// Allocate and fill matrices...
Enter – a modern solution designed to handle the messy, heterogeneous reality of advanced computing. The Problem with Traditional Batched GEMM Imagine training a recommendation system with embedding tables of varying sizes, or running inference on a transformer model with variable sequence lengths. In these scenarios, you might have 1,024 independent GEMM operations, each with different M, N, or K dimensions. cublaslt grouped gemm
cublasLtMatmulDesc_t matmulDesc; cublasLtMatmulDescCreate(&matmulDesc, CUDA_R_32F, CUDA_R_16F); // Allocate and fill matrices
If you're building a transformer-based model, a recommender system, or any application that requires many small, independent matrix multiplications, Grouped GEMM should be your default choice. As NVIDIA continues to optimize cuBLASLt for Hopper and future architectures, the performance gap between irregular and regular workloads will only shrink further. For implementation details, refer to the NVIDIA cuBLASLt Developer Guide (CUDA 12.x and later). In these scenarios, you might have 1,024 independent
Traditional cuBLAS offers batched GEMM (e.g., cublas<t>gemmBatched ), which runs a list of independent matrix multiplications. However, it comes with a major limitation: (M, N, K) and data types.
In the world of High-Performance Computing (HPC) and Deep Learning (DL), the General Matrix Multiply (GEMM) operation is the undisputed king. From large language models (LLMs) to scientific simulations, performance often hinges on how efficiently you can compute C = α*A*B + β*C .