Traditional chips and AI chips

Traditional chips and AI chip
Ai Chip is now a hot field , Compared with traditional chips , There are great differences between algorithms and architectures , Bring infinite creative space to the market , Make many impossibilities possible .
Ai Chip specific algorithms have more advantages than traditional chips
Ai The chip has Al Professional chip of Algorithm , Traditional chips run Ai Algorithm , The performance will be very low , Not able to handle . At the mobile terminal , Perform face recognition 、 intelligence Ai Skin care 、 speech recognition , You have to go through Ai Algorithm development GPU To execute .Ai The algorithm is very different from the traditional chip Algorithm , By convolution 、 Full connection 、 These types of network residuals , Then add and multiply , If the size of the operation graph is determined , The total number of operations can be determined .、

Ai chip NPU Powerful unit , Need a lot of data support
Ai The chip has a built-in network carrier NPU, The computing speed per second has reached 1000 Ten thousand times , Faster than traditional chips 30 times , The speed of processing pictures per minute 2000 Zhang , Ordinary chips can only handle 90 Zhang . because Ai Chips are used in the cloud of data centers and consumer terminals , High requirements for data , Through a large number of data operations to complete various tasks .

Ai Chips are more intelligent than traditional chips , Simulate the computing mechanism of human brain
Traditional chip application software is programming , There is a fixed operation mode , Calculate by executing instructions .Ai A chip is a nerve that mimics the human brain , The basic control system simulates the mechanism of human brain , There is no need to write a large number of fixed programs to solve the calculation ability . Tradition CPU Computation is the mode of instruction , It takes thousands of instructions to complete ,Ai The chip can complete the operation task with only one instruction .
Ai The more intelligent computing power of the chip has subverted the functions of the traditional chip , Replace the position of traditional chips , Give Way Ai The Internet of things 、 Artificial intelligence gets better development .
AI Chip system architecture
The following is the core of the answer to this question .
Tradition CPU In structure , In addition to data operations , You also need to perform data storage and reading 、 Command analysis 、 Branch jump and other commands .AI Algorithms usually need to deal with massive data , Use CPU Execute algorithm , It will take a lot of time , On the reading and analysis of data instructions , Computational efficiency is very low .

With AI Industrial Development , There have been 4 Kind of AI Chip architecture . Yifeng · Neumann is based on traditional computing architecture , Mainly used to accelerate hardware computing power , Yes GPU、FPGA、ASIC 3 Two types represent , The other is to subvert Feng · Neumann architecture , Independent design with brain like nerve structure , Improve computing power . Let's expand in detail 4 There are different types of architectures .

The first is GPU, General image processing unit .GPU use SIMD Single instruction multiple data stream mode , That is, one instruction operates on multiple data , It has a large number of computing units and an ultra long graphics and image processing pipeline , When it was first invented, it mainly processed parallel and accelerated operations in the field of image , because GPU Inside , Most transistors can form all kinds of special circuits 、 Multiple lines , bring GPU Is much faster than CPU, And has a more powerful floating-point computing ability , It can alleviate the training problem of deep learning algorithm , Release AI Potential , It is widely used in the field of deep learning algorithms . It's worth noting that ,GPU Lack of complex arithmetic logic unit , Must be CPU To schedule .

NVIDIA as GPU giant , Occupy 70% of GPU and AI market share . In recent years GTC At the conference ,CEO Huang Renxun's mouth is full of AI, so AI Yes GPU The importance of development is self-evident .

The second is FPGA, Commonly understood as , The hardware design can be repeatedly burned in the programmable memory , send FPGA The chip can perform different hardware designs and functions , So it's called 「 Field programmable logic array 」.FPGA Lock instructions on hardware architecture , Then use the hardware instruction stream to run the data , A simple understanding is to make AI The computing architecture is realized by hardware circuit , Then continuously input the data flow into the system , And complete the calculation . And GPU The difference is ,FPGA It can have hardware pipeline parallel and data parallel processing capability at the same time , It is suitable for processing data stream in hardware pipeline mode , Therefore, it is very suitable for AI The reasoning stage , be relative to CPU And GPU Have obvious performance or energy consumption advantages .

Currently in use FPGA Used to design AI The chip has Shenjian technology in China 、 Microsoft Catapult project . Shenjian technology in 2018 year , With 3 It's sold to FPGA Giant Xilinx .
because FPGA It's hard to program , High requirements for developers , And there it is ASIC, ASIC , Mainly to achieve AI Specific algorithms , Require customized chips . The so-called customization , Just for AI Algorithm , Designed architecture , It can help to improve the chip performance and power consumption ratio , The disadvantage is that the circuit design is customized , Resulting in a relatively long development cycle , Cannot extend beyond , The advantage is in power consumption 、 reliability 、 Chip size 、 Performance and other aspects have great advantages .

since 2016 year Google The release is based on ASIC The first generation of Architecture TPU after , Huawei's shengteng series chips 、 The Cambrian 、 Bit continent 、 Horizon and other manufacturers have entered the game , Even if AI Algorithms are developing rapidly , But based on ASIC Of AI Chip is still the mainstream today .
Some people say , Real AI chips , Future development direction , Could it be a brain chip ？ Finally, let's talk about , Brain like chip Exhibition , Brain like chips are designed directly based on neuromorphological architecture , It is used to simulate the function of human brain for perception 、 Calculation of behavior and thinking mode . But research and development is very difficult .
2014 year ,IBM Launch the second generation TrueNorth chip , use 28nm Process technology , It includes 54 Billion transistors and 4096 A processing core , amount to 100 Ten thousand programmable neurons , as well as 2.56 Billion programmable synapses , The chip works in a way similar to the synergy between neurons and synapses in the human brain .

AI Chip industry chain
Now let's talk about AI chip , It must be inseparable from AI Position of chips in the industrial chain . From the perspective of the overall industrial links of chips , The most upstream is chip design , The midstream is manufacturing and sealed testing , Finally, the downstream system integration and application . But how is the division of labor ？AI In the chip industry chain , Huawei is rising here AI Take industry as an example .

The first is upstream , Rise 910 The chip uses ASIC ASIC , Based on Da Vinci architecture , Da Vinci built this IP Well , It is designed by Huawei Hisilicon , So Haisi is da Vinci's model IP Designer of .
After the design , Just to the middle reaches , Namely AI Wafer manufacturing and packaging testing of chips , But wafers are not just tested in packaging , There will be a test after manufacture , Do it again after packaging . Now, most chip manufacturing depends on Taiwan's TSMC, It's the famous TSMC , And SMIC SIMC Wait for the chip manufacturer .
And finally AI Downstream of the industry , The downstream is mainly system integration and application , Huawei shengteng AI Industry as AI The main integrator of the system set provides shengteng Atlas The server , Then co developers , It's also known as ISV, To provide the upper AI Solution .
AI Future development trend of chips
The last is AI The development trend of chips , Whether it's the Da Vinci architecture of Huawei shengteng products 、 NVIDIA Tensor Core、 still Google, Deep learning requires massive data for calculation , Memory bandwidth constraints , It has become the performance bottleneck of the whole system . The second is massive memory and computing units , Frequent access switching , It is difficult to reduce the overall power consumption . Finally, with AI Rapid changes in industry , How the hardware adapts to the algorithm is a difficult problem .
Let's make a prediction AI Chip 4 Big trends .
future 10 Is a new decade to accelerate the transformation of computing architecture . In terms of computing and storage integration , Put computing units and storage units together , bring AI The computing and data throughput of the system increases , It can also significantly reduce power consumption . Will there be a new type of nonvolatile memory device , Is to add AI Computing function , Save data moving operation ？ Now hardware computing power is greater than data reading and access performance , When the cell is not there, it is the bottleneck , How to reduce memory access delay , It will become the next research direction .
Usually , The closer to computing, the faster the memory speed , The higher the cost per byte , At the same time, the capacity is more limited , Therefore, new storage structures will emerge .

The second trend is , Sparse computing . With hundreds of billions 、 To trillion network model , The model is getting bigger and bigger , But not every neuron , Can effectively activate , At this time, sparse Computing , It can efficiently reduce useless energy efficiency . Especially in the application of recommended scene and graph neural network , Sparsity has become the norm .
for example , Harvard University proposed an optimized five stage pipeline structure for this problem , The trigger signal is output at the last stage . stay Activation Make a pre judgment on the necessity of the next calculation after the layer , If it is found that this is a sparse node , Trigger SKIP The signal , Avoid the power consumption of multiplication , In order to reduce useless power consumption .

The third trend is to support more complex AI operator . In standard SIMD On the basis of ,CNN Special structure reuse , It can reduce the data communication of the bus ,Transformer Structure switches between computing and storage of big data , Or in NLP And voice fields often need to support dynamic shape, All need to be reasonably decomposed 、 Operators that map these different complex structures , To effective hardware has become a research direction .
Finally, faster reasoning delay and storage bit width . With apple 、 qualcomm 、 Huawei is working on mobile phone chips SoC It launches AI Reasoning hardware IP, In recent years, in mobile phones SoC On , It also introduces the learnable function . How to use mobile phones in the future SoC Faster execution on is a point of great concern in the industry , Tiktok including frequent video watching. 、bilibili, You need to do the video AI codec , be based on ISP Conduct AI image processing . In addition, in the field of theoretical calculation , The bit width calculated by neural network 32bit To 16bit, There has been mixing accuracy so far 8bit, Even lower number of bits , Are slowly entering the field of practice .
AI chip , What will ultimately determine success or failure ？ You should choose , NVIDIA GPU The hardware architecture of 、 Huawei Da Vinci architecture 、Google TPU The systolic array architecture ？
in general , stay ZOMI From the point of view of , The choice of chip architecture should serve the success of the whole chip project , It is the result of the game of many factors . NVIDIA can be here today AI The field occupies the head Market , Thanks to the underlying hardware architecture , Or a perfect software and hardware ecosystem ？ This question , I think everyone should see clearly .

Reference link ：
http://www.getfun001.com/net/typeB/91zhuomianA/3492950
https://blog.csdn.net/m0_37046057/article/details/121172739