NVIDIA Grace CPU Debuts in Bullsequana XH3000 Supercomputer

Last year NVIDIA introduced us to its first ISA ARM-based CPU for data centers, baptized as Grace, however a processor of this type requires a supercomputer to show off. This is the BullSequana XH3000 from Atos, which will also integrate the Hopper next generation HPC GPU. Let’s see what this almost entirely NVIDIA hardware supercomputer will look like.

If there is a word that is repeated continuously in the supercomputers created with the collaboration of NVIDIA, Intel or AMD, it is Exascale. This refers to the fact that these complex systems have the ability to exceed the ExaFLOP of power, 1018 operations per second. This equates to one million TeraFLOPS or 100,000 times the GPU power of the PlayStation 5.

After Intel with the Aurora supercomputer and AMD with El Capitan showed their muscle from NVIDIA, they could not be less and have managed to get the people of Atos to bet on their next generation hardware for HPC in their BullSequana XH3000 supercomputer. Which will be deployed at the end of this year 2022. Let’s not forget that the race for the ExaFLOP has become in recent years the great obsession of the big brands. Which is evident in this type of systems.

BullSequana XH3000, supercomputer with NVIDIA Grace and Hopper

NVIDIA GRACE Hopper BullSequana XH3000

While we wait for the next GTC Jensen Huang to present us with his next GPU architecture for HPC, baptized as Hopper, the supercomputers that will implement it are already waiting to receive this graphics card. Although let’s not forget that there is a paradigm shift in which both Intel and AMD are offering integrated CPU and GPU solutions under their own brand and without relying on third parties. This phenomenon has left NVIDIA totally displaced, who taking advantage of their previous experience in their Tegra series have created their first ARM CPU for HPC. Named after legendary computer scientist Grace Hopper, NVIDIA will offer an integrated system that combines its new data center processor and new HPC GPU in one system.

And what can we expect in terms of performance? At the moment we do not know the details of the Hopper architecture, but the BullSequana XH3000 can reach a power of 1 64-bit or double-precision floating-point ExaFLOP. Although what is striking is that the people of Atos do use of liquid cooling to get up to 1000 W of consumption for each of the computing nodes or racks on the server. This does not imply that the NVIDIA solution has that level of consumption, but one thing that is clear is that there will be significant growth in terms of the energy consumption of these systems, specifically due to the costs of moving large quantities of data that needs to be processed from your memories to your CPUs and GPUs.

Not buying from ARM has not affected NVIDIA’s plans

NVIDIA’s failed attempt to buy ARM Holdings has not demoralized an NVIDIA that does not want to be the last in a market for said ISA that is expanding looking for other horizons beyond traditional pocket devices. Not only in the automotive and medical market where Jensen Huang already finds itself with its Tegra and Clara solutions, but also in servers with Grace and we should not rule them out in order to integrate their technology into autonomous virtual reality units.

So the next NVIDIA GTC is going to be of great interest, although we do not expect them to present the RTX 40 for gaming, for that we will have to wait a few more months.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *