CUDA(Compute Unified Device Architecture),是显卡厂商 NVIDIA 推出的运算平台。作为一种通用并行计算架构,CUDA 使 GPU 能够解决复杂的计算问题。 它包含了 CUDA 指令集架构(ISA)以及 GPU 内部的并行计算引擎。

CUDA (Compute Unified Device Architecture) is a computing platform launched by the graphics card manufacturer NVIDIA. As a general-purpose parallel computing architecture, CUDA enables GPUs to solve complex computing problems. It contains the CUDA instruction set architecture (ISA) and the parallel computing engine inside the GPU.

Currently, NVIDIA CUDA 11.7 is released, CUDA 11.7 brings compatibility support for the new NVIDIA Open GPU Kernel Modules, and another important highlight is lazy loading support. And the CUDA 11.7 Toolkit is available as the latest feature update to NVIDIA's proprietary compute stack.

The main features of CUDA 11.7 are updated as follows:

Support for using NVIDIA Open GPU kernel modules with CUDA 11.7 and R515 beta drivers.
Lazy loading of compute kernels saves device-side memory by deferring loading at the beginning of the program until the first call to the compute kernel. The "CUDA_MODULE_LOADING=LAZY" environment variable controls the lazy loading behavior of this CUDA module.
The NVCC host compiler supports Clang 13.
Various new features for Nsight Compute, CUPTI, NVPROF / Visual Profiler and other CUDA development tools.
Vulkan-CUDA interop on L4T on Android and Tegra now supports all color formats.
Deprecated 32-bit compilation of NVCC for all GPUs.
CentOS Linux 8 has also been removed from the CUDA toolkit due to upstream CentOS 8 EOL and replaced by Rocky Linux 8 support.

For more details on CUDA 11, check out the release notes.

目前,NVIDIA CUDA 11.7 发布了,CUDA 11.7 为新的 NVIDIA Open GPU Kernel Modules 带来了兼容性支持,另一个重要的亮点是延迟加载支持。 而 CUDA 11.7 Toolkit 已作为 NVIDIA 专有计算堆栈的最新功能更新提供。

CUDA 11.7 主要功能更新如下:

  • 支持将 NVIDIA Open GPU 内核模块与 CUDA 11.7 和 R515 beta 驱动程序一起使用。
  • 延迟加载计算内核,将程序开始时的加载推迟到首次调用计算内核时,可节省设备端内存。“CUDA_MODULE_LOADING=LAZY” 环境变量控制这个 CUDA 模块延迟加载行为。
  • NVCC 主机编译器支持 Clang 13。
  • Nsight Compute、CUPTI、NVPROF / Visual Profiler 和其他 CUDA 开发工具的各种新功能。
  • 现在,Android 和 Tegra 的 L4T 上的 Vulkan-CUDA 互操作支持所有颜色格式。
  • 弃用 NVCC 对所有 GPU 的 32 位编译。
  • 由于上游 CentOS 8 EOL,CentOS Linux 8 也已从 CUDA 工具包中删除,并由 Rocky Linux 8 支持取代。

有关 CUDA 11 的更多详细信息,可查看发行说明 。

声明:本站所有文章,如无特殊说明或标注,均为本站原创发布。任何个人或组织,在未征得本站同意时,禁止复制、盗用、采集、发布本站内容到任何网站、书籍等各类媒体平台。如若本站内容侵犯了原著者的合法权益,可联系我们进行处理。