NVIDIA also unveiled its first CPU, Grace, which is aimed at the new data centre. Grace Hopper, a computer scientist and one of the first programmers of the Harvard Mark 1 and inventor of the first linkers, was one of the founders of computer science and one of the first programmers of the Harvard Mark 1.
The Grace CPU, NVIDIA's first ARM-based processor, is aimed at data centres and features Neoverse Cores.
The Grace CPU is NVIDIA's first design to use the Neoverse cores, which have been in production for years. The SOC is made up of several chips, each of which contains three main components: the CPU, GPU, and memory/IO subsystems. NVIDIA hasn't released much data, but it has stated that the Grace CPU will be able to achieve a SPECrate2017 Int base score of 300.
Grace, according to NVIDIA, is a highly advanced processor designed for workloads including training next-generation NLP models with over 1 trillion parameters. A Grace CPU-based system can provide 10x faster performance than today's state-of-the-art NVIDIA DGX-based systems, which operate on x86 CPUs when closely coupled with NVIDIA GPUs.
Grace is being introduced by NVIDIA as the amount of data and the complexity of AI models grows exponentially. The most advanced AI models today have billions of parameters and double every two and a half months. To remove device bottlenecks, they'll need a new CPU that can be closely combined with a GPU.
Grace was designed by NVIDIA by using Arm's data centre architecture's unprecedented versatility. NVIDIA is advancing the mission of technological diversity in the AI and HPC communities, where choice is critical to providing the creativity required to address the world's most pressing challenges, by launching a new server-class CPU.
Grace's Early Adopters Test the Boundaries between Science and AI
Grace-powered supercomputers, designed by Hewlett Packard Enterprise, will be operational in 2023 at both CSCS and Los Alamos National Laboratory.
“This next-generation architecture will shape our institution's computing strategy with a creative combination of memory bandwidth and capacity,” said Thom Mason, director of the Los Alamos National Laboratory. “We will be able to provide sophisticated science analysis using high-fidelity 3D models and analytics with databases that are bigger than historically possible thanks to NVIDIA's recent Grace CPU.”
Delivering up to 10 times faster performance than x86 servers
Grace's success is powered by NVIDIA's fourth-generation NVLink interconnect technology, which offers a record 900 GB/s connectivity between Grace and NVIDIA GPUs, allowing for 30x more combined bandwidth than today's leading servers.
Grace will also use a cutting-edge LPDDR5x memory subsystem, which will have twice the bandwidth and ten times the energy consumption of DDR4 memory. Furthermore, the latest architecture offers unified cache coherence with a single memory address space, simplifying programmability by integrating framework and HBM GPU memory.
Grace will be supported by the NVIDIA HPC software development kit, as well as the entire suite of CUDA and CUDA-X libraries, which accelerate over 2,000 GPU programmes, allowing scientists and engineers working on the world's most pressing problems to make faster discoveries. Grace CPUs are projected to be available in the first half of 2023.
Post a Comment