Autonomous driving is a significant application of edge computing, requiring 100-1000 TOPS of edge AI computing power, which has become an industry barrier due to its high performance and low power consumption characteristics.
AI computing requires domains to optimize algorithms and data flow architectures. The limits of Moore’s Law are approaching, and without the correct algorithms and architectures, performance driven solely by processing technology will not achieve the expected results.
The overall edge computing market is experiencing rapid growth. Image source: IDC
Future Computing Platforms
Type 1: Von Neumann AI Architecture
-
Harvard University has launched the parameterized deep learning benchmark suite ParaDNN, a systematic and scientific cross-platform benchmarking tool that not only compares the performance of various platforms running different deep learning models but also supports in-depth analysis of cross-model interactions, hardware design, and software support.
-
The TPU (Tensor Processing Unit) is a processor developed by Google, specifically designed for machine learning, requiring fewer transistors for each operation and achieving higher efficiency. The TPU is highly optimized for large batches of data in CNNs and DNNs, providing the highest training throughput.
-
GPUs exhibit performance similar to TPUs but offer 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.
Type 2: Non-Von Neumann AI Architecture
-
Compute-in-Memory (CIM): CIM arrays based on SRAM, NAND flash, and emerging memories (such as ReRAM, CeRAM, MRAM) are viewed as reconfigurable and reprogrammable accelerators for neural network computing. Advantages of CIM include high performance, high density, low power consumption, and low latency. Current challenges include readout bitline analog signal sensing and dedicated RAM processing technology’s ADC.
-
Neuromorphic Computing: Neuromorphic computing extends AI into areas corresponding to human cognition, such as interpretation and autonomous adaptation. Next-generation 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. Quantum computing allows for more than two states based on quantum entanglement, with entanglement speeds being extremely fast (e.g., Google Sycamore, Quantum Supremacy, 53 qubits, 15 trillion times faster, completing a task in 200 seconds that would take classical computers 10,000 years). Current challenges include 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 accelerate computation speed.
Edge AI and Vertical Applications
-
Edge AI will dominate future computing, as AI is a technology that can achieve future horizontal and vertical applications.
- Horizontal AI applications address a wide range of issues across various industries (e.g., computer vision and speech recognition); vertical AI applications are highly optimized for specific industries (e.g., high-definition maps, autonomous driving positioning and navigation).
- With deep domain knowledge, efficient AI models and algorithms can increase computation speed by 10-100,000 times. This is the core and most important technology for autonomous driving in future AI.
-
All vertical application solutions require multi-level AI models for multitasking.
AI Models and Algorithms
-
DNNs are the foundation of AI, and today’s DNNs use a learning form called backpropagation. Current DNN training speeds are slow, and after training, they are static, sometimes unable to adapt flexibly in practical applications.
-
Transfer learning is a method of “recycling” previously developed DNNs as a starting point for learning a second task, allowing DNNs to train models with less data.
-
Continual (lifelong) learning refers to the ability to continuously learn by adapting to new knowledge while retaining previous learning experiences. For example, autonomous driving interacting with the environment must learn from its experiences and gradually acquire, fine-tune, and transfer knowledge over time.
-
Reinforcement Continual Learning (RCL) seeks the optimal neural structure for each new task through carefully designed reinforcement learning strategies. The RCL approach not only performs well in preventing catastrophic forgetting but also adapts well to new tasks.
Autonomous Driving System (ADS) – Functional Block Diagram. Image source: ARM
Breakthroughs Needed in Autonomous Driving Technology:
-
Edge precise positioning and navigation – lightweight, fingerprint-based precise positioning and navigation.
-
Critical real-time response – 20-30 milliseconds, similar to the human brain.
-
Elimination of blind spots – V2X, V2I, DSRC, 5G.
-
Upgradable – low power consumption and low cost.
Image source: ARM
Autonomous driving requires processing large amounts of data in high-definition maps, positioning, and environmental perception, with all edge processing needing to be completed within critical milliseconds. Smartly reducing data in perception, positioning, navigation, and reinforcement interaction (driving strategy) will shorten latency for autonomous driving systems and enable rapid responses to changing traffic conditions.
Powerful, high-performance edge AI is one of the main barriers in the field of autonomous vehicles. 5G connectivity supports reliable MIMO connections, low latency, and high bandwidth. With the support of 5G, powerful edge AI, along with innovations in high-definition maps, positioning, and perception, will make true autonomous driving a reality.
