The era of chip is coming

The driving force of AI

The first electronic devices were made of wood, on which components were connected by wires and pins. They need to be hand-made by professionals, and can produce at most a dozen circuit boards a day.

With the development of technology and the appearance of printed circuit board (PCB), workers with lower technical level can also make circuit boards. The great progress of technology is often accompanied by the increase of product output and the decrease of manufacturing cost.

Artificial intelligence (AI) is a very broad concept, and many engineers have different opinions on its exact meaning. However, it usually occurs at the same time as machine learning (ML). Machine learning allows computers to receive information, analyze information and learn from it, and then continuously improve its response.

For example, we can train the AI system to recognize written words by showing it many different pictures of words and telling it what the words are. This learning capability of AI has been integrated into many products, including safety systems, car safety features, speech recognition, and even content review.

Running simple AI programs (such as word recognition) on devices like smartphones and desktop computers is easy, but what about more advanced systems? How to do face recognition at city level? How to manage the production line and maximize its profit?

With the increasing demand for AI, dedicated hardware for running AI software may become essential. So, how is the current AI system trained?

How does AI train at present?

Small non resource intensive AI systems (handwriting recognition, etc.) can usually be implemented on standard desktop PCs and mobile devices, but their functions are limited because only AI systems with a lot of training will be practical. Here is a typical case of problem AI, which is handwriting recognition. This means that every 100 words are written, the user needs to correct 4 of them. Taking this paragraph as an example, there are more than 100 characters.

A number of GPU based graphics designers may be able to run this task on CPU intensive GPUs. Unlike CPU, GPU is designed to solve complex matrices and polynomials, which is similar to AI systems (especially in image-based AI systems such as face recognition). NVIDIA has recognized the important role of GPU in the AI field. It has launched the fastest and most effective embedded AI module Jetson TX2. The module includes HMP dual Denver 2 / 2MB L2 + Quad arm a57 CPU, 4K x 2K video encoder / decoder and 8 GB 128 bit lpddr4. It also includes various peripherals for hardware developers, including 5 UART ports, 3 SPI ports, 8 I2C ports, 7 USB ports, SATA, Ethernet and HDMI.

Jetson TX2 module

If the machine runs AI locally, the biggest problem is that there is not enough data to learn. This is why many AI solutions are suitable for the cloud; that is, these AI algorithms run on large computers with hundreds of GPUs and CPUs. When millions of users interact with AI, it can learn from each interaction. After learning, it can provide more reliable and accurate results. However, these AI systems do not use devices designed specifically for AI.

AI solutions

Soon after, AI will be more and more common in the market of things. But how will these special equipment be made? Are there any ready-made products?

At present, there are many reasons why there is no special AI silicon device (such as AI CPU). First of all, the AI systems used in practical applications can basically run on general-purpose chips such as CPU and GPU; however, this does not mean that these chips can also cope with future AI problems.

But there are many ways to realize deep learning. This means that hardware solutions for AI applications may need to include elements of dynamic reprogramming and reconfiguration.

Field programmable gate array (FPGA) may be a potential solution to AI problems. An FPGA consists of millions of gates and logic blocks that are isolated and disconnected when the chip is turned off (assuming a volatile FPGA). When the FPGA starts up, it loads configuration data telling it how all the gates and blocks should be connected together. Now FPGA is very powerful, engineers can also configure 32-bit ARM CPU with internal RAM and program memory in FPGA. So what does this have to do with AI?

Neural network is a common method to realize AI, and its construction is similar to the neurons in the brain. The input layer is connected to the hidden processing unit layer using weighted connections, and then to the output layer.

Neural network structure diagram

Unlike neural networks, the brain can change its physical structure by forming new connections between neurons and breaking old ones. However, engineers learn by physically changing the interconnections between logical blocks when designing AI systems using FPGA. One of the main advantages of using logical systems rather than software solutions is that custom hardware is often faster than software implementation, especially when the logical system is asynchronous.

Custom AI chip

Although the market scale of customized AI chips is not large, many manufacturers either have not developed AI hardware or are still in the development stage. However, this has not prevented some companies from developing their own hardware! For example, Google has developed the tensor processing unit, an AI ASIC that specializes in neural networks. The chip, launched on Google I / O 2016, is specifically designed for Google's tensorflow framework. It is worth noting that it is used for machine learning. Unlike CPU and GPU devices, tensor processing units are designed for high-capacity, low-precision computing (as low as 8 bits); however, this may be a valuable solution for future AI devices. When dealing with AI, parallelism is more important than the power of any single CPU; like the human brain, individual neurons are very simple, but when combined with 100 billion other neurons, they make up a person's consciousness.

Of course, Google is not alone in designing its own custom AI chips; Alibaba is also designing its own hardware. In April 2018, Alibaba announced that it was designing an AI chip called Ali NPU. The price of the chip is 40 times lower than that of traditional chips (CPU and GPU), and the performance is 10 times higher than that of CPU and GPU. The reason why Alibaba launched AI hardware was to improve its cloud services. It's worth noting that the company says its chips are ready to serve all users of its cloud network.

The future of AI

Robots replace human physical labor, computers prevent traffic accidents People will rely more and more on computer assistance in their daily life. It is obvious that AI will become an integral part of human society. Since custom AI hardware has not yet been announced, it is unclear how AI will be integrated into the device. But judging from the fact that Internet connections are becoming faster and more widespread, the future AI system may be completely cloud based, that is, a supercomputer in Iceland will solve all the AI problems in the world.