The Israel-1 supercomputer is based on the new NVIDIA Spectrum-X platform developed in Israel, and will be used for internal research and development and with the company's partners. Later, it will be launched in Israel as part of the supercomputer service for AI in the cloud, NVIDIA DGX Cloud * At the same event, it also announced a dedicated supercomputer for artificial intelligence applications
Nvidia announced a new type of memory-intensive supercomputers for large and complex artificial intelligence tasks - a series supercomputer NVIDIA DGX based on sample superchips NVIDIA Grace Hopper Superchips and communication technology NVIDIA NVLink Switch System, created to enable the development of next-generation generative artificial intelligence models especially for language applications, recommender systems, and massive data analysis and processing tasks.
The NVIDIA DGX GH200, which is expected to be available towards the end of the year, has a huge amount of memory. It uses NVLink technology with the NVLink switch system to combine 256 Grace Hopper Superchips, which work together as one giant GPU. This unique combination provides 144 terabytes of memory – 500 times more than the NVIDIA DGX A100 system.
"Creative artificial intelligence, large language models (LLMs), and recommendation systems, are the digital engines of the modern economy", said Jensen Huang, founder and CEO NVIDIA. "The DGX GH200 artificial intelligence supercomputers combine NVIDIA's most advanced accelerated computing and communication technologies to expand the boundaries of AI."
The GH200 super-chips eliminate the need for traditional PCIe connections between central processing units (CPU) and graphics processing units (GPU) in computing systems, by combining NVIDIA Grace central processor And NVIDIA H100 Tensor Core graphics processor in the same packaging, using a connection NVIDIA NVLink-C2C. The advanced connection increases the bandwidth between GPU and CPU by seven times compared to the latest PCIe technology, cuts the power consumption of the connections by more than five times, thus providing the building blocks of the DGX GH200 supercomputers, with a GPU of 600 giga- Byte in the innovative Hopper architecture.
The DGX GH200 is the first supercomputer to pair the Grace Hopper superchips with the NVIDIA NVLink Switch System, a new connectivity technology that allows all the GPUs in the DGX GH200 to work as one giant GPU. The previous generation of the system, by comparison, only allowed eight graphics processors to work as one without compromising performance.
Google, Meta and Microsoft are among the first companies to have access to the DGX GH200 for research and development of very large artificial intelligence models.
NVIDIA Helios
To power the activities of its research and development teams, NVIDIA is building the NVIDIA Helios supercomputer that will be based on the DGX GH200. The NVIDIA Helios supercomputer will consist of 4 DGX GH200 systems, connected to each other via NVIDIA Quantum-2 InfiniBand, the communication technology that was developed in Israel and offers a bandwidth of up to 400 gigabits per second. Helios will feature a total of 1,024 Grace Hopper superchips and is expected to be operational by the end of the year.
The supercomputer ISRAEL1
Nvidia announced today the establishment of Israel-1, the most powerful supercomputer in Israel for AI calculations, which will be used by the company for cooperation with industry, for internal R&D, and also as a template for building supercomputers based on the Spectrum-X platform. Down the road, it will be available as part of the NVIDIA DGX Cloud supercomputing service, delivering the power of cloud AI infrastructure to the on-premises ecosystem.
When completed, Israel-1 will incorporate 256 HGX H100 systems from Dell Technologies, including 2,048 NVIDIA H100 80GB GPUs with more than 34 million CUDA cores and over one million 2,560th generation Tensor cores. The supercomputer will also incorporate 3 BlueField-80 DPU processors and 4 Spectrum-XNUMX switches - all developed in Israel.
Israel-1, unveiled at the Computex conference, is expected to deliver AI performance of up to 8 Exaflops, making it one of the world's fastest supercomputers for artificial intelligence calculations. In addition, it is expected to provide performance of more than 135 petaflops (Petaflops) in classic supercomputing tasks.
The supercomputer, whose value is estimated at hundreds of millions of dollars, is expected to enter initial use by the end of the year.
"Artificial intelligence is the most important technological force in our lives, and NVIDIA is committed to maintaining its leadership at the forefront of AI," said Gilad Sheiner, Senior Vice President of Supercomputing and Communications at Nvidia. Israel-1 represents a huge investment for Nvidia, which will allow us to drive innovation in Israel and around the world," he added.
The new supercomputer will be based on the NVIDIA Spectrum-X platform that was just announced at Computex, and represents a new type of Ethernet communication technology developed to improve and accelerate the performance of Ethernet-based cloud computing for artificial intelligence (AI Clouds).
The platform is powered by the tight integration of NVIDIA BlueField-3 DPU chips and Spectrum-4 switches, which together result in up to 1.7 times improvement in artificial intelligence performance and energy efficiency. Spectrum-X also includes the set of software solutions from Nvidia that will allow developers to build new AI applications adapted to work in the cloud.
Providing end-to-end artificial intelligence capabilities reduces the runtimes of Transformer-based Generative AI models, enabling engineers, data scientists and cloud providers to reach accurate results and make informed decisions quickly.
