Command-List Processing for External-to-Local Memory Synchronization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information processing devices fail to synchronize the data transfer progress with the timing of the calculation processing unit, leading to increased latency and reduced performance, particularly in AI accelerators with external and internal memory transfers.
Innovation Solution
Incorporating a transfer size monitoring unit into the DMAC to detect the completion of data transfer and notify a command list processing unit, allowing the calculation processing unit to initiate operations during data transfer, thereby synchronizing the data transfer and processing units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data is transferred from external memory to local memory without synchronization mechanism, then data transfer can proceed independently, but the calculation processing unit cannot start operations until data transfer completes, increasing processing time
Solution Approach 1:
The command list processing unit generates and stores calculation commands in advance before data transfer completes. The calculation processing unit can then start executing these pre-prepared commands as soon as data is ready, eliminating idle waiting time and improving overall processing efficiency
Solution Approach 2:
The transfer size monitoring unit provides feedback about data transfer progress to the command list processing unit. This feedback mechanism enables the system to synchronize data transfer with calculation operations, allowing the calculation processing unit to start operations at the optimal timing when data is available
2Reliability
If the calculation processing unit waits for data transfer completion before starting operations, then data availability is ensured, but processing time increases due to sequential execution
Solution Approach 1:
Commands are prepared and stored in advance in the command list. When data transfer completes, the calculation processing unit can immediately start executing pre-prepared commands without waiting for additional setup, ensuring both data availability and efficient processing timing
Solution Approach 2:
The command list processing unit acts as an intermediary between data transfer completion and calculation execution. It manages the transition by generating appropriate calculation commands based on transfer status, enabling seamless coordination between data availability and processing initiation
3Device complexity
If data transfer and calculation operations are executed sequentially, then coordination is simple, but overall processing time increases
Solution Approach 1:
The system maintains continuous useful action by overlapping data transfer with command preparation and calculation execution. The transfer size monitoring unit continuously tracks transfer progress and triggers appropriate commands, ensuring that data transfer, command generation, and calculation operations proceed in parallel without coordination complexity
Data Source
AI summary
An information processing device is provided that synchronizes the progress of data transferring unit with the timing of processing by calculation processing unit. The information processing device comprises an external bus connecting unit for connecting data transferring unit, a local memory, and a command list processing unit to the external memory, the data transferring unit for transferring data, which are conditions for calculation stored in the external memory, to the local memory, the local memory for storing data and a list of commands, the command list processing unit for generating commands that cause the calculation processing unit to execute calculations, by reading the list of commands from the local memory while data is being transferred from the external memory to the local memory, and the calculation processing unit for executing calculations and processing data.