It was at CES 2022 earlier this year that NVIDIA revealed some of its news for what it showcased today, its new servers based on the Grace platform, the latest in HPC and AI. . And it is that these servers are going to redefine all the market sectors they enter, since Nvidia showed the models HGX Grace Data Center What do you plan to include and… With over 12,000 cores and 1TB of RAM, they will be the benchmark to beat.
Several companies will make available to their customers one of the four types of servers designed by NVIDIA: ASUS, Foxconn, GIGABYTE, OCT, Supermicro and Wiwynn, where their customization will elevate the four types of Huang models to more than a dozen different companies. waiters. They will arrive in a year, 2023, but we already have everyone’s hearts and maximum configuration on the table: HGX Grace CPU Superchips and Grace Hopper Superchip.
NVIDIA HGX Grace: the monstrous server with 12,096 Cores
As always, the details that are not said are the most important, because they reveal data that the company does not want to show even if it is in front of us. The HGX Grace type servers represent the biggest giant leap in this world that we can remember.
They will ship the NVIDIA Grace Superchip processor, i.e. two processors coherently connected via NVLink C2C based on ARM V9 Neoverse and designed for AI and HPC infrastructure. It therefore consists of what NVIDIA calls a CPU-CPU module and counts for each PCB that integrates no less than 144 hearts with memory LPDDR5X with up to 1TB per rack and a bandwidth of at least 1 TB/s.
while consuming 500 watts and can be cooled by air or water. The most impressive thing here is that NVIDIA allows servers with up to 84 nodes per rack, which is huge 12,096 cores in all. Considering the HGX Grace is 1.5 times faster than the DGX A100 to begin with, one can already imagine the beast the green team has created.
HGX Grace Hopper: CPU and GPU on a PCB
This type of server bets on NVIDIA’s second option. And it is that here we do not have two CPUs, but CPU and GPU on the same substrate, which communicate again by NVLink C2C in order to have a consistent high performance memory model that will be interconnected to 900 GB/s being 7 times faster than the PCIe 5.0 bus.
What NVIDIA has created is the ultimate multitasking server, capable of working with any enterprise software stack, whether for HPC, AI and Omniverse, so it is multifaceted. The pattern to follow is simple here with these servers HGX Grace Hoppereach will integrate a Hopper GPU to 4nm with a Grace processor, where each will have its own memory, the first having no less than 80 GB of HBM3
This added to the total bandwidth would give us 3.5 TB/s with a total consumption of 1000 watts per rack and having the option of being air or water cooled. NVIDIA guarantees that it can be installed 42 knots Grace Hopper HGX rack.
Server designs and your portfolio
There will be four specific designs depending on the workload that manufacturers may require, where within these four types anyone can configure and customize them to suit their needs, leaving a wide range of benefits and price:
- Systems NVIDIA HGX Grace Hopper for AI training, inference and HPC are available with Grace Hopper Superchip and NVIDIA BlueField-3.
- systems NVIDIA HGX Grace for HPC and supercomputing: they feature a CPU-only design with Grace Superchip, BlueField-3 processors and NVIDIA GPUs.
- Systems NVIDIA OVX for digital twins and collaborative workloads are powered by Grace CPU Superchip, BlueField-3 and NVIDIA.
- Systems NVIDIA CGX for graphics and cloud gaming, they are equipped with Grace Superchip processors, BlueField-3 GPUs and NVIDIA A16.
As expected, no price has yet been revealed, since we are one year away from its official launch, so expect NVIDIA to reveal the costs over this period, which surely will not be cheap.
