Integrated Circuit Memory Management for Neural Network Calculations

The patent segments the neural network calculation process into multiple operator engines (first operator engine, second operator engine, etc.) that can perform calculations in parallel. Each operator engine processes specific operations independently, allowing the system to achieve high calculation capability without requiring a single large memory block. The segmentation also applies to memory management, where different memory regions are allocated to different operator engines based on their specific needs.

The patent implements dynamic memory management through a memory management unit (MMU) that dynamically allocates and deallocates memory resources based on the actual calculation requirements of different operator engines. The MMU maintains mapping tables that are updated in real-time to reflect the current state of memory usage, allowing the system to adapt memory allocation to changing computational demands rather than requiring fixed large memory allocations.

The patent creates a universal memory pool that serves multiple operator engines simultaneously. Instead of dedicating separate memory resources to each operator engine, a single memory pool is shared among all operators through the MMU's address translation and allocation mechanisms. This multi-functional memory system allows the same physical memory resources to support calculations for multiple different neural network models and operations, reducing the total memory capacity required.

The patent changes the parameter of memory allocation from fixed to variable. The MMU dynamically adjusts the amount of memory allocated to each operator engine based on the current calculation task's requirements. When one operator engine finishes its computation, its memory resources are released and reallocated to other operators that need them. This parameter change allows the system to achieve high calculation capability with smaller total memory capacity by efficiently reusing memory resources.

The patent ensures continuous utilization of memory resources through the MMU's allocation mechanism. When one operator engine completes its calculation, its memory resources are immediately made available for allocation to the next operator engine that requires them. This continuous reuse of memory resources eliminates idle time and ensures that the limited memory capacity is constantly productive, allowing the compact device to maintain high calculation capability despite small memory size.

The patent performs preliminary actions by pre-configuring the MMU with mapping tables and allocation strategies before the actual neural network calculations begin. The system prepares the memory management infrastructure in advance, so that when calculations start, the operator engines can immediately access their required memory resources without delay. This preliminary setup allows the compact device to maximize its calculation capability from the outset despite limited memory resources.