In the midst of a maelstrom about the value and risks that AI poses to society, nvidia has announced the launch of a new supercomputer that allows it to consolidate itself as a benchmark technology company in the field of high performance computing (HPC). All this is possible thanks to DGX GH200with which it is possible to develop giant models capable of powering generative AI.
Along with this, DGX GH200 will be focused on the recommender systemsand the data processing and analysisbeing available from the end of 2023. These are the digital engines of the modern economy and that is why it is so important to devise a technology according to it.
Its technical features
The DGX GH200, unlike the DGX A100 of 2020, employs Grace Hopper superchips from Nvidia, all connected to each other. Specifically, it is made up of 256 GH200 chips shaping a gigantic machine with high power and performance. It also combines interconnection technologies NV Link and NVLink Switch System to function as a single GPU.
With more than one exaflop of performance (one trillion floating point operations per second) and 144 terabytes of memory, it is estimated that the DGX GH200 can outperform the most powerful PC games. The memory that characterizes it is 500 times greater than the previous generation of DGX A100.
Through a statement, Nvidia has detailed that DGX GH200 uses GH200 super chips, which together with NVLink technology removes any need for a traditional CPU to GPU connection. And it is that with the chip Nvidia NVLink-C2Ceach superchip is capable of combining an Nvidia Grace CPU, based on the 32-bit ARM architecturewith a GPU Nvidia H100 Tensor Core.
DGX GH200 is a pioneer in the combination of superchips Grace Hopper (GH) and NVLink Switch Systemthus achieving that all the GPUs of the superchips work in unison as a single GPU and with a NVLink bandwidth 48 times greater than the previous generation.
Thanks to NVLink technology, it is possible to multiply up to seven times the bandwidth between the GPU and the CPU and reduce up to five times the consumption of interconnection energy. Thus, a 600 GB Hopper architecture GPU building block is provided.
Jen Hsun Huanghead of the American company has presented all the news related to AI at the international technology event ‘Computex’. He has been highly satisfied with the DGX GH200, as it allows you to get the power of a massive AI supercomputer, but with the simplicity of programming a single GPU.
the first customers
Nvidia has detailed that Microsoft, Goal and Google clouds They will be the first technology companies to have the DGX GH200 supercomputer to explore all its capabilities and apply it to their workloads related to generative AI. In this sense, the Vice President of Computing at Google Cloud, Mark Lohmeyerhas qualified that the new scale NVLink and the shared memory of the Grace Hopper superchips address the big problems present in large-scale AI.
At the same time, Girish Bablanicorporate vice president of Azure Infrastructure at Microsoft, said that the potential demonstrated by the DGX GH200, in being able to work with terabyte-sized data sets, will enable developers to undertake advanced research at greater scale and speed.
Nvidia’s forecasts
The estimated success rate is such that the American company plans to provide the DGX GH200 design to cloud service providers and other hyperscalers, on the basis that they could adapt and customize it to their infrastructure. Its release is expected by the end of 2023although new innovations will be incorporated to improve its efficiency and performance.
The latest from Nvidia
Similarly, Nvidia has announced that it has been working on its own AI supercomputer based on the technology provided by the DGX GH200. named Nvidia Heliosaims to enhance the work of researchers and increase the yield of data, since has four DGX GH200 systems.
Nvidia Helios will work by connecting each DGX GH200 system to the network Nvidia Quantum-2 InfiniBandbeing endowed with 1024 super chip Grace Hopper. As with the DGX GH200, Nvidia Helios is expected to be available in the market later this year.
