The potential of edge AI computing for autonomous vehicles

Autonomous driving is an important application of edge computing. Autonomous driving requires 100-1000TOPS of edge AI computing power. Its edge AI (Edge AI) with high performance and low power consumption has become an industry barrier.

AI computing requires a domain to optimize algorithms and data flow architecture. Moore's Law has reached its limit. Without the correct algorithms and architecture, driving performance based solely on processing technology will not achieve the expected results.

The overall edge computing market size is growing rapidly. Image source: IDC

Future computing platform

Category 1: Von Neumann AI Architecture

Harvard University has launched ParaDNN, a parameterized deep learning benchmark suite, which is a systematic and scientific cross-platform benchmark tool that can not only compare the performance of various platforms running various deep learning models, but also support in-depth analysis of cross-model attribute interactions, hardware design and software support.

TPU (Tensor Processing Unit) is a processor built by Google, tailored for machine learning, requiring fewer transistors to perform each operation and more efficiently. TPU is highly optimized for large batches of data for CNN and DNN, and has the highest training throughput.

GPUs exhibit similar performance to TPUs, but have better flexibility and programmability for irregular computations such as small batches and non-MatMul computations.

CPUs achieve the highest FLOPS utilization for RNNs and support the largest models due to their large memory capacity.

Category 2: Non-von Neumann AI architecture

Computing in Memory (CIM): CIM arrays based on SRAM, NAND flash, and emerging memories such as ReRAM, CeRAM, and MRAM are seen as reconfigurable, reprogrammable accelerators for neural network computing. CIM advantages: high performance, high density, low power consumption, and low latency. Current challenges: ADCs for reading bitline analog signal sensing and dedicated RAM processing techniques.

Neuromorphic computing: Neuromorphic computing extends AI to areas that correspond to human cognition, such as explanation and autonomous adaptation. The next generation of AI must be able to handle new situations and abstractions to automate common human activities.

Quantum computing: In quantum computing, the smallest unit of data is a qubit based on magnetic field spin. Based on quantum entanglement, quantum computing allows more than 2 states, and the entanglement speed is very fast (for example: Google Sycamore, Quantum Supremay, 53 Qbits, 1.5 trillion times faster, completing a task that would take a classical computer 10,000 years in 200 seconds). Current challenges: error rates and decoherence in noisy intermediate-scale quantum (NISQ) computers.

Quantum neuromorphic computing: Quantum neuromorphic computing physically implements neural networks in brain-like quantum hardware to speed up computing.

Edge AI and vertical applications

Edge AI will dominate future computing, and AI is a technology that will enable future horizontal and vertical applications.

• Horizontal AI applications solve a wide range of problems across many different industries (e.g., computer vision and speech recognition);

• Vertical AI applications are specific industries that are highly optimized for specific fields (such as high-definition maps, autonomous driving positioning and navigation). With deep domain knowledge, efficient AI models and algorithms can increase computing speed by 10-100,000 times. This is the core and most important autonomous driving technology in the future of artificial intelligence.

All vertical application solutions require multi-level AI models for multiple tasks.

AI Models and Algorithms

DNNs are the foundation of artificial intelligence. Today’s DNNs use a form of learning called back-propagation. Today’s DNNs are slow to train, static after training, and sometimes not flexible enough for practical applications.

Transfer learning is a method of “recycling” a previously developed DNN as the starting point for the DNN to learn a second task. With transfer learning, the DNN model can be trained with less data.

Continuous (lifelong) learning refers to the ability to continuously learn by adapting to new knowledge while retaining previous learning experiences. For example, an autonomous vehicle interacting with its environment needs to learn from its own experience and must be able to gradually acquire, fine-tune, and transfer knowledge over a long period of time.

Reinforced Continuous Learning (RCL) searches for the best neural architecture for each new task through a carefully designed reinforcement learning strategy. The RCL method not only has good performance in preventing catastrophic forgetting, but also adapts well to new tasks.

Automated Driving System (ADS) – Functional block diagram. Image source: ARM

Autonomous driving technology requires breakthroughs:

Edge Precise Positioning and Navigation – Lightweight, fingerprint-based precise positioning and navigation.

Critical real-time response – 20-30 milliseconds, similar to the human brain

Eliminating blind spots – V2X, V2I, DSRC, 5G

Scalable – low power and low cost

Image source: ARM

Autonomous driving requires processing large amounts of data in high-definition maps, positioning, and environmental perception, and all data processed at the edge needs to be completed within critical milliseconds. Intelligent and precise data reduction in perception, positioning, navigation, and enhanced interaction (driving strategy) will enable autonomous driving systems to reduce latency and respond quickly to changing traffic conditions.

Powerful, high-performance edge artificial intelligence (Edge AI) is one of the main barriers in the field of autonomous vehicles. 5G connections support reliable MIMO connections, low latency, and high bandwidth. With the support of 5G, powerful edge AI, coupled with innovations in high-definition maps, positioning, and perception, will make true autonomous driving a reality.