Autonomous driving is an important application of edge computing. Autonomous driving requires 100-1000TOPS edge AI computing power, and its edge AI (Edge AI) with high performance and low power consumption has become an industry barrier.
AI computing requires domains to optimize algorithms and data flow architectures, and Moore’s Law is approaching its limits. Without the right algorithms and architectures, driving performance based solely on processing technology will not achieve the expected results.
The overall edge computing market is growing rapidly. Image source: IDC
future computing platform
Category 1: Von Neumann AI Architecture
Harvard University has launched ParaDNN, a parametric deep learning benchmark suite, which is a systematic and scientific cross-platform benchmarking tool that can not only compare the performance of various platforms running various deep learning models, but also support cross-platform benchmarking. In-depth analysis of Model attribute interactions, hardware design and software support.
TPU (Tensor Processing Unit, or TPU Tensor Processing Unit) is a processor built by Google, tailored for machine learning, requiring fewer transistors per operation and higher efficiency. TPUs are highly optimized for large batches of data for CNNs and DNNs with the highest training throughput.
GPUs exhibit similar performance to TPUs, but have better flexibility and programmability for irregular computations such as mini-batches and non-MatMul computations.
The CPU achieves the highest FLOPS utilization against RNNs and supports the largest models due to its large memory capacity.
Category 2: Non-von Neumann AI Architectures
Computing in Memory (CIM): CIM arrays based on SRAM, NAND flash, and emerging memories (eg, ReRAM, CeRAM, 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 read-out bit line analog signal sensing and dedicated RAM processing techniques.
Neuromorphic computing: Neuromorphic computing extends AI to areas that correspond to human cognition, such as interpretation and autonomous adaptation. The next generation of AI must be able to handle new situations and abstractions to automate ordinary human activities.
Quantum computing: In quantum computing, the smallest unit of data is a qubit based on the spin of a magnetic field. 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, and it takes 10,000 years for a classical computer to complete an item in 200 seconds. completed tasks). Current challenges: Error rate 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 computation.
Edge AI and Vertical Applications
Edge AI will dominate future computing, and AI is a technology that enables future horizontal and vertical applications.
-
Horizontal AI applications address 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 speed up computation by a factor of 10-100,000. This is the most core and important autonomous driving technology in future artificial intelligence.
All vertical application solutions require multi-level AI models for multi-tasking.
AI models and algorithms
DNNs are the foundation of artificial intelligence, and today’s DNNs use a form of learning called backpropagation. Today’s DNNs are slow to train, static after training, and sometimes inflexible in practical applications.
Transfer learning is a method to “recycle” a previously developed DNN as the starting point for the second task of DNN learning. With transfer learning, the DNN can train the DNN model 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, autonomous driving that interacts with the environment needs to learn from its own experience and must be able to gradually acquire, fine-tune and transfer knowledge over long periods of time.
Reinforcement continuous learning (RCL) finds the optimal neural architecture for each new task through carefully designed reinforcement learning strategies. RCL methods not only have good performance in preventing catastrophic forgetting, but also adapt well to new tasks.
Automated Driving System (ADS) – functional block diagram. Image source: ARM
Autonomous driving technology needs to break through:
Precise positioning and navigation at the edge – lightweight, fingerprint-based precise positioning and navigation.
Critical real-time response – 20-30 milliseconds, similar to the human brain
Eliminate dead zones – V2X, V2I, DSRC, 5G
Scalable – low power and low cost
Image source: ARM
Autonomous driving requires massive amounts of data to be processed in high-definition mapping, localization, and environment perception, all of which needs to be processed at the edge in critical milliseconds. Intelligent and precise data reduction in perception, localization, navigation, and enhanced interaction (driving strategy) will allow autonomous driving systems to reduce latency and respond quickly to changing traffic conditions.
Powerful, high-performance edge AI (Edge AI) is one of the main barriers to autonomous vehicles. 5G connectivity supports reliable MIMO connectivity, low latency, and high bandwidth. Powered by 5G, powerful edge AI, coupled with innovations in high-definition maps, positioning, and perception, will make true autonomous driving a reality.
The Links: CM400DY-12NF SKIM304GD12T4D