Shulou(Shulou.com)11/24 Report--

With the development of artificial intelligence technology, machine learning applications are more and more extensive, from intelligent voice assistant to autopilot, from intelligent recommendation to image recognition, all need a lot of computing resources to support. As an efficient computing resource, GPU has attracted more and more attention and become an important tool for machine learning to accelerate computing. However, cross-hardware general acceleration is lack of cross-platform and cross-hardware general API, so it is very difficult and complex for different graphics cards to achieve efficient operators.

As a head technology company, Tencent has been committed to promoting the development of artificial intelligence technology. Therefore, as a member of khronos, Tencent actively participates in the formulation of new extended standards to provide a better solution for machine learning accelerated computing. In Vulkan 1.3.255, Tencent joined ARM, NVIDIA, AMD, Google and many other technology companies around the world to bring a new extension VK_KHR_cooperative_matrix. This is the first time that Tencent has participated in the contribution of khronos standards.

Vulkan is a general-purpose, cross-platform, new generation of graphics acceleration API, supporting Windows, Linux, macOS, Android, iOS and other operating systems. VK_KHR_cooperative_matrix extension brings mesoscale matrix type to Vulkan, which is used to accelerate matrix calculation and neural network reasoning. This new expansion makes it possible to accelerate general AI computing without a single hardware binding, breaking the industry monopoly and bringing about industrial innovation. Nvidia, ARM, AMD and other graphics card manufacturers will release new drivers to support this extended standard, which will further promote the development of machine learning accelerated computing.

Tencent YouTu Lab participated in the formulation of the VK_KHR_cooperative_matrix extension standard. In the process of revising the standard, Tencent YouTu Lab, combined with the practical experience of vulkan acceleration in the ncnn project, advocated that the stride parameter of the matrix loading function in the new standard is allowed to be 0 in order to support automatic broadcast behavior. This parameter can effectively improve the efficiency of bias data processing in neural network convolution and linear layer computation. Nihui, an expert from Tencent YouTu Lab and author of ncnn, a well-known open source neural network reasoning library in the industry, said: "in the khronos online meeting, the proposal was recognized and approved by the technical staff of the participating manufacturers and became one of the hard standards, requiring that the behavior be implemented in the GPU driver."

Ncnn uses Vulkan API as its general acceleration scheme for cross-platform GPU. By using the VK_KHR_cooperative_matrix extension, ncnn runs over-score AI on the AMD graphics card, with a speed increase of about 2.3 times. This is also the first application of VK_KHR_cooperative_matrix, which has been released in the new version of ncnn, bringing a wider range of GPU acceleration across hardware vendors.

Tencent participated in the formulation of the VK_KHR_cooperative_matrix extension standard, promoted the development of artificial intelligence technology, made the computing power of machine learning more efficient, inclusive and sustainable, provided a better solution for machine learning accelerated computing, realized general API across platforms and hardware, and made machine learning accelerated computing more efficient and flexible.

For a long time, Tencent has been actively participating in the standard construction of all kinds of AI and big models. In 2020, Tencent was elected as a member and deputy secretary-general of the artificial Intelligence Sub-Committee of the National Beacon Commission, which means that Tencent is serving as a core member in the process of promoting "new infrastructure", including artificial intelligence. to take on more standard-setting work and technology leadership. Not long ago, Tencent Cloud, together with China Institute of Information and Communication, launched an industry large model ecological plan and took the lead in the formulation of China's first large model standard for the financial industry, providing important support for the intelligent and high-quality standardized development of the financial industry.

In the future, Tencent will actively participate in the formulation of more industry standards, help more industries improve quality and efficiency, and bring more well-being to human society.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.