Parallel Task-Based Tensor Network Contraction for Quantum Circuit Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current classical simulations of quantum circuits face inefficiencies, including overestimation and high computational costs, particularly in simulating random quantum circuits, which are challenging due to low CPU usage and redundancy in calculations, and are not well-suited for the heterogeneity and concurrency expected in exascale computing environments.
Innovation Solution
The implementation of task-based parallelism, using asynchronous task-parallelism and libraries like Taskflow, allows for efficient decomposition of tensor network simulations across heterogeneous nodes, enabling concurrent contraction on CPUs and GPUs, and optimizing shared work reuse to reduce redundant calculations and maximize parallelism.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If classical simulation methods are used for quantum circuits, then simulation capability is provided, but computational efficiency is low and redundancy in calculations occurs
Solution Approach 1:
The patent segments the quantum circuit simulation into multiple tensor networks representing different paths or components of the circuit. Each tensor network can be contracted independently or in parallel, dividing the monolithic computational problem into manageable segments that reduce overall computational overhead and eliminate redundancy.
Solution Approach 2:
The patent introduces a new dimension of parallelism by utilizing heterogeneous computing architectures (CPU, GPU, FPGA) simultaneously. This multi-dimensional approach allows different subsets of tensor network contractions to execute concurrently on different device types, achieving order-of-magnitude speedups by exploiting architectural diversity rather than just increasing single-device power.
2Adaptability or versatility
If conventional tensor network contraction is used, then quantum circuit simulation is achieved, but it does not utilize heterogeneous architectures effectively
Solution Approach 1:
The patent creates a universal simulation framework that can execute on multiple heterogeneous architectures (CPU, GPU, FPGA). The system divides tensor network contraction tasks and distributes them across different device types, allowing each architecture to contribute its strengths. This multi-functional approach enables the same simulation code to adaptively utilize available hardware resources, achieving both versatility and high efficiency.
Solution Approach 2:
The patent implements dynamic task distribution and load balancing across heterogeneous devices. The system can adaptively assign different subsets of contraction tasks to different device types based on their current state and capabilities, allowing the simulation to dynamically optimize performance across the heterogeneous architecture rather than using a static allocation scheme.
3Ease of operation
If sequential processing is used for tensor network contraction, then simplicity is maintained, but parallelism and concurrency are not exploited
Solution Approach 1:
The patent segments the tensor network contraction into independent sub-tasks that can be executed in parallel. By identifying independent contraction paths and subsets of operations, the system maintains the conceptual simplicity of tensor network methods while enabling concurrent execution on multiple processing units, thus improving throughput without significantly complicating the implementation.
Data Source
AI summary
A method includes receiving a representation of a quantum circuit at a processor and identifying multiple contraction trees based on the representation of the quantum circuit. Each of the contraction trees represents a tensor network from a set of tensor networks. A first subset of multiple tasks, from a set of tasks associated with the plurality of contraction trees, is assigned to a first set of at least one compute device having a first type. A second subset of multiple tasks mutually exclusive of the first subset of multiple tasks is assigned to a second set of at least one compute device having a second type different from the first type. The quantum circuit is simulated by executing the first subset of tasks via the first set of at least one compute device and executing the second subset of tasks via the second set of at least one compute device.