New Technology / Automotive Technology
Unclear topic
The focus on self-driving technology has transitioned from theoretical computing power to user experience and practical scenarios. Effective computing power often falls short of claimed capabilities due to limitations like data storage and processing speed. The SA8650P chip utilizes a multi-layer storage architecture with high-speed caches to enhance computing efficiency. This shift towards effective computing power reflects a broader industry trend prioritizing efficiency over sheer computational capacity.
Source material: As Computing Power Increases, Why Isn't Smart Driving Smarter? [Critique King]
Summary
The focus on self-driving technology has transitioned from theoretical computing power to user experience and practical scenarios. Effective computing power often falls short of claimed capabilities due to limitations like data storage and processing speed. The SA8650P chip utilizes a multi-layer storage architecture with high-speed caches to enhance computing efficiency. This shift towards effective computing power reflects a broader industry trend prioritizing efficiency over sheer computational capacity.
The company aims to develop end-to-end VLA models and world models that can understand the environment like humans, necessitating an exponential increase in effective computing power. This shift emphasizes the importance of utilizing computing power effectively rather than merely increasing its numerical value.
Metrics
performance
32.0 TOP
performance of the chip
This performance metric indicates the chip's capability in processing tasks efficiently.
In April 2014, the Off-Road Plus with the desired control was released, first verifying the 32TOP control chip.
memory_access
1.0 access
memory access requirements
Reducing memory access from three to one significantly decreases bandwidth demand.
Now it only requires one time.
computing_power
0.0
the required increase in effective computing power
This indicates a significant shift in the industry's approach to computing capabilities.
These high-level technologies require exponential growth in computing power.
technology_scope
0.0
the expansion of technology exploration beyond passenger vehicles
This reflects a broader application of technology in various sectors.
Instead, they gradually extend to new fields like heavy-duty autonomous vehicles and L4 Robotaxis.
Key entities
Timeline highlights
00:00–05:00
The focus on self-driving technology has transitioned from theoretical computing power to user experience and practical scenarios. Effective computing power often falls short of claimed capabilities due to limitations like data storage and processing speed.
- The focus on self-driving technology has shifted from absolute computing power metrics, such as TOPs, to the actual user experience and the scenarios covered by the technology
- Theoretical computing power is useful for marketing but does not directly correlate with effective computing power experienced in real-world applications
- Effective computing power is often only 20% to 40% of the claimed capabilities due to limitations such as data storage and processing speed
- Two main barriers to effective computing power in vehicles are the storage wall, which limits data handling capabilities, and the power wall, which restricts chip performance due to overheating
- Companies are exploring methods to optimize computing efficiency, such as using quantization-aware training to maintain model performance while reducing data size
- Qualcomms SA8650P chip exemplifies a multi-core computing platform that integrates various specialized cores, allowing for high efficiency but presenting challenges in development
05:00–10:00
The SA8650P chip utilizes a multi-layer storage architecture with high-speed caches to enhance computing efficiency. This shift towards effective computing power reflects a broader industry trend prioritizing efficiency over sheer computational capacity.
- The SA8650P chip features a multi-layer storage architecture with high-speed caches that significantly reduce the need for DDR memory access, allowing the chip to maintain a continuous computing state and minimize idle time
- Engineers implemented a block-based technique to manage large image data, enabling AI algorithms to process data efficiently within the chips high-speed cache and creating a dedicated high-speed pathway that alleviates pressure on external storage
- A new algorithmic toolchain was developed to streamline deep learning operations, merging three separate steps into one, which reduces memory access requirements from three to one and significantly decreases the demand for memory bandwidth
- The model optimization process involves pruning unnecessary connections within the neural network, enhancing overall efficiency while ensuring that only the most impactful connections remain to improve decision-making capabilities
- To enhance data quality, a comprehensive data selection and cleaning process was implemented, focusing on low-variance data changes, ensuring a diverse and representative dataset for training
- The industry is shifting towards effective computing power over sheer computational capacity, as demonstrated by the successful deployment of advanced driving features in vehicles, reflecting a trend towards maximizing efficiency within existing computational limits
10:00–15:00
The company aims to develop end-to-end VLA models and world models that can understand the environment like humans, necessitating an exponential increase in effective computing power. This shift emphasizes the importance of utilizing computing power effectively rather than merely increasing its numerical value.
- The next step for the company is to develop end-to-end VLA models and world models that not only perceive the environment but also understand it like humans. This requires an exponential increase in effective computing power, focusing on vast amounts of effective computation rather than just numerical values
- The industry has recognized that the upper limits of experience are not solely determined by the size of computing power but by how effectively that power can be utilized. As computing power evolves beyond single chips and specific vehicle models, it becomes a systematic foundation for technology
- The company is building a robust technical foundation using real data to develop mobile intelligent models that integrate perception, decision-making, and control capabilities. This foundation enables stable transitions across various scenarios
- The exploration of technology is expanding beyond just passenger vehicles to include heavy-duty trucks, L4 robotaxis, and L4 logistics vehicles. This shift indicates that high computing power is not the endpoint but rather a solid foundation for further advancements
- To progress towards the next era of computing, both theoretical computing power and advanced algorithms/frameworks must improve together. This dual enhancement is essential for a stable transition into the future