I believe you must still remember the Google "Alpha Go" that defeated Lee Sedol and Ke Jie, so do you know what drives Alpha Go?
If you think that Alpha Go is similar to a human, except that the human brain is replaced by a chip, then you are very wrong. The Alpha Go that beat Lee Sedol is equipped with 48 Google AI chips, and these 48 chips are not installed in the Alpha Go body, but in the cloud. So, the device that actually drives the Alpha Go, looks like this...
The picture comes from the Internet, the copyright belongs to the author
Therefore, Li Shishi and Ke Jie lost not to "robots", but to cloud workstations equipped with AI chips.
However, in recent years, the application scenarios of AI technology have begun to shift to mobile devices, such as autonomous driving on cars and face recognition on mobile phones. The needs of the industry have contributed to the advancement of technology, and as the foundation of the industry, AI chips must achieve stronger performance, higher efficiency, and smaller size in order to complete the transfer of AI technology from the cloud to the terminal.
At present, the research and development directions of AI chips are mainly divided into two types: one is FPGA (field programmable gate array) and ASIC (application specific integrated circuit) chips based on the traditional Von Neumann architecture, and the other is designed to imitate the neuron structure of the human brain Brain-like chips. Among them, FPGA and ASIC chips have already formed a certain scale in terms of research and development and application; and although brain-like chips are still in the early stage of research and development, they have great potential and may become the mainstream in the industry in the future.
The main difference between the two development routes is that the former follows the von Neumann architecture, while the latter uses the brain-like architecture. Every computer you see uses a von Neumann architecture. Its core idea is to separate the processor and the memory, so there is a CPU (central processing unit) and memory. The brain-like architecture, as the name suggests, mimics the neuron structure of the human brain, so the CPU, memory, and communication components are all integrated.
Next, Xiaotan will introduce the brief development history, technical characteristics and representative products of the two architectures to the readers.
From GPUs to FPGA and ASIC chips
Before 2007, limited by factors such as algorithms and data at that time, AI did not have a particularly strong demand for chips, and general-purpose CPU chips could provide sufficient computing power. For example, if you are reading this article now, there is a CPU chip in your mobile phone or computer.
After that, due to the rapid development of high-definition video and game industries, GPU (graphics processing unit) chips have achieved rapid development. Because the GPU has more logical operation units for processing data, it is a highly parallel structure, and it has more advantages than the CPU in processing graphics data and complex algorithms, and because AI deep learning has many model parameters, large data scale, and large amount of calculation , GPU replaced CPU for a period of time after that and became the mainstream of AI chips at that time.
GPUs have more logical operation units (ALUs) than CPUs
The picture comes from the Internet, the copyright belongs to the author
However, the GPU is only a graphics processor after all, not a chip dedicated to AI deep learning. Naturally, there are deficiencies. For example, when executing AI applications, the performance of its parallel structure cannot be fully utilized, resulting in high energy consumption.
At the same time, the application of AI technology is increasing day by day, and AI can be seen in fields such as education, medical care, and unmanned driving. However, the high energy consumption of GPU chips cannot meet the needs of the industry, so FPGA chips and ASIC chips are replaced.
So what are the technical characteristics of these two chips? What other representative products are there?
"Universal Chip" FPGA
FPGA (FIELD-PROGRAMMABLE GATE ARRAY), namely "Field Programmable Gate Array", is the product of further development on the basis of PAL, GAL, CPLD and other programmable devices.
FPGA can be understood as a "universal chip". The user defines the connection between these gate circuits and the memory by burning the FPGA configuration file, and uses the hardware description language (HDL) to design the hardware circuit of the FPGA. Every time the programming is completed, the hardware circuit inside the FPGA has a certain connection method and has a certain function. The input data only needs to pass through each gate circuit in turn to get the output result.
In the vernacular, a "universal chip" is a chip that has what functions you need it to have and what functions it can have.
Despite being called a "universal chip," FPGAs are not without flaws. Because of the high flexibility of the FPGA structure, the cost of a single chip in mass production is also higher than that of an ASIC chip, and in terms of performance, the speed and energy consumption of an FPGA chip have also been compromised compared to an ASIC chip.
That is to say, although the "universal chip" is an "all-rounder", its performance is not as good as that of an ASIC chip, and its price is higher than that of an ASIC chip.
However, when the chip demand has not yet reached a large scale and the deep learning algorithm needs to be iteratively improved, FPGA chips with reconfigurable characteristics are more adaptable. Therefore, using FPGA to implement semi-custom artificial intelligence chips is undoubtedly an insurance choice.
At present, the FPGA chip market is divided up by American manufacturers Xilinx and Altera. According to the statistics of foreign media Marketwatch, the former accounts for 50% of the global market share and the latter accounts for about 35%. The two manufacturers occupy 85% of the market share and have more than 6,000 patents. There is no doubt that they are two big mountains in the industry.
Xilinx's FPGA chips are divided into four series from low-end to high-end, namely Spartan, Artix, Kintex, Vertex, and the chip technology also ranges from 45 to 16 nanometers. The higher the chip technology level, the smaller the chip. Among them, Spartan and Artix are mainly aimed at the civilian market, and their applications include unmanned driving, smart home, etc.; Kintex and Vertex are mainly aimed at the military market, and their applications include national defense, aerospace, etc.
Xilinx's Spartan series of FPGA chips
The picture comes from the Internet, the copyright belongs to the author
Let's talk about Xilinx's old rival, Altera. Altera's mainstream FPGA chips are divided into two categories. One focuses on low-cost applications, with medium capacity and performance that can meet general application requirements, such as Cyclone and MAX series; the other focuses on high-performance applications, with large capacity and high performance. Various high-end applications such as Startix and Arria series. Altera's FPGA chips are mainly used in consumer electronics, wireless communications, military aviation and other fields.
Application Specific Integrated Circuit ASIC
Before the large-scale rise of AI industrial applications, using general-purpose chips such as FPGAs suitable for parallel computing to achieve acceleration can avoid the high investment and risk of developing custom chips such as ASICs.
But as we just mentioned, since the original design of general-purpose chips is not specifically for deep learning, FPGAs inevitably have bottlenecks in terms of performance and power consumption. With the expansion of the scale of artificial intelligence applications, such problems will become increasingly prominent. In other words, all our good ideas about artificial intelligence require chips to catch up with the rapid development of artificial intelligence. If the chip cannot keep up, it will become a bottleneck in the development of artificial intelligence.
Therefore, with the rapid development of artificial intelligence algorithms and application fields in recent years, as well as the gradual maturity of research and development achievements and processes, ASIC chips are becoming the mainstream of artificial intelligence computing chip development.
ASIC chips are specialized chips tailored to specific needs. Although it sacrifices versatility, ASIC has advantages over FPGA and GPU chips in terms of performance, power consumption and volume, especially on mobile devices that require chips with high performance, low power consumption, and small size. Like our mobile phones.
However, the high R&D cost of ASIC chips may also pose high risks because of their low generality. However, if you consider market factors, ASIC chips are actually the development trend of the industry.
Why do you say that? Because from servers and computers to driverless cars, drones, and all kinds of home appliances in smart homes, a large number of devices need to introduce artificial intelligence computing capabilities and perception and interaction capabilities. Due to the real-time requirements and the privacy of training data, these capabilities cannot be completely dependent on the cloud, and must be supported by local software and hardware infrastructure. The high performance, low power consumption, and small size of ASIC chips can just meet these needs.
A hundred schools of thought contend in the ASIC chip market
In 2016, Nvidia released the Tesla P100 chip specially designed to accelerate AI computing, and it was upgraded to the Tesla V100 in 2017. When training very large neural network models, Tesla V100 can provide up to 125 trillion tensor calculations per second for deep learning-related model training and inference applications (tensor calculations are the most frequently used calculations in AI deep learning) . However, in the highest performance mode, the power consumption of Tesla V100 reaches 300W. Although the performance is strong, it is undoubtedly a "nuclear bomb" because it consumes too much electricity.
Nvidia Tesla V100 chip
The picture comes from the Internet, the copyright belongs to the author
Also in 2016, Google released the TPU (Tensor Processing Unit) chip to accelerate deep learning, and later upgraded to TPU 2.0 and TPU 3.0. Unlike Nvidia's chips, Google's TPU chips are set up in the cloud, as the article puts it in the Alpha Go example, and "rent only, not sell," with service charged by the hour. However, the performance of Google TPU is also very powerful, the computing power reaches 180 trillion times per second, and the power consumption is only 200w.
Google TPU chip
The picture comes from the Internet, the copyright belongs to the author
Google CEO Sundar Pichai and Nvidia CEO Jen-Hsun Huang have previously debated online about the performance of their respective AI chips. Don't look at the two big guys supporting their own products and fighting for each other. In fact, many netizens pointed out that there is no need to "hardly compare" these two products, because one is in the cloud and the other is in the terminal.
In addition to big companies, startups are also fiercely competing in the ASIC chip market. So how do startups survive in the industry? In this regard, Zhou Bin, CEO of AI chip start-up Novumind in China, told Xiaotan: Innovation is the core competitiveness of start-ups.
In 2017, NovuMind launched its first self-designed AI chip: NovuTensor. This chip uses the Native Tensor Processor as the core architecture, which was independently developed by NovuMind and obtained a US patent in just one year. In addition, NovuTensor chips use different heterogeneous computing modes to deal with 3D tensor computing in different AI application fields. In the second half of 2018, Novumind just launched a new generation of NovuTensor chip. While this chip can achieve 15 trillion calculations per second, the power consumption of the whole chip is controlled at about 15W, and the efficiency is extremely high.
NovuTensor chip from Novumind
Although the paper computing power of the NovuTensor chip is not as good as that of NVIDIA's chip, its computing delay and power consumption are much lower, so it is suitable for edge AI computing, that is, serving the Internet of Things. Although everyone is pursuing high computing power, not all chips actually need high computing power. For example, the chips used in mobile phones and smart glasses have certain requirements for computing power, but they also need low energy consumption. Otherwise, your mobile phone, smart glasses and other products will run out of power after a few uses, which is also very troublesome. one thing. And according to EE Times, when running industry-standard neural network inferences such as ResNet-18, ResNet-34, ResNet70, and VGG16, the throughput and latency of the NovuTensor chip are better than that of NVIDIA's other high-end chip, Xavier.
Combined with the success of Novumind at this stage, it is not difficult to see that: in the cloud market currently dominated by giant companies such as NVIDIA and Google, and the terminal application chips competing for each other, we should focus on technological innovation and significantly lead all competitors in key indicators. Perhaps The survival of AI chip startups.
brain chip
As mentioned at the beginning of the article, all current computers, including all the chips mentioned above, are based on the von Neumann architecture.
However, this architecture is not perfect. The design of separating CPU and memory will lead to the so-called von Neumann bottleneck: the data transfer rate between CPU and memory is very small compared with the capacity of memory and the working efficiency of CPU. Therefore, when the CPU needs to execute some simple instructions on huge data, the data transfer rate becomes a very serious limit to the overall efficiency.
Since artificial intelligence chips are to be developed, some experts return to the problem itself and begin to imitate the structure of the human brain.
There are hundreds of billions of neurons in the human brain, and each neuron is connected to other neurons through thousands of synapses, forming a super huge neuron circuit, transmitting signals in a distributed and concurrent manner, It is equivalent to super-large-scale parallel computing, so the computing power is extremely strong. Another feature of the human brain is that not every part of the brain is working all the time, resulting in low overall energy consumption.
neuron structure
Image source: Wikipedia
This brain-like chip is different from the traditional von Neumann architecture. Its memory, CPU and communication components are fully integrated, using the digital processor as a neuron and the memory as a synapse. In addition, on the brain-like chip, the processing of information is completely performed locally, and since the amount of data processed locally is not large, the bottleneck between the memory and the CPU of traditional computers no longer exists. At the same time, as long as neurons receive pulses from other neurons, these neurons will act at the same time, so neurons can communicate with each other conveniently and quickly.
In the research and development of brain-like chips, IBM is a pioneer in the industry. In 2014, IBM released the TrueNorth brain-like chip. This chip integrates 4096 cores, 1 million "neurons" and 256 million "synapses" in a space of only a few centimeters in diameter. To 70 mW, it is the perfect interpretation of high integration and low power consumption.
DARPA SyNAPSE motherboard with 16 TrueNorth chips
So how does this chip perform in actual combat? The IBM research group has demonstrated using the NeoVision2 Tower dataset from DARPA. It can identify people, bicycles, buses, trucks, etc. in Street View videos in real time at a speed of 30 frames per second, with an accuracy rate of 80%. By contrast, a laptop programmed to complete the same task takes 100 times slower and consumes 10,000 times the energy of an IBM chip.
However, one of the challenges in the development of brain-inspired chips is to imitate the synapses in the human brain at the hardware level, in other words, to design perfect artificial synapses.
In existing brain-like chips, voltages are usually applied to simulate information transmission in neurons. But the problem is that in most artificial synapses made of amorphous materials, there are infinite possibilities for ions to pass through, and it is difficult to predict which path the ions will take, resulting in differences in the current output of different neurons.
In response to this problem, a research team at MIT this year created a brain-like chip with artificial synapses made of silicon germanium and each synapse is about 25 nanometers long. When a voltage was applied to each synapse, all synapses exhibited nearly identical ion currents, with a difference of about 4% between synapses. Compared to synapses made of amorphous materials, their performance was more consistent.
Even so, the brain-like chip is still a long way from the human brain. After all, there are hundreds of billions of neurons in the human brain, and the neurons in the most advanced brain-like chips are only a few million. , even less than 1/10,000 of the human brain. Therefore, the research on such chips still has a long way to go before it becomes a mature technology that can be widely used on a large scale in the market, but in the long run, brain-like chips may bring about a revolution in computing systems.
PVC Power Cable NYY
Standard: VDE0276, VDE0271, IEC 60502
Rated Voltage: 0.6/1kV
Others: Fire Cable and other property Low Voltage Power Cable can be available
Applications: Those LV Power Cables used for electricity supply in fixed installation, power networks,underground and in cable ducting where where mechanical damages are not to be expected.
Nyy Cables,Nyyj Cable,Nyy Flexible Cable,Nyy Power Cable
Shenzhen Bendakang Cables Holding Co., Ltd , https://www.bdkcables.com