Trust the Machine — A New Faster but Safe Processor for Artificial Intelligence
Being compliant with the highest safety standards and tailored for high performance in AI applications such as self-driving cars, a novel processor architecture shows much promise
Daejeon, July 17, 2020
Two of the reasons that artificial intelligence (AI) applications have not seen a boom are low processor speeds and safety considerations. Now, researchers from Korea present a new processor architecture tailored for enhancing both the performance and safety of neural networks, which are the bread and butter of modern AI. Their design is highly reliable and could easily find a home in self-driving cars and other technologies such as mobile phones.
Over the last decade or so, artificial intelligence (AI) has taken the world by storm. Of all its varied applications — such as big-data analysis, advertising, surveillance, smart devices, etc. — self-driving cars seem to be gaining considerable spotlight. However, a standard processor is not enough to get an autonomous vehicle up and running safely. The computers on-board self-driving cars must not only operate at adequate speeds to ensure that calculations are made in a timely manner, they also require mechanisms that ensure errors are detected to guarantee the safety of its occupants and others.
Advanced AI applications like autonomous vehicles need special processor architectures that are tailor-made with these issues in mind. And researchers from the Electronics and Telecommunications Research Institute (ETRI) in Korea have designed a novel AI-oriented processor that could just be a solution. Dr. Jinho Han, who led the study, says: “Our design is aimed at increasing the processing speed for artificial intelligence, which is one of the obstacles to the proliferation of AI-based services.”
The layout of the processor architecture was developed considering the fundamental operations required by neural networks, the most commonly used type of algorithm in AI. The processor has a ‘super-thread core’ composed of a matrix of 128 by 128 nanocores that are connected to memory modules. These memory modules contain all the inputs necessary for the neural networks to perform their calculations. The integration of the memory modules into the super-thread core greatly reduces the number of required reads to the external memory, thus improving performance. In addition, the nanocore matrix provides much flexibility for multi-thread processing.
As for safety, the architecture allows the processor to meet the highest level of the international ISO 26262 standard, which is related to the functional safety of road vehicles. Dr. Han explains: “In automotive applications, processors have to be fault-tolerant to avoid running into problems due to voltage fluctuations, temperature variations, and even radiation. The processors for advanced driver-assistance systems and autonomous cars should be extremely robust and stable during operation to guarantee both safety and convenience.”
This high level of compliance with the standard is partly achieved through multiple specific considerations in the design of the architecture’s general-purpose processor. For one, its cache — the internal memory that stores frequently used data — is designed to be self-recovering. It can also discover faults in itself and take the necessary actions to correct them. Moreover, the processor has ‘dual modular redundancy’, meaning that it has every component in duplicate, running the same instruction in parallel, so that faults can be detected by simply comparing both outputs.
These and other safety features make this new architecture more reliable than existing solutions. When Dr. Han and his team tested an actual prototype of it, on which errors could be injected as desired into different parts of the system, the robustness of their design was experimentally proved.
Overall, the proposed AI-oriented architecture provides the best safety to date without incurring a loss in performance. Its high speed and compact design should make it a welcome addition in emerging AI applications, which will hopefully make the world a safer and more convenient place.
Reference
Titles of original paper:
Names of authors: Jinho Han*1, Minseok Choi 1, Youngsu Kwon2
Affiliations:
1 AI Processor Research Section, ETRI
2 AI SoC Research Division, ETRI
About Dr. Jinho Han
Dr. Jinho Han studied at KAIST, Korea, where he earned Bachelor’s and Master’s degrees, as well as a PhD, in Electrical and Electronics Engineering in 1998, 2001, and 2020, respectively. Since 2001, he has worked at ETRI in the fields of system architecture, processor design, system-on-chip designs, and, more recently, AI processors for autonomous cars. He is currently a Principal Researcher at ETRI.
Media contact:
E-mail: soc@etri.re.kr
Office Number: +82 42–860–6558
