Most recently, the best Frontier supercomputer has reached a consistent performance of 1.1 exaflops, but more powerful systems may be coming soon.
We are Focus. We are from Ukraine
From the first day of the war, Focus did not stop working for a minute. Our team considers it its duty to inform the reader about what is happening, to collect and analyze facts, to resist enemy propaganda. Today Focus needs your support to continue its mission. Thank you for being with us.
New NVIDIA Grace chips will increase the performance of supercomputers up to 10 exaflops. The developers announced this on May 30 in their press release on the official website.
Grace features two Arm-based advanced architecture CPUs interconnected via a high-speed NVIDIA NVLink-C2C connection for low latency and high power efficiency. The CPUs combine up to 144 high-performance Arm Neoverse cores and a 1 terabyte per second memory subsystem to significantly speed up resource-intensive processes.
The Grace chip features the latest PCIe Gen5 (Fifth Generation Peripheral Component Interconnect Express) computer bus protocol, ensuring compatibility with the most powerful GPUs, as well as NVIDIA ConnectX-7 smart network cards and NVIDIA BlueField-3 processors, allowing you to safely withstand heavy workloads in complex computing and artificial intelligence.
NVIDIA also offers the Grace Hopper superchip, which combines the Grace CPU and NVIDIA Hopper GPU in a single module via an NVLink-C2C connection. Such a complex allows the supercomputer to cope even better with resource-intensive tasks and supports large-scale AI. NVIDIA Grace-based machines will run NVIDIA AI and NVIDIA HPC software.
NVIDIA Vice President Jan Buck says the revolutionary chips will allow supercomputers to solve large-scale tasks that were simply inaccessible before. Grace and Hopper capabilities offer applications in climate science, energy, space exploration, digital biology, quantum computing, and more.
As NVIDIA announced, its superchips will soon be used for the first time by leading centers in the US and Europe to create next-generation servers. The NVIDIA HGX platform will provide manufacturers with blueprints for building supercomputers with the highest performance, power efficiency, and twice the memory bandwidth of other advanced data center processors.
In the US, Los Alamos National Laboratory (LANL) will launch the first system with NVIDIA Grace called the Venado. It will be built on top of the HPE Cray EX supercomputer, combining the Grace processor and the Grace Hopper module to power a vast and growing arsenal of applications. As the developers suggest, the performance of AI in such a supermachine will exceed 10 exaflops (an exaflops is a quintillion (billion billion) operations per second, – ed.).
“By giving LANL researchers the performance of NVIDIA Grace Hopper, Venado will continue this lab’s commitment to pushing the boundaries of scientific breakthroughs. NVIDIA’s accelerated computing platform and vast ecosystem remove performance barriers, enabling LANL to make new discoveries that will benefit the country and society as a whole,” commented Associate Director National Laboratory for Modeling and Computing Irene Kwalters.
In Europe, the Swiss National Computing Center will lead the way by embedding the Grace superchip into a system called Alps, also built on the HPE Cray EX supercomputer. The organization promises to use the high power of the processor in various areas, making the system open to the scientific community in Switzerland and other countries of the world.
Earlier it became known that for the first time engineers were able to overclock a supercomputer to 1.68 exaflops. Researchers at Oak Ridge National Laboratory in the US have achieved a breakthrough on a Frontier system with high performance AMD processors, achieving 1.1 exaflops of sustained power. Thanks to this, Frontier took first place in the Top 500 supercomputers, doubling ahead of its main competitor.