GEMM Operations

Benchmarking double-precision general matrix multiplication (DGEMM) is crucial as it is a fundamental building block for many HPC applications. This test measures sustained floating-point performance for large square matrices (N x N).

To be updated

Batched GEMM operations involve performing many independent matrix multiplications on small matrices. This is common in applications like deep learning (tensor contractions), scientific simulations with many small systems (e.g., astrophysics, materials science), and block-oriented algorithms. This benchmark evaluates the throughput (e.g., GFLOPS or matrices per second) for various batch sizes and matrix dimensions.

To be updated

For matrices too large to fit in the memory of a single compute node, or to accelerate computations, GEMM operations must be distributed across multiple nodes. This benchmark evaluates the strong and weak scaling performance of distributed GEMM implementations (e.g., ScaLAPACK's PDGEMM, SLATE, DLA-Future) using metrics like sustained GFLOPS and parallel efficiency.

To be updated