Nvidia has announced the new DGX GH200 supercomputer, designed for the development of a “Giant Model” that is power-generating artificial intelligence (AI), and focused on recommendation systems and data analysis workloads, expected by 2023. Will be available eventually.
The technology company highlights that generative AI, pervasive language models and recommendation systems are “the digital engines of the modern economy.” In this sense, they have created a supercomputer capable of performing the tasks required to develop this type of technology.
Thus, Nvidia has introduced its new DGX GH200 supercomputer model, which houses 256 GH200 superchips that together with NVLink and NVLink Switch System interconnection technology act as a single GPU.
Thanks to this, supercomputers are able to achieve exFLOPS of performance, that is, one trillion floating point operations per second. Likewise, it gets 144 terabytes of shared memory, almost 500 times more memory than the previous generation Nvidia DGX A100, as the company detailed in a statement on its website.
Specifically, Nvidia has explained that the use of the GH200 superchips in conjunction with NVLink technology eliminates the need for a traditional CPU-to-GPU connection. This is because with the Nvidia NVLink-C2C chip, it is possible for each superchip to combine an Nvidia Grace CPU based on the 32-bit ARM architecture with an Nvidia H100 Tensor Core GPU in a single package.
Thus, with NVLink technology, the bandwidth between the GPU and the CPU is increased up to seven times compared to a traditional CPU to GPU connection. This interconnect reduces power consumption by more than five times and provides a 600Gb Hopper architecture GPU building block.
On top of all that, the DGX GH200 model is the first to combine Grace Hopper (GH) superchips with the NVLink switch system, so it’s a new interconnect that enables all GPUs on the superchips to work together as a single GPU. . In fact, this allows the new supercomputer to deliver up to 48 times more NVLink bandwidth than the previous generation.
According to Nvidia, this design allows “to achieve the power of a massive AI supercomputer with the simplicity of single GPU programming.”
Similarly, the company has insisted that some large technology companies are among the first to field the DGX GH200 supercomputer, such as Google Cloud, Meta and Microsoft, to explore its capabilities for generative AI workloads.
As VP of Computing at Google Cloud, Mark Lohmeyer elaborated, “The new shared memory from the NVLink and Grace Hopper superchips addresses key bottlenecks in large-scale AI.”
Similarly, Girish Bablani, corporate vice president of Azure infrastructure at Microsoft, highlighted that the DGX GH200’s ability to work with terabyte-sized data sets “allows developers to conduct advanced research at scale and at an accelerated pace.” will allow.” ,
Along these lines, Nvidia also plans to provide the design of this model to cloud service providers and “other hyperscalers” so that they can “further optimize it for their infrastructure.” And the DGX GH200 supercomputer is expected to be available at the end of this year 2023.
Helios supercomputer
On the other hand, the technology company has also announced that it is building its own AI supercomputer based on the DGX GH200 technology, Nvidia Helios, with the aim of promoting the work of its own researchers.
The Nvidia Helios is powered by four DGX GH200 systems and will boost data performance for training large AI models. To do this, each DGX GH200 system will be connected to an Nvidia Quantum-2 InfiniBand network. Together, the Helios supercomputer will have 1,024 Grace Hopper superchips and is expected to be available by the end of the year.
