At the SIGGRAPH 2023 event held in Los Angeles, USA, NVIDIA announced the upcoming release of the next-generation Grace Hopper Superchip, named GH200. It also launched new RTX workstations for professional computing needs, as well as new professional graphics cards using the Ada display architecture, including the RTX 5000, RTX 4500, and RTX 4000. Furthermore, it launched the new L40S graphics card using the Ada display architecture for the OVX server built for the Omniverse digital twin technology.
The GH200 Grace Hopper Superchip, announced this time, is expected to officially enter the market in the second quarter of 2024. It is equipped with a Grace CPU with 72 cores and a Hopper GPU with 4PFLOPS computing performance. It is also the first acceleration component to adopt HBM3e memory design, with a total of 141GB HBM3e memory capacity, corresponding to a data transmission speed of 5TB per second. Compared with the HBM memory equipped with H100, the capacity has increased by 1.7 times, and the transmission bandwidth has increased by 1.5 times.
In addition, the data transmission part is equipped with a new generation NVLink design, which can connect two GH200 Grace Hopper Superchips in series to form a Grace CPU with a 144-core design and a Hopper GPU with a computing power of up to 8PFLOPS. At the same time, it can increase the HBM3e memory capacity to 282GB, corresponding to a data transmission speed of 10TB per second. Compared with H100, the memory capacity can be increased by 3.5 times, the transmission bandwidth is increased by 3 times, and the memory capacity can even be used with 1.2TB of capacity.
NVIDIA emphasizes that if combined with the stackable and NVLink-connected features, 256 GH200 Grace Hopper Superchips can also be constructed in the form of a SuperPod to form a supercomputer with a computing power of up to 1 EFLOPS and a memory capacity of up to 144TB.
If built with a budget of US$1 million, using the computing power of 8800 x86 CPUs executing the Llama 2 artificial intelligence model, the Vector DB vector database, and the SDXL artificial intelligence image generation model as a 1x computing basis, and consuming 5 megawatts (mW) of power, a supercomputer operated by 2500 GH200 Grace Hopper Superchips can achieve 12 times the computing power, while reducing power consumption to 3 mW, achieving up to 20 times the energy efficiency compared to a CPU-only operation mode.
Even if the goal is to achieve the same computing power, only 210 GH200 Grace Hopper Superchips are needed, and the construction budget is only US$800 million. The power loss can also be reduced to 0.26 megawatts, which can reduce the cost to one-twelfth compared to the CPU operation mode alone.
As for the new RTX workstation launched for professional work needs, it can be equipped with up to four RTX 4 professional graphics cards built with the Ada display architecture, and has a built-in NVIDIA CX6000 6Gbps-level smart network card (SmartNIC). It can also use the latest versions of NVIDIA AI Enterprise and NVIDIA Omniverse software resources. It is expected to be officially shipped this fall.
To meet the expansion and upgrade needs of existing workstations, NVIDIA also announced the launch of professional graphics cards including RTX 5000, RTX 4500 and RTX 4000. They are also built with the Ada display architecture and can be used for automatic generative artificial intelligence training, graphics rendering processing, and Omniverse real-time ray tracing effect rendering. The suggested retail prices are US$4000, US$2250 and US$1250 respectively.
Among them, the RTX 5000 is available from now on, with a dual-slot width and 32GB of display memory design. The RTX 4500 will be available in October, also with a dual-slot width design and a display memory capacity of 10GB. As for the RTX 24, it will enter the market in September, with a single-slot width design and a display memory capacity of 4000GB.
For the OVX servers built for Omniverse digital twin technology, NVIDIA also announced the launch of the new L40S graphics card using the Ada display architecture, which is equipped with 48GB of GDDR6 display memory. Each server can be equipped with up to 8 L40S graphics cards for computing acceleration. Corresponding server products will be launched by companies such as ASUS, Dell, GIGABYTE, HPE, Lenovo, Quanta and Supermicro, and will enter the market as soon as possible.
