NVIDIA cuTENSOR is a GPU-accelerated tensor linear algebra library designed for high-performance tensor computations. It supports tensor contraction, reduction, and elementwise operations, enabling applications to harness the specialized tensor cores on NVIDIA GPUs. cuTENSOR accelerates deep learning training and inference, computer vision, quantum chemistry, and computational physics workloads.

Features

• Just-in-Time Compiled Kernels: Offers just-in-time compiled kernels for tensor contraction, significantly boosting performance.
• Plan-Based Multi-Stage APIs: Provides plan-based multi-stage APIs for all operations, enhancing flexibility and efficiency.
• Support for Arbitrarily Dimensional Tensor Descriptors: Handles tensors of any dimensionality, making it versatile for various applications.
• 3xTF32 Compute Type: Supports 3xTF32 compute type for improved performance in tensor computations.
• Int64 Extents: Supports int64 extents, allowing for large-scale tensor operations.
• Mixed Precision Support: Enables mixed precision computations, optimizing performance and resource usage.
• Expressive API: Allows for elementwise operation fusion, simplifying complex tensor operations.

Benefits

• High Performance: Delivers significant performance gains, leveraging NVIDIA GPUs' tensor cores.
• Flexibility: Supports a wide range of tensor operations and dimensionalities, making it suitable for diverse applications.
• Efficiency: Just-in-time compiled kernels and mixed precision support enhance computational efficiency.
• Scalability: Handles large-scale tensor operations, making it ideal for high-performance computing (HPC) environments.
• Versatility: Applicable to various fields, including AI, deep learning, computer vision, quantum chemistry, and computational physics.