From NVIDIA: 2024-08-26 12:00:49

NVIDIA’s new libraries in accelerated computing deliver order-of-magnitude speedups and reduce energy consumption and costs in various applications such as data processing, generative AI, recommender systems, and 6G research. The libraries include tools like NeMo Curator for creating custom datasets, cuVS for vector search, and Polars GPU Engine in open beta. Other tools support physics simulation, wireless network simulation, and link-level wireless simulation. Companies worldwide are increasingly turning to NVIDIA accelerated computing to speed up applications previously only run on CPUs, achieving extreme speedups and energy savings. For example, CPFD in Houston uses NVIDIA GPUs for its Barracuda Virtual Reactor software for designing recycling facilities. Plastic recycling plants have seen simulations run 400x faster and 140x more energy efficiently using NVIDIA GPUs. In a shift from CPUs to GPUs in the cloud, a popular video conferencing app increased throughput from 3 to 200 queries per second for live captions, achieving a 66x speedup and 25x energy-efficiency improvement. Similarly, an e-commerce website connected millions of shoppers with deep learning-powered recommendations on NVIDIA accelerated cloud computing, reducing latency by 33x and boosting energy efficiency by nearly 12x. NVIDIA estimates that converting all AI, HPC, and data analytics workloads still on CPU servers to CUDA GPU-accelerated systems could save 40 terawatt-hours annually, equivalent to the energy consumption of 5 million U.S. homes per year. NVIDIA accelerated computing leverages GPUs’ parallel processing capabilities to complete tasks much faster than CPUs, enhancing productivity and reducing energy consumption. GPUs offer 20x greater energy efficiency than CPUs for superior performance. NVIDIA AI computing has achieved 100,000x more energy efficiency in processing large language models in the last decade. In comparison, if cars improved efficiency by the same rate, they’d get 500,000 miles per gallon. NVIDIA accelerated computing enables impressive speedups over CPUs, with customers experiencing 10-180x faster performance across various real-world tasks on the cloud. GPUs’ performance advantage over CPUs has led to a growing performance gap with increasing data growth, driving demand for accelerated computing. The gap between data growth and CPU compute performance per watt has widened, emphasizing the energy-saving benefits of GPU acceleration. Specialized algorithm software libraries are essential to accelerate specific workloads on GPUs, offering optimized tools for parsing and executing data calculations. NVIDIA provides a range of CUDA libraries for diverse use cases, including LLM applications, data processing tasks, and physical AI simulations. New updates continue to expand the CUDA platform’s capabilities, supporting developers in various fields like computer vision, data science, and wireless network design. NVIDIA’s diverse library offerings cater to both general and specialized workloads, harnessing GPUs’ hardware features for optimized performance. Additional libraries focus on accelerating specific tasks like computer vision, data frames, and silicon computational lithography. For researchers seeking a streamlined deployment process, NVIDIA NIM packages multiple libraries and AI models into optimized containers for production use. The containerized microservices enhance workflow efficiency for researchers across industries like CAD, CAE, and EDA.



Read more at NVIDIA: CUDA Libraries Expand Accelerated Computing Into New Science, Industrial Applications