Comment
Author: Admin | 2025-04-28
Files. cuda-libraries-dev This meta package contains library binary files and header files. 2.3.1. cuBLAS Library cuBLAS 9.0.333 is an update to CUDA Toolkit 9 that improves GEMM computation performance on Tesla V100 systems and includes bug fixes aimed at deep learning and scientific computing applications. The update includes optimized performance of the cublasGemmEx() API for GEMM input sizes used in deep learning applications, such as convolutional sequence to sequence (seq2seq) models, when the CUBLAS_GEMM_DEFAULT_TENSOR_OP and CUBLAS_GEMM_DEFAULT algorithm types are used. cuBLAS 9.0.282 is an update to CUDA Toolkit 9 that includes GEMM performance enhancements on Tesla V100 and several bug fixes targeted for both deep learning and scientific computing applications. Key highlights of the update include: Overall performance enhancements across key input sizes that are used in recurrent neural networks (RNNs) and speech models Optimized performance for small-tiled GEMMs with support for new HMMA and FFMA GEMM kernels Improved heuristics to speed up GEMMs across various input sizes The Volta architecture is supported, including optimized single-precision and mixed-precision Generalized Matrix-Matrix Multiply (GEMM) matrices, namely: SGEMM and SGEMMEx with FP16 input and FP32 computation for Tesla V100 Tensor Cores. Performance enhancements have been made to GEMM matrices that are primarily used in deep learning applications based on Recurrent Neural Networks (RNNs) and Fully Connected Networks (FCNs). GEMM heuristics are improved to choose the most optimized GEMM kernel for the input matrices. Heuristics for batched GEMM matrices are also fixed. OpenAI GEMM kernels and optimizations in GEMM for small matrices and batch sizes have been integrated. These improvements are transparent with no API changes. Limitations on New Features of the cuBLAS Library in CUDA 9 Batching GEMM matrices for Tesla V100 Tensor Cores is not supported. You may not be able to use cublasgemmBatched() APIs on Tesla V100 GPUs with Tensor Cores, but these functions will use the legacy FMA and HFMA instructions. Some GEMM heuristic optimizations and OpenAI GEMM kernels for small matrices are not available on Tesla V100 Tensor Cores. For cublasSetMathMode(), when set to CUBLAS_TENSOR_OP_MATH, cublasSgemm(), cublasGemmEx(), and cublasSgemmEx() will allow the Tensor Cores to be used when A/B types are set to CUDA_R_32F. In the CUDA 9 RC build, the behavior was to perform a down conversion from FP32 to FP16 with Round to Zero. In the production release of CUDA 9, the behavior is to perform a down conversion with Round to Nearest instead. To use single-precision
Add Comment