CUDA 13.1
تُحدث ثورة في حوسبة GPU من خلال نماذج برمجة جديدة وأداء مُحسّن عبر أحدث المعماريات.
NVIDIA CUDA 13.1: تحديث شامل لبيئة CUDA، يقدم CUDA Tile لتبسيط البرمجة، وتعزيز الأداء، وتحسين أدوات المطورين، مع دعم مكتبات الرياضيات المُحسّنة.
المزيد عن CUDA 13.1
CUDA 13.1
NVIDIA CUDA 13.1 introduces a new tile-based programming model for GPU kernels. It offers performance improvements and new features for accelerated computing. With this release, developers can write code that is compatible with current and future GPU architectures.
- •CUDA Tile Programming:Enables writing GPU kernels at a higher level of abstraction using tiles instead of individual threads. Simplifies the use of specialized hardware like tensor cores, ensuring code compatibility with future GPU architectures.
- •Green Contexts in Runtime API:Provides a lightweight alternative to traditional CUDA contexts for finer-grained spatial partitioning and resource management on the GPU. Allows developers to define and manage distinct partitions of GPU resources, improving performance for latency-sensitive code.
- •Memory Locality Optimization:Introduces MLOPart on select GPUs, creating specialized CUDA devices to improve memory locality. Each partition appears as a distinct CUDA device with its own compute and memory resources.
- •Static SM Partitioning:Provides a way to create exclusive SM partitions for MPS clients on NVIDIA Ampere and newer GPUs. Delivers deterministic resource allocation and improved isolation between MPS clients.
- •cuBLAS FP Emulation:Improves the performance of double-precision (FP64) matrix multiplications using floating-point emulation on Tensor Cores. Offers significant speedups on Blackwell GPUs.
- •CUDA Tile Kernel Profiling:Adds support for profiling CUDA Tile kernels in NVIDIA Nsight Compute. Includes a new “Tile Statistics” section to summarize Tile dimensions and utilization.
- •Compile-time Patching:Integrates error detection directly into NVCC for faster runs while catching more subtle memory issues. Debug memory problems without sacrificing speed, run more tests, and maintain productivity.
- •Deterministic Reductions:Provides additional floating-point determinism options for CUB algorithms. Offers GPU-to-GPU determinism, ensuring bitwise-identical results.
- •Single-phase CUB APIs:Adds new overloads to CUB algorithms that accept a memory resource, eliminating the need for a separate temporary storage query/allocate/free pattern.
By choosing CUDA 13.1, you gain access to advanced features that streamline GPU programming and optimize performance. The introduction of CUDA Tile simplifies kernel development, while enhancements to libraries like cuBLAS and cuSOLVER deliver significant speedups. This release empowers developers to achieve greater efficiency and unlock the full potential of their hardware.
