Nvidia has been focusing entirely on productive artificial intelligence lately. In addition to its tools that use productive AI, the firm also offers powerful systems to train, research and process AI. Within the scope of Computex 2023, Nvidia came up with the new DGX GH200 artificial intelligence supercomputer.Nvidia DGX GH200 introduced
This new DGX supercomputer features the company’s GH200 Grace Hopper Superchip and Nvidia NVLink Switch System solutions. Nvidia combines the DGX GH200’s large shared memory space, NVLink Switch System and 256 GH200s using NVLink interconnect technology, enabling them to perform as a single GPU. This results in 1 exaflop performance and 144 terabytes of shared memory. It means about 500 times more memory than the previous generation NVIDIA DGX A100 introduced in 2020.
“Generative AI, major language models, and recommendation systems are the digital engines of the modern economy,” CEO Jensen Huang said. According to Huang’s statements, the DGX GH200 artificial intelligence supercomputers use the company’s most advanced accelerated computing and networking technologies to expand the frontiers of artificial intelligence.DGX GH200 artificial intelligence supercomputer in details
The GH200 Grace Hopper Superchips use the company’s NVLink-C2C chip interconnects to combine an Arm-based Grace CPU and an Nvidia H100 Tensor Core GPU in the same package, eliminating the need for traditional CPU-GPU PCIe connectivity. Thus, the bandwidth between GPU and CPU is increased by 7 times compared to the latest PCIe 5.0 technology, interconnect power consumption is reduced by more than 5 times, and 600GB Hopper architecture GPU building block is provided for DGX GH200 supercomputers.
The DGX GH200 is the first supercomputer to pair Grace Hopper Superchips with the NVLink Switch System, a new interconnect that enables all GPUs in the DGX GH200 system to operate as a single GPU. The previous generation system, on the other hand, only allowed eight GPUs to be combined into a single GPU with NVLink without sacrificing performance. The DGX GH200 architecture delivers 48 times more NVLink bandwidth than the previous generation, promising the power of a massive AI supercomputer with the simplicity of programming a single GPU.Giant companies are next
Google Cloud, Meta and Microsoft are among the first to access the DGX GH200 to explore its capabilities for productive AI workloads. Nvidia also plans to make this system available as a blueprint to cloud service providers and other hyperscalers so they can further customize it for their infrastructure.On the other hand, Nvidia is building its own DGX GH200-based artificial intelligence supercomputer to power the work of its researchers and development teams. Each will be interconnected via the Nvidia Quantum-2 InfiniBand network, increasing the throughput needed to train large AI models. This supercomputer, named Helios, will contain 1,024 Grace Hopper Superchips and will be operational by the end of the year.
Returning to the DGX GH200 supercomputers, this product is actually a complete package. Nvidia ships the product with its own software to provide a turnkey, full-stack solution for AI and data analytics workloads. Finally, the DGX GH200 supercomputers are expected to be available by the end of the year.
7237
