THE Nvidia it will always be synonymous with video card, but this is not the only segment to which the company is dedicated. Proof of this is that, on Monday (12), the first day of the GTC 2021 conference, the company announced the Nvidia Grace, a CPU for servers and supercomputers that has ARM technology.
It sounds strange to associate Nvidia with CPUs, but the company has a good reason to invest in this niche: the Nvidia Grace chip was developed to meet the growing demand for applications based on artificial intelligence or that require high performance processing.
As if to prove that it takes this aspect seriously, the company revealed that a Grace CPU machine can offer up to ten times more performance than next-generation Nvidia DGX systems that, in turn, are based on x86 processors.
All of this performance can be achieved without the need for a huge increase in energy consumption.
Like? The technical characteristics of the novelty have not yet been fully released (the number of cores is still a mystery, for example), but Nvidia said that the chip combines low-consumption ARM architecture cores with a memory subsystem that also requires little power, but without affecting overall performance.
The memory subsystem referred to by Nvidia consists of support for LPDDR5x modules, which, compared to DDR4 memories, are up to ten times better in energy efficiency and can have twice the bandwidth on servers and the like.
In addition, Grace can communicate with Nvidia GPUs at rates of up to 900 GB / s (gigabytes per second) because it has the fourth generation of NVLink technology.
Thanks to these and other attributes, CPU Grace is able to handle advanced workloads, including model training for natural language processing (NLP) with more than a trillion parameters, according to Nvidia.
In the official announcement, it is noteworthy that the company highlights that the advances applied in Grace are possible, in large part, to the flexibility of ARM technology.
Probably, Nvidia highlights this point because it announced, in 2020, the purchase of ARM and has been trying, since then, to obtain approval of the deal by regulatory bodies.
But this aspect also makes us question whether the project can be a threat to Intel, which is in the high-performance computing segment.
Still according to Nvidia, the Grace chip will only be available in early 2023, but there are already interested. One of them is the Swiss National Supercomputing Center (CSCS), which announced a supercomputer based on the novelty: the Alps.
Designed in conjunction with HPE and, obviously, Nvidia, Alps is already in the development phase and should be ready in 2023. This machine will replace Piz Daint, the most powerful supercomputer in the CSCS today.
Alps will be used in research for several fields, such as meteorology, materials science, life sciences, particle physics and quantum chemistry.
To this end, the project will be based on the HPE Cray EX line of supercomputers and the Nvidia HGX supercomputing platform, which includes the company’s high-performance GPUs and the Grace CPU.
The expectation is that, with these resources, Alps can offer up to seven times more performance in artificial intelligence tasks than Nvidia Selene, a supercomputer that is a reference for the subject.
In other words, Alps promises to be the world’s most powerful supercomputer for artificial intelligence.
Tecnocast 187 – The global chip crisis
The increase in electronics consumption during the pandemic is causing a shortage in the supply of semiconductors. Manufacturers were not prepared to meet this demand and this gap should affect the price of products in the coming months.
This is the theme of Tenocast 187. Play and check it out!