•A Programming Architecture Built on C
CUDA (Compute Unified Device Architecture) is an architecture and technology provided by NVIDIA for the GPU, offered as an extension used on top of the C programming language. It is not a standalone language but rather a C compiler based on PathScale, accompanied by a suite of development tools that enable algorithms written in C to run on the GPU. The technology was developed by NVIDIA and requires both an NVIDIA GPU and the latest version of its drivers to function.
Supported Hardware and Forward Compatibility
CUDA runs on every GPU within the G8X family, including GeForce, Quadro, and Tesla lines. NVIDIA states that due to forward code compatibility in its graphics card architectures, programs developed for the GeForce 8 series can automatically benefit from performance improvements on next-generation cards without any modifications. This forward compatibility is a defining characteristic of the platform’s hardware strategy.
Dezvoltator Control Over GPU Resources
The CUDA library enables developers to control the memory and Stream Processors located on CUDA-capable GPUs. The first CUDA Dezvoltator Kit (SDK) was released on February 15, 2007, marking the initial public availability of the development tools.
Benefits of the Architecture
CUDA enables the use of the graphics processor’s cores in general-purpose GPU computations. It uses the C language with only a few simple extensions. Additionally, code can write to randomly accessible memory, providing flexibility for certain computational patterns.
Technical Limitations
The source outlines several constraints on the architecture. Texture creation is not supported. Recursive (repeated) functions are not supported and must be converted into loops. The IEEE 754 standard is not supported for double precision floating point numbers. For best performance, execution units should run in groups of 32, and different branchings of the program flow within the group of 32 negatively affect performance.
Supported Cards
CUDA is supported only on NVIDIA Quadro and GeForce 8, 9, and 200 series cards, as well as all newer cards. A list of supported GPUs and graphics cards is maintained, with the official list of CUDA-supported devices referenced separately. The source notes that information was archived on the Wayback Machine on January 22, 2016.
Related Topics
The source connects CUDA to related concepts including GPGPU, OpenCL, BrookGPU, and Shader, as well as a comparison of deep learning software. External references include NVIDIA’s CUDA homepage (archived August 6, 2008), developer forums (archived August 3, 2008), and a Beyond3D article titled “Introducing CUDA – NVIDIA’s Vision for GPU Computing” dated February 16, 2007. A University of Illinois CUDA course taught by Wen-mei Hwu and David Kirk is also referenced, dated March 9, 2007.
This explanation is adapted from Wikipedia and is available under CC BY-SA 4.0.
Alătură-te conversației