Novumind heterogeneous Intelligence: artificial intelligence is redefining edge computing. The development of IOT is overwhelming, and intelligent edge computing has also become the hottest market field. AI + aiot + edge computing will break the disadvantages of traditional IOT data centers that are not on edge devices, and enable AI to become smarter and more agile on edge devices to meet the needs of more application scenarios. Local computing on edge devices means that a large amount of data needs to be effectively reasoned, analyzed and extracted. The demand for computing power is very high, and low latency and low power consumption are required
in the keynote speech of the 2019 SV connect summit held recently, Dr. Wu Ren, founder and CEO of heterogeneous intelligence, shared the view of how artificial intelligence can achieve high computing power and low power consumption at the edge of the network. He believes that the semiconductor industry is constantly expanding to the field of artificial intelligence, and the traditional chip architecture can no longer meet the various application needs of artificial intelligence, especially in the application of edge devices
Dr. Wu Ren's keynote speech on SV connect 2019
making a special domain architecture for artificial intelligence
"the era of general-purpose processors is coming to an end, and the improvement of computing performance in the future will take the special processor architecture (DSA) as a breakthrough, which is suitable for machine learning and inference neural network processors." Dr. Wu also quoted the view of Hennessy and Patterson, the Turing prize winners, "DSA is the only path to move forward". DSA is a form of heterogeneous computing that makes the piston drop for a distance. The direct interpretation of heterogeneous computing refers to the combination of several processors with different architectures for computing. The core idea is to use professional hardware to do professional things. DSA aims to optimize specific areas and give full play to the highest effectiveness of hardware performance. In order to better realize man-machine running in in the process of operation
traditionally, deep learning technology relies on high-performance GPU to meet the needs of high computing power. However, when it comes to specific consumer market requirements, there are often more stringent power consumption and cost constraints. Especially when it comes to the application of artificial intelligence in different scenarios, the requirement for hardware is often to meet accurate goals. "Traditional chip architecture requires matrix expansion for artificial intelligence computing such as neural networks, which will lead to a lot of unnecessary computing and power consumption. Our independently designed novutensor chip architecture supports native tensor computing, and can achieve a power consumption ratio far higher than that of other common architectures while ensuring high computing power." Dr. Wu Ren also mentioned that, "Many chip companies mention that their products have high peak computing power in testing the mechanical properties of materials, but they don't mention the accuracy of calculation. Behind many high computing power and low power consumption, it is often the sacrifice of calculation accuracy. Such products undoubtedly can't stand the test of the market. Novutensor uses the original dynamic half precision floating-point calculation, which has little accuracy loss compared with the standard half precision floating-point number, but can greatly save hardware costs."
doing business with heterogeneous computing thinking is our own everywhere. With the advent of the era of
ai, the demand for solutions is further clarified and growing. Dr. Wu believes that AI is an enabling industry. The vision of heterogeneous intelligence is to use AI to help people in different fields, which will also greatly reduce its use value. Enterprises will do their products better. Whether it is intelligent security cameras, smart homes, smart factories, data centers, etc., all need cost-effective computing engines that can process large amounts of data and video images. "The core concept of heterogeneous computing is to give full play to the advantages of processors with different architectures and let those who are good at doing what they are good at, and so are heterogeneous intelligent products." Dr. Wu said, "what we are good at is to help different industries empower AI technology, so that their products can be closer to market demand and better serve their customers."
with the development of artificial intelligence technology, both cloud applications and terminal applications will become more mature. The two are not antagonistic to each other, but complementary and win-win. The advantage of heterogeneous intelligence is that it can provide AI full stack services, from the cloud to the edge, and then from the support of underlying hardware to model training. In the future, with the progress of computing power and the development of the market, heterogeneous intelligence will play an important role in the wave of artificial intelligence