The title "1" does not distort the original words and context of the experts.
Why do you maliciously say that a local expertise in China is ahead of the world? Some people just can't stand China, okay? If you are in doubt, please show evidence to convince everyone.
This expert is obviously full of confidence in his work and publicizes China's achievements. What is the malice of the media? Why is it asserted that this expert's sentence is not good?
The following is the latest interpretation article published by Observer Network:
Liu Tie: CAMBRIAN processor is the technical achievement of many years' efforts of China Academy of Sciences.
Recently, I interviewed researcher Chen of Institute of Computing Technology, Chinese Academy of Sciences, and reported the instruction set DianNaoYu of deep learning processor and CAMBRIAN neural network processor. Many readers are happy that "China's smart chips lead the world", but some readers have raised doubts-some readers suspect that CAMBRIAN processors are marketing hype; Some readers think that "the Cambrian will be over as soon as NVIDIA takes the shot"; Other readers think that DianNaoYu, an independent instruction set, is too easy. The Cambrian didn't jump out of the category of traditional chips and couldn't simulate the synapses of the brain. Only IBM's "True North" is the real neural network processor.
In view of this, the author consulted Dr. Chen of Beijing Zhongke Cambrian Technology Co., Ltd. and made some clarifications according to his own views.
Cambrian breakthrough of classical von Neumann structure
Artificial neural network is the general name of a kind of computer algorithm that imitates the construction of biological neural network, and it is composed of several artificial neuron nodes (hereinafter referred to as "neurons"). Neurons are connected by synapses, which record the strength (weight) of connections between neurons.
Each neuron can be abstracted as an excitation function, and the input of this function is determined by the output of the neuron connected to it and the synapse connecting the neuron. In order to express specific knowledge, users usually need to adjust (through some specific algorithms) the values of synapses in artificial neural networks and the topological structure of the networks. This process is called "learning". After learning, artificial neural network can solve specific problems through the acquired knowledge.
In order to make the neural network algorithm convenient for ordinary people to use through cloud services, mobile phones and other carriers, computer hardware has become a key bottleneck-Google's neural network for cat face recognition is slow to train and use, and consumes a lot of computing resources, which ordinary users can't afford; Advertising recommendation applications must calculate the products that end users may care about within 100 milliseconds or even less. For many large-scale deep neural network calculations, it is impossible for current CPU and GPU to achieve this speed.
Therefore, traditional processors (including x86 and ARM chips) are not efficient in processing deep learning, and it is necessary to find another way to break through the classic von Neumann structure.
When training, the neural network can realize the induction and summary of the existing knowledge by automatically adjusting the weight of synapses between neurons. When using, the output result corresponding to the current input can be calculated according to the current synaptic weight. In other words, storage and processing in neural networks are integrated, and they are all reflected by synaptic weights.
In Von Neumann's structure, storage and processing are separated and realized by memory and arithmetic unit respectively. There is a huge difference between the two. When using the existing classic computers based on Von Neumann architecture (such as X86 processor and NVIDIA GPU) to run neural network applications, it is inevitably restricted by the separation of storage and processing structures, and the efficiency is low.
This is the basis of developing a neural network processor dedicated to artificial intelligence.
CAMBRIAN processors are not hype marketing.
Some readers questioned that the CAMBRIAN processor was beating Li Shishi with an Alpha dog to hype itself, which was commercial packaging and hype marketing. But in fact, CAMBRIAN processor is the technical achievement of many years' efforts in computing by Chinese Academy of Sciences.
As early as 2008, the Institute of Computing Technology of Chinese Academy of Sciences took the lead in developing CAMBRIAN series deep neural network processors. The related work has won the best paper award from the top conferences in the field of computer hardware: ASPLOS' 14 and MICRO' 14. This is also the first time that Asia has won the best paper award in the top conference in this field.
CAMBRIAN 1 was also selected as the research hotspot in the computer field in 20 14 by the newsletter of the Computing Machinery Association (about 20 articles per year, the mainland was selected for the first time). This indicates that China has entered the international leading ranks in the field of brain like computing.
In addition, DianNaoYu, the instruction set of deep learning processor, was accepted by ISCA 20 16 (International Symposium on Computer Architecture), ranking first among all nearly 300 submissions.
So far, Dr. Chen and researcher Chen Yunyong have gained two ASPLOS, two ISCA, one MICRO and one HPCA in the technical achievements of the photoCAMBRIAN series. This is the four top international conferences on computer architecture, but only the scientific research community pays attention to it, and ordinary people don't understand its significance.
Therefore, CAMBRIAN processor was not born in the hot spot of Alpha Dog vs. Li Shishi, but the result of long-term technology accumulation, which won numerous awards at high-end international conferences a few years ago. It has never been seen in the mainstream media before. It's just that domestic media people are more willing to report the so-called "high technology" abroad. The so-called "high-tech" is full of articles and reports, even if the size of a certain brand's mobile phone has increased a little. All kinds of praises are endless, but the significance of top conferences in many professional fields is unclear, which leads to a blind eye to the truly valuable independent technology selectivity.
The meaning of computer language in autonomous instruction set
An instruction set is a collection of codes. It is a set of command standards that use some codes to express reading and writing operations and command computers to do various operations.
The technical difficulty of redefining the instruction set is close to zero. In the United States, it was forbidden to apply for a single instruction set as a patent, and only by combining the instruction set with the implementation method was it allowed to apply for a patent. However, the influence of instruction set on integrated circuit design can not be ignored. For example, it is impossible to design high-performance chips with ARM's incomplete instruction set. In recent years, after ARM updated its instruction set and purchased MIPS 498 instruction set license, it was able to show its strength in the field of high-performance chips.
The value of the simple instruction set itself is very limited. What is really valuable is the software ecology built around instruction set, the implementation method of instruction set and the power to freely expand instruction set. Everyone knows the software ecology, so I won't elaborate on it, focusing on the right to develop instruction sets.
For example, Huawei bought an ARM instruction set license, but it didn't have the power to independently expand the instruction set, which led to profits being controlled by others, just like a joint venture car factory transfused foreign capital-buying an ARM instruction set license is not only expensive, but also has a license period of only five years and a limited scope of use. The most important thing is "endless purchase"-buy when the instruction set expires, and update the instruction set.
In addition, the ARM instruction set can't be changed except the profit from purchasing the instruction set license. Even if it is modified at the risk of infringing intellectual property rights, it cannot be supported by AA software vendors, which is equivalent to an invalid extension of the instruction set.
In sharp contrast, Loongson. Godson has obtained the permanent authorization of MIPS, which avoids the embarrassment of "endless purchases" and is not subject to others in terms of profits.
More importantly, it has the right of independent expansion-godson ISA has 1907, MIPS has 527, and the rest instructions are independently expanded by godson. Due to the self-built ecology, the self-expanding instruction set can naturally be supported by software, and it has also been supported by many foreign open source software. Godson can reach the original peak of special application through its own extended vector instruction. Moreover, the authorization to purchase the ARM instruction set can only be what ARM sells and what you use; It is impossible for IC design companies that buy ARM instruction set to expand the instruction set independently by Godson and quadruple the peak value of specific applications.
It can be seen that the significance of independent instruction set DianNaoYu is that profit is not subject to people and development is not subject to people.
Although DianNaoYu will have no influence on X86 and ARM, Harvard, Stanford, MIT, Columbia and IBM in the United States are all making neural network processors. It can be predicted that in the neural network processor, there will be a similar battle among Alhpa, MIPS, SPARC, X86 and Power, and the final winner will gain a position similar to X86 on the desktop chip, and then gain excess profits.
About "NVIDIA finished the Cambrian as soon as he took the shot"
Neural network processor and general processor, DSP, FPGA and GPU are different computing devices, just as the existence of GPU will not make DSP die, neural network processor and GPU do not conflict. Of course, if NVIDIA makes a better neural network processor than CAMBRIAN, CAMBRIAN will probably be affected.
In addition, NVIDIA has been striving to enter the field of intelligence, but its graphics processing architecture is far from that of neural network processing, and it will have an energy consumption disadvantage of more than 100 times compared with Cambrian. Take Diannao and Bignaonao as examples. Diannao is a single-core processor with a main frequency of 0.98GHz, peak performance of 452 billion basic neural network operations per second, power consumption of 0.485W, and an area of 3.02mm^2 under 65nm technology. The experimental results on several representative neural networks show that the average performance of DianNao is 100 times higher than that of mainstream CPU cores, while the area and power consumption are only110, and the efficiency can be improved by three orders of magnitude. The average performance of DianNao is comparable to that of mainstream GPGPU, but its area and power consumption are only 1% of that of mainstream GPGPU. On the basis of Dot Brain, Big Dot Brain further expands the processor scale, including 16 processor cores and larger on-chip storage, and supports direct high-speed interconnection among multiple processor chips, thus avoiding high memory access overhead.
Under 28nm technology, the main frequency of Dadian brain is 606MHz, the area is 67.7mm^2, and the power consumption is about 16W. The performance of single chip is 2 1 times higher than that of mainstream GPU, while the energy consumption is only 1/330 of that of mainstream GPU. Compared with the mainstream GPU, the performance of the 64-chip high-performance computing system can even be improved by 450 times, but the total energy consumption is only1150.
Therefore, this sentence is as meaningless as "NVIDIA will be finished as soon as Intel makes a move".
About CAMBRIAN replacing Intel
Although CAMBRIAN processor has great market potential in artificial intelligence, it is not a subversion of traditional CPU. As far as the current technology is concerned, the neural network chip can't be better than the traditional CPU in all application fields, but it has advantages over the traditional CPU only in the field of artificial intelligence, and it is more like a special chip.
Even if mobile phones and PCs use heterogeneous computing technology in the future, the role of CAMBRIAN processors is only to complete intelligent cognition and other functions, such as running databases, scientific computing, office, WeChat and so on. Cambrian processor can't do better than existing traditional CPU. Because even the human brain itself, many things can't be done by CPU.
Therefore, the traditional CPU will be the core of the computer in the future, but it is only a key, and a lot of time-consuming and laborious work will be handed over to other computing devices, such as DSP, FPGA, GPU and CAMBRIAN chip (neural network chip). If someone wants CAMBRIAN chips to replace Intel chips on home PCs, it can only be an unrealistic fantasy.
On the possible difficulties in CAMBRIAN
The risk of CAMBRIAN processor failure lies in the excessive hype of artificial intelligence by society and media. If the development speed of artificial intelligence can't reach the expectations of the public (investors) (which is bound to happen, for example, many media and even Google itself have said that Skynet will be built), then the whole field will fall into a big trough and there will be no eggs under the nest. This kind of thing happened once in the 1980s.
Although the road of CAMBRIAN may not be smooth sailing, the future is bright-because CAMBRIAN is completely different from traditional processors such as Godson and Shenwei. In the words of Dr. Chen from Beijing Zhongke Cambrian Technology Co., Ltd., "We used to be followers and accidentally fell into someone else's pit and were blocked by someone else's barriers. Now we are the leaders. Without intellectual property barriers, we have a vast sky. At present, the Cambrian spends a lot of money on patents, but it is only to set barriers for followers. "
As long as the artificial intelligence industry develops healthily, the electronics industry has strong support, and the marketing can not be inferior to foreign technology companies, the future of Cambrian is very worthy of Chinese people's expectation, so please wait and see.