GPU-SSD Direct Data Path to Bypass CPU Transfer Bottlenecks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In conventional systems processing large data, such as machine learning and big data, graphics processing units (GPUs) face performance degradation and excessive resource consumption due to the need to request and transfer data through central processing units (CPUs), leading to long waiting times and data duplication across storage devices and memory.
Innovation Solution
A system where the application processor generates read/write requests and includes a data memory, allowing direct data transmission from a data storage device to the application processor's data memory without passing through the host processor, utilizing a bus circuit for P2P data exchange and shared memory to manage offset information for efficient data operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If data is transferred through the host processor (CPU) from the data storage device to the application processor (GPU), then data can be read and written between devices, but the waiting time at the GPU increases significantly and system performance deteriorates
Solution Approach 1:
The patent introduces a direct data transmission path between the data storage device and the application processor that bypasses the host processor as an intermediary. The data storage device can directly transmit data to the application processor's data memory without requiring the host processor to read and rewrite the data, thereby eliminating the time loss and performance degradation caused by the traditional CPU-mediated data transfer path.
2Ease of operation
If data is transferred through the host processor and main memory, then data exchange between storage device and application processor is enabled, but system resources are excessively consumed for data duplication
Solution Approach 1:
The patent extracts the data transmission function from the host processor and main memory path, creating a dedicated direct transmission channel between the data storage device and the application processor. This extraction eliminates the unnecessary data duplication in the host processor and main memory, thereby reducing system resource consumption and energy usage while maintaining full data exchange capability.
3Reliability
If the GPU requests data through the CPU, then data can be accessed from the SSD, but the operation sequence (GPU→CPU→SSD→CPU→GPU) creates excessive complexity and delays
Solution Approach 1:
The patent inverts the traditional data access architecture by enabling the data storage device to directly transmit data to the application processor without requiring the host processor to mediate the data flow. This inversion simplifies the operation sequence from a multi-hop path (GPU→CPU→SSD→CPU→GPU) to a direct path (SSD→GPU), reducing both complexity and potential failure points while maintaining data access reliability.
Data Source
AI summary
A system includes an application processor configured to generate a read request and including a data memory; a host processor configured to generate a read command corresponding to the read request; and a data storage device including a data storage memory, wherein the data storage device transmits read data output from the data storage device according to the read command to the data memory of the application processor without passing the host processor.