AI and Chips: Trends and Scenarios
Several discussions have summarized the trends and scenarios of the integration of AI and chips. - Three major trends: moving from general-purpose to specialized, integrating storage and computing to address the memory wall, and optimizing through hardware-software collaboration. - Application scenarios: embodied intelligence (robots/drones), edge-side multimodality (phones/cars/glasses), AI-native scientific computing, and edge intelligent networks. - It highlights the crucial role of chips in the next 3-5 years for embodied intelligence and edge deployment, listing domestic and international companies and technological pathways.
Key insights
Chip Development is Shifting from General-purpose GPUs to Specialized AI Chips
Discussions indicate that the demand for training, inference, and edge-side chips differs, leading to a future differentiation of training chips, inference chips, and low-power edge-side chips. General-purpose GPUs will no longer meet all scenarios, and specialized chips will become key to driving embodied intelligence and edge multimodal deployment.
Integrated Storage and Computing is Seen as the Key Path to Solve the Memory Wall
Integrated storage and computing can reduce data transfer overhead, thereby lowering inference costs and energy consumption. If the technology matures and is mass-produced, it will significantly enhance the efficiency of large model inference and may lead to new application scenarios.
Embodied Intelligence and Edge-side Multimodality are the Main Application Areas for Chips
Embodied intelligence (robots) and edge-side multimodality (phones/cars/glasses) require chips with low latency, low power consumption, and real-time perception control. These scenarios are driving the expansion of chips from cloud to edge and edge-side.