the SoC+IP model, that is, compared with ASIC, the SoC+IP model has a shorter time to market and lower cost, and IP can be more Flexible to meet user needs. IP companies focus on the design of IP modules, while SoC companies focus on chip integration, division of labor and cooperation to improve efficiency. Previously, the application of Huawei's Kirin chip and Cambrian IP combined on smartphones belonged to this model. But the premise is scale (huge shipments of Huawei mobile phones) and SoC support. So for the Chinese market, how many users can there be as large as Huawei? ASIC alone can hardly become a forest.
What makes ASIC prospects even more unpredictable is that there is an analysis and view in the industry that FPGAs benefit from the scale effect brought about by the exponential increase in chip NRE costs. With the continuous improvement of the manufacturing process and the exponential increase in the cost of chip NRE, more and more ASIC chips will be forced to abandon because they cannot reach the economies of scale, and thus turn to directly based on FPGA development and design.
According to TracTIca's estimates, as of last year, FPGAs have hardly been found in deep learning applications. However, by 2025, its deployment will be equivalent to the deployment of CPUs (if not more than CPU). As a result, by 2025, FPGAs will gain a significant market share in the deep learning chipset market with a total scale of 12.2 billion US dollars.
The so-called change is inseparable. Although the current names of AI chips vary widely, they are still not separated from the cores of CPU, GPU, FPGA and ASIC. Among these cores, it is obvious that traditional chip manufacturers, such as Intel, Nvidia, Xilinx and other foreign manufacturers, are still in the world. .
And through the Xilinx acquisition of Shenjian Technology, we see that a considerable number of the so-called Chinese AI chip companies are only doing secondary development, optimization and application aspects on top of other people’s architectures. They just changed It’s just a novel name and name. Just like the traditional chip industry competition, the seemingly noisy Chinese AI chip is still a dependent survival model.
UCOAX manufactures a wide spectrum of wire and cable assemblies, each custom-designed for each individual client. Coaxial cable, at the industrial level, is used for instrumentation, control, RF, and microwave applications. It is also commonly used to connect home video equipment and cable television. These cable assemblies are available both rigid and flexible options, and consist of four layers: a round conducting wire, insulating spacer, cylindrical conducting sheath, and a final insulating layer. For any cable fabrication needs, contact UCOAX.
Customized Cables,Custom Pc Cables,Custom Computer Cables,Custom Network Cables
UCOAX , https://www.jsucoax.com