PIM Task Batching Across CPU and GPU for Memory-Bound Workloads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computing systems face challenges in efficiently processing applications or workloads in environments where Processing-in-Memory (PIM) is mixed with traditional computing devices like GPUs and CPUs, lacking suitable solutions for parallel computation and memory-bound issues.
Innovation Solution
A method and device that enable parallel computation between PIM, CPU, and GPU by parsing user requests into tasks, grouping them into batched units, and scheduling execution based on data movement, utilizing a heterogeneous computing framework for synchronization and asynchronization of computation order, and optimizing data movement and replication between memories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If PIM is mixed with traditional computing devices (CPU/GPU), then processing performance for memory-intensive computations is improved, but system complexity and difficulty of coordinating computation between heterogeneous devices increases
Solution Approach 1:
The patent introduces a computing device that acts as an intermediary between PIM and traditional computing devices. This mediator coordinates task distribution, data movement, and computation scheduling, simplifying the complexity of heterogeneous system integration while enabling parallel processing and improving overall productivity.
2Productivity
If computation is performed sequentially on single devices, then system complexity is reduced, but processing time and productivity decrease
Solution Approach 1:
The patent segments computation into multiple independent tasks that can be distributed across PIM and traditional computing devices. By dividing the workload into parallelizable task units and scheduling their execution across heterogeneous devices, the system achieves faster processing while managing coordination complexity through structured task management.
Solution Approach 2:
The patent transitions from sequential single-device computation to parallel multi-device computation by adding the dimension of spatial distribution across heterogeneous devices. This enables simultaneous execution of multiple tasks on different devices, dramatically improving processing speed while the scheduling mechanism manages the increased coordination requirements.
3Adaptability or versatility
If data is frequently moved between memory and processing units, then computation flexibility is improved, but memory-bound problems and processing efficiency worsen
Solution Approach 1:
The patent performs preliminary actions by pre-processing and preparing data in advance within the computing device before transferring to PIM. This includes task parsing, data formatting, and computation setup, which reduces the frequency and volume of data movements during actual computation, thereby improving processing efficiency while maintaining computation flexibility.
Data Source
AI summary
The present disclosure relates to a method, a device, and a computer program for performing computation using PIM. The present disclosure presents a method for performing computation using PIM, the method including: parsing a user request into multiple tasks; grouping the multiple parsed tasks into at least one batched task unit; scheduling execution of the batched task unit, based on data movement between the multiple tasks; and executing the multiple tasks according to the scheduling.